Authors: Yuqi Liu, Chen Sun, Yuning Niu, Xu Wang, Zehua Yue, Tieqiang Zhang, Jiang Li, Xiudong Guan, Dainan Zhang, Wang Jia
Categories: Article, Cancer, Computational biology and bioinformatics, Engineering, Mathematics and computing, Medical research, Oncology
Source: NPJ Digital Medicine
Authors: Yuqi Liu, Chen Sun, Yuning Niu, Xu Wang, Zehua Yue, Tieqiang Zhang, Jiang Li, Xiudong Guan, Dainan Zhang, Wang Jia
Brain tumor MRI segmentation and classification are essential for preoperative boundary assessment, lesion burden quantification, postoperative response monitoring, and radiotherapy planning, yet edema overlap, sequence heterogeneity, and artifacts often blur lesion margins. Together with the high cost of pixel-level annotation, these factors limit robust, cross-institution deployment. We propose DARE-FUSE (Domain Aligned Representation with Evidence-guided FUSE), a unified framework for pixel-level segmentation and image-level classification under limited samples and labels. Dual encoders with a feature-interaction bridge learn a shared embedding, and a Domain Alignment Refiner maps it to task-aligned representations for the segmentation and classification branches. For segmentation, U-SEG decodes features and SEGU outputs pixel-wise uncertainty to regularize boundary over/under-segmentation. For classification, CPG produces predictions and multi-scale Grad-CAM++ evidence. A Generative Lesion Removal Prior reconstructs a tumor-free counterpart to yield a difference prior, and FUSE combines this prior with Grad-CAM++ under uncertainty attenuation to guide segmentation and suppress hallucinations. DARE-FUSE achieves stable, leading performance on BraTS segmentation benchmarks and several classification datasets; ablations and label-reduction experiments confirm complementary gains and smooth degradation as pixel annotations decrease. The resulting uncertainty maps and continuous priors support interpretable decision assistance in surgery, radiotherapy contouring, triage, and longitudinal follow-up.
Brain tumors are life-threatening central nervous system diseases that often cause severe neurological deficits, long-term disability, and high mortality^1,2^. Frequent recurrence and progression impose a heavy global burden on individuals and healthcare systems^1,2^. Despite advances in imaging, surgical, and chemoradiation techniques, timely, accurate identification, characterization, and quantitative assessment of brain tumors remain clinically challenging^3–5^. MRI, the primary imaging modality for brain tumor evaluation, provides multi-parametric information to support lesion detection, preoperative grading, and treatment response assessment^3,4,6^. However, image interpretation is constrained by radiologist availability, fatigue, case complexity, and the time-consuming reading and contouring process, causing inter-observer variability^7–10^. Additionally, overlapping tumor infiltration and surrounding edema, sequence and inter-center protocol variations, and metal artifacts blur lesion boundaries, increasing the risk of misdiagnosis and further undermining the reliability of critical therapeutic choices^5,11–13^.
Timely and comprehensive imaging assessment is crucial for improving prognosis^3,4^. Current clinical management relies heavily on scoring systems and rules based on limited variables, including response and progression criteria dependent on 2D linear measurements and subjective interpretation, and risk stratification tools using a small set of clinical and laboratory indicators^14–16^. These approaches struggle to adequately characterize key factors influencing disease severity and long-term outcomes, such as infiltrative margins, subregional heterogeneity, treatment-related changes, and recurrence risk, hindering clinicians’ ability to accurately evaluate disease status and tailor treatment strategies^17,18^. Accurate assessment of these aspects requires sophisticated quantitative imaging and systematic reporting^19–21^. Consequently, there is an urgent clinical need for more efficient, stable, and reproducible imaging analysis and reporting workflows to enhance the quality of diagnostic and therapeutic decision making in complex real-world settings^10,20,22^.
Existing methods in pixel-level segmentation primarily center on U shape encoder–decoder architectures and multi-scale context, exemplified by U-Net and related methods, as well as ViT-based UNETR and SwinUNETR^23–25^. In parallel, recent years have seen the transfer of general-purpose segmentation paradigms represented by SAM and MedSAM to medical scenarios^26,27^. These approaches achieve high scores on standard benchmarks, yet they still show insufficiency in boundary robustness and calibration at the infiltration^11,28^. Meanwhile, the zero-shot behavior of general-purpose segmentation models on medical imaging is dependent on large-scale dataset for training and requires further evaluation^26,29,30^.
Concurrently, the discrimination pathway for brain tumors is represented by the ViT family and contrastive learning, and has gradually incorporated medical-specific vision-language pretraining models represented by MedCLIP and BiomedCLIP and weakly supervised multiple instance learning represented by Transformer-based MIL^31–35^. This line of work emphasizes global semantics and cross-domain robustness but lacks direct modeling of pixel-level spatial evidence; CAM-based weakly supervised localization often exhibits a bias toward only the most discriminative regions, leading to incomplete foreground contours and delayed responses at boundary transitions^36,37^. Furthermore, in a unified framework, directly merging data that contain only classification labels into a shared trunk may cause task-level negative transfer when decoupling and constraints are absent, thereby affecting subsequent segmentation quality^38,39^.
In summary, the key to overcoming current bottlenecks is to align and fuse discriminative evidence and structural evidence within a unified system, and to propagate uncertainty as a unified weight throughout evidence construction and segmentation prediction.
To this end, this paper proposes Domain Aligned Representation with Evidence guided FUSE (DARE-FUSE), a unified framework built around task-oriented alignment, an evidence fusion closed loop, and uncertainty-aware decoding. The framework consists of multiple tightly coupled modules within a single pipeline, including Dual encoders with a cross-layer Feature interaction bridge to extract and interactively fuse complementary multi-scale representations from the perspectives of segmentation and classification; Domain-Alignment Refiner (DAR) performs statistical alignment and adversarial correction in the medical imaging domain, mitigates cross-center and cross-year distribution shifts, and outputs task-oriented embeddings. Classification with Pseudo mask Generator (CPG) generates image level predictions in the classification branch while extracting multi scale evidence heatmaps that provide initial cues to tumor location; Generative Lesion Removal Prior (GLRP) reconstructs tumor free images with a generative model and computes their difference from the original image to produce a differential prior that is highly sensitive to lesion structural changes, thereby providing strong boundary related evidence without pixel level annotations; Fusion based Uncertainty weighted Seed Estimator (FUSE) fuses multi source evidence at the pixel level, including class activation maps and the differential prior, and generates a continuous tumor seed map using an uncertainty weighted combination of intersection and union; finally, U-shape Decoder and Segmentation Head with Uncertainty (U-SEG with SEGU) injects the above priors during decoding to produce fine-grained pixel level segmentation while simultaneously outputting uncertainty estimates of the predictions. The pipeline of our proposed method is shown in Fig. 1.Fig. 1DARE-FUSE: a unified pipeline for brain-tumor MRI that couples discriminative evidence with structural priors under weak supervision.Generative Lesion Removal Prior (GLRP) reconstructs a tumor-free image, the differential prior is computed. Dual encoder consists of a Seg-Encoder and a Cls-Encoder, the features are fused by a cross-attention Feature Bridge. Task-oriented Domain-Alignment Refiner (DAR) combines higher-order rectification and adversarial constraints (GRL) and outputs weighted task-specific embeddings. The segmentation path decodes pixel-level masks and produces an uncertainty map, while the classification path (CPG) yields image-level probabilities and multi-scale Grad-CAM++ heatmaps. The Fusion based Uncertainty weighted Seed Estimator (FUSE) merges removal prior and multi-scale CAMs into a continuous prior P and an ignore mask Mignore, pixel-wise uncertainty down weights unreliable regions. Then P and prior are injected into decoder features with learnable gates at aligned scales, sharpening boundaries while suppressing spillover.
We conduct a systematic evaluation on six BraTS BraTS 2015 and BraTS 2016 are treated strictly as external test sets, and the official splits are used for BraTS 2018–2021^40–45^. The results show that DARE-FUSE achieves consistent improvements over recent strong baselines on whole tumor Dice and HD95, with particularly notable advantages on complex boundaries and small lesions^46^. Meanwhile, on two classification datasets, AUC and F1 at the Youden point surpass representative lightweight ViT and weakly supervised methods from the past two years^47^. Modular ablations indicate that DAR contributes the most to cross-task decoupling and representation stability; the GLRP and FUSE cascade reduce over-segmentation near boundaries and lead to sustained convergence of HD95^46^. The uncertainty from SEGU both lowers predictive variance and, as an evidence down-weighting signal, substantially reduces false positives. Cross-center and cross-protocol experiments further demonstrate that DARE-FUSE maintains robustness under domain shift without increasing model size or requiring additional supervision^11^.
In summary, DARE-FUSE provides a structured solution to the core challenges of clinical application for brain lesions. While remaining lightweight and interface-unified, the system offers a practical technical pathway for the clinical usability and cross-domain deployment of brain tumor MRI. Future work will focus on finer pathological subregions and longitudinal follow-up, extending the evidence closed loop to the temporal dimension and multi-sequence joint modeling to further enhance interpretable identification of boundary dynamics and treatment response.
The contributions of this paper are summarized as follows. (1) Task-oriented representation we propose DAR, which combines first-order and second to higher order statistical alignment on shared embeddings and imposes multi-scale adversarial non-separability regularization for domain invariance, and finally obtains task-oriented outputs through a dual-head lightweight projection. This design ensures that when only classification labels participate in training, freezing the segmentation path still preserves subsequent segmentation performance, addressing the long-standing pain point of task interference in a unified framework. (2) Evidence-fusion closed we propose the GLRP with the FUSE scheme for fusing structural and discriminative evidence. GLRP provides a differential prior that is in phase with structural changes. FUSE adopts a closed-form construction, where intersection ensures purity, union with distance-based attenuation complements coverage, and unified uncertainty-based downweighting is applied. This converts two coarse evidence sources into a single, continuous, and interpretable prior P that aligns with the decoder’s scale, effectively alleviating contour breaks and false positives under weak supervision. (3) Uncertainty-aware SEGU integrates U-shape multi-scale fusion with MC-Dropout sampling. Uncertainty is used for segmentation calibration and risk exposure in boundary-blurred regions, and also permeates the evidence pathway and loss scheduling to form a closed-loop consistent confidence scheme. (4) Unified interfaces and all modules interface through stable “scaled tensors plus single-channel priors and masks,” without relying on extra interactions or specific sequence combinations, facilitating training and deployment under multi-center, heterogeneous protocols and mixed multi-task data conditions.
To ensure strict data isolation and prevent subject leakage, we implemented a rigorous subject-level de-duplication protocol across the four Kaggle MRI classification datasets. We first grouped images by patient identifiers and metadata, and then employed image-level perceptual hashing as a secondary verification step to flag and remove near-identical images across cohorts. All images belonging to a duplicated subject were removed from all but one dataset, ensuring that each real-world subject contributes to at most a single benchmark. Following de-duplication, we constructed stratified train, validation, and test splits strictly at the subject level, assigning all slices from the same subject to a single split to preserve the original class distribution without data leakage.
We use four publicly available image level MRI datasets from Kaggle: Brain Tumor Classification (MRI) (four glioma, meningioma, pituitary, no tumor), Brain Tumor MRI Dataset (7,023 images across the same four classes), Brain MRI Images for Brain Tumor Detection (two tumor, no tumor), and Brain Tumor MRI Images 17 Classes (Tombackert collection; 4,449 axial images with an expanded taxonomy of 17 classes). All classification datasets are used with image-level labels only; when author-provided splits exist, we adopt them, otherwise we construct stratified train and test splits at the subject level to prevent leakage. Datasets are described in Table 1.Table 1Datasets used in this studyDatasetModalitiesSamples**Segmentation cohorts (multi-parametric MRI, 4 modalities)**BraTS-2015 (external test only)mpMRI274 subjects ^a^BraTS-2016 (external test only)mpMRI274 subjects^a^BraTS-2018mpMRI285 training cases^b^BraTS-2019mpMRI460subjects^c^BraTS-2020mpMRI494 subjects^d^BraTS-2021mpMRI1251 training cases^e^**Classification datasets (image level labels)**Brain Tumor Classification (MRI) — BhuvajiJPG3064 images^f^Brain Tumor MRI Dataset — NickparvarJPG7023 images^g^Brain Tumor MRI Data – TombackertJPG7153 images^h^Brain Tumor MRI Images – 17 ClassesJPG4449 images^i^Resolution and modalities and sample sizes are reported as available from official pages or dataset descriptions.^a^ BraTS 2015/2016 training set composition reported as 220 HGG + 54 LGG (total 274).^b^ BraTS-2018 training set reported as 285 cases; standard held-out evaluation.^c^ BraTS-2019 training and validation sizes reported as 335/125 cases.^d^BraTS-2020 training and validation sizes reported as 369/125 cases.^e^BraTS-2021 training cohort reported as 1251 cases (four modalities).^f^Widely used “Sartaj” four-class brain MRI set; 3064 images (Figshare lineage).^g^Combined four-class set aggregating Figshare/SARTAJ/BR35H; 7023 images.^h^Curated Kaggle compilation with provided splits; counts from dataset notebook.^i^“Brain Tumor MRI Images–17 Classes” reports 4449 axial MRIs with expanded taxonomy.Note: Several Kaggle classification sets partially overlap in provenance (Figshare/SARTAJ/BR35H); we treat them as distinct benchmarks due to differing curation and splits.
All experiments run on a single NVIDIA RTX 4090 with 24 GB VRAM under Linux. The codebase uses Python 3.10 with PyTorch 2.3 and CUDA 12.1 together with cuDNN 9. Mixed precision relies on torch.amp autocast and GradScaler. Medical imaging utilities follow MONAI 1.5.1 for data I O and sliding window inference. Metric computation uses TorchMetrics with synchronized reduction on a single device.
For BraTS segmentation tracks each subject contains the four standard modalities T1 T1Gd T2 and FLAIR as provided by the organizers. Volumes are resampled to one millimeter isotropic spacing, then z-score normalized within a brain mask and cropped to foreground. Training uses three-dimensional patches of size 160 by 160 by 128 with a foreground oversampling probability of 0.6. Spatial augments include random flip probability 0.5, small rotation up to ten degrees small-scale jitter up to ten percent, elastic deformation with alpha 90 and sigma 10 gamma correction magnitude up to ten percent, Gaussian noise standard deviation in the range 0 to 0.05, Gaussian blur sigma in the range 0.5 to 1.5 and random bias field probability 0.3. For image-level classification, tracks two-dimensional MRI frames, are center-cropped or padded, then resized to 224 by 224, followed by per-image z-score normalization. Color jitter is disabled. Geometric augments mirror the three-dimensional settings in two dimensions with the same magnitude ranges.
Optimization uses AdamW with cosine decay and linear warmup. Global gradient norm clipping equals 1.0. Weight decay equals 0.05 on encoders and 0.01 on decoders and heads. An exponential moving average of model weights with decay 0.999 is maintained for evaluation. Mixed precision is enabled during forward and backward, while normalization layers and loss reductions run in float32. Segmentation on BraTS uses a base learning rate of 2e−4 with a ten-epoch warmup and a total of 100 epochs. Batch size equals two patches per step with gradient accumulation to emulate batch four when VRAM is tight. Classification on Kaggle datasets uses a base learning rate of 3e−4 with a five-epoch warmup and a total of 100 epochs. Batch size equals sixty-four images.
Dual encoders and the feature interaction bridge follow the architecture defined in the Methods section. DAR applies AdaIN and IBN style whitening, then CORAL and MMD alignment, together with a gradient reversal-based adversarial discriminator at each scale. The default alignment weights are CORAL 0.10 MMD 0.05 and adversarial 0.05. During optimization, segmentation and classification losses both use unit weight and are active for all epochs. All DAR losses share a global coefficient λDAR and are linearly ramped from 0 to their final value over the first 5 epochs; within this budget, the feature-alignment terms receive 0.8 λDAR and the adversarial term 0.2 λDAR, so that DAR acts as a regularizer rather than a competing objective. U SEG adopts a four-level U style decoder with one-by-one lateral projections and three-by-three convolutions. MC Dropout with a rate of 0.3 is enabled in the last two decoding stages. For BraTS segmentation experiments that rely on FUSE and uncertainty maps, both validation and test-time inference perform K = 8 stochastic forward passes and aggregate the mean and variance of logits to derive the uncertainty map reused by the other modules. For all classification results, as well as ablation rows where FUSE/uncertainty is disabled, we instead use a single deterministic forward pass with dropout deactivated. SEGU outputs a multi-channel foreground probability for WT TC and ET masks using a sigmoid activation. The main loss combines Dice and binary cross entropy with an additional narrow band boundary cross entropy weight 0.3. A weak seed loss guided by the continuous prior P weight 0.3 is applied only on confident pixels. An empty mask consistency term on the no tumor reconstruction from GLRP weight 0.2 reduces over-segmentation and hallucinations. CPG aggregates multi-scale features by adaptive pooling and a linear classifier with label smoothing of 0.1, then computes multi-scale Grad CAM plus on fixed intermediate blocks. Class activation maps are normalized and can be down weighted by one minus the normalized uncertainty before being consumed by FUSE. GLRP relies on a LaMa like inpainting backbone to produce a no-tumor reconstruction. The difference between the input and the reconstruction is normalized by the L1 norm to obtain the structural prior Delta. The inpainting objective combines reconstruction loss, identity consistency weight 0.10, perceptual alignment weight 0.05, Fourier domain alignment weight 0.05, and total variation regularization weight 0.01. FUSE converts Delta together with multi-scale CAM evidence into a continuous prior P and an ignore mask. In practice, the voxel-wise uncertainty map Umap ∈ [0, 1] produced by SEGU is shared across modules and partitioned into three operating regimes. Voxels with low uncertainty (Umap(x) < τconf) are treated as reliable foreground or background seeds, participate in the P-guided weak seed loss, and are eligible to be reinforced by the continuous prior. Voxels with intermediate uncertainty are kept in the prediction but excluded from weak supervision so that ambiguous boundary bands are not over-constrained. Voxels with high uncertainty (Umap(x) ≥ τignore), together with locations where CAM and GLRP evidence strongly disagree, are collected into the ignore mask Mignore. During training, Mignore suppresses gradients from P and prevents contradictory pseudo-labels in these regions; during inference, it is exported alongside the probability map as a binary overlay that indicates regions where the model output should be regarded as assistive only rather than fully automatic. In this work, τconf and τignore are chosen once on the BraTS validation split and then fixed for all ablations and comparison runs. Soft thresholds use temperature two point zero with tau Delta equal 0.30 and tau cam equal 0.40. The uncertainty attenuation exponent gamma equals 1. The core and extension region weights are 1.0 and 0.5. P and Delta are injected into all decoder scales by gated additive modulation with learned gates and one-by-one channel alignment.
Three-dimensional inference follows a sliding-window strategy with region-of-interest size 160 × 160 × 128, window batch size 4, and overlap 0.75. Windowed logits are blended using Gaussian weights for smooth tiling. Under this setup, the computational cost of BraTS segmentation scales approximately linearly with the number of stochastic passes K. Our default choice K = 8 trades higher wall-clock time for more stable uncertainty maps and slightly improved small-lesion sensitivity, while remaining practical for offline batch processing on a single RTX 4090. In high-throughput deployment scenarios, K can be reduced to recover near-baseline runtime at the expense of less stable uncertainty estimates; a systematic runtime–accuracy characterization over different K and batch configurations is left for future work. Classification uses center crop evaluation. Five crop testing is optional for robustness analysis.
To keep the training objective manageable, all loss weights and thresholds are selected once and then held fixed across datasets, ablations, and label-scaling experiments. The coefficients for DAR regularization, the segmentation stack and the GLRP reconstruction are set by a small number of pilot runs on the BraTS validation split, following standard practice for deep medical segmentation.
For comparison models with publicly released implementations, we retrain the official code on our subject-level splits using the same preprocessing, augmentation, and optimization schedules as DARE-FUSE, and we adjust batch size and learning rate only within the ranges recommended by the original authors. For methods that are available only through published summaries or challenge leaderboards, we report the numbers provided in the corresponding papers under the stated evaluation protocol. Whenever retraining is possible, all models share an identical train, validation and test partition so that differences in performance are not confounded by custom data splits.
For segmentation, the best checkpoint is selected by the highest mean Dice on the validation fold with HD95 as the tie breaker. For classification, the primary criterion is AUC on the validation fold with F1 at the Youden threshold as the secondary criterion. Exponential moving average weights are used for reporting. All experiments are repeated with three independent random seeds. For each run, early stopping is triggered on the fixed subject-level validation split described in the section “Datasets”, using mean Dice over WT, ET, and TC with HD95 as a tie-breaker for BraTS and AUC with F1 at the Youden threshold as a secondary criterion for Kaggle. The train/validation/test partition is identical across all baselines, ablations, and comparison model. Across runs, we re-sample only network initialization and mini-batch ordering. Unless otherwise specified, all metrics in the tables are reported as the mean value over these three independent runs. Standard deviations were calculated to verify experimental stability and were found to be consistently low; however, they are omitted from the main comparison tables to strictly adhere to page limits and preserve legibility. Regarding inference protocols, for all BraTS segmentation results relying on the fusion module, we performed K = 8 stochastic forward passes to generate the final prediction. In contrast, classification results utilized a single deterministic forward pass.
Figure 2 summarizes the classification results. Detailed Accuracy and AUC across four public brain MRI datasets are put in Table 2. The proposed head achieves the best results on all datasets with a mean accuracy of 0.957 and a mean AUC of 0.991, improving over the strongest non-ours baselines by small yet consistent margins. On Masoudnickparvar, the accuracy and AUC reach 0.968 and 0.995, exceeding supervised ViT B/16^31,48,49^. On Bhuvaji, the gains are 0.004 accuracy and 0.006 AUC. On Tombacker,t the gains are 0.004 accuracy and 0.006 AUC. On the 17 Classes benchmark, our method attains 0.934 accuracy and 0.983 AUC, surpassing the best baseline by 0.006 and 0.005, respectively, which is notable given the expanded taxonomy and class imbalance.Fig. 2Overview of dataset composition, ROC performance, heatmap refinement, and summary indices.a Class-distribution pie charts for 17-Classes, Tombackert (4 classes), Masoudnickparvar (4 classes), and Bhuvaji (4 classes) display the category proportions across each dataset, providing an imbalanced context for subsequent performance metrics. b Average ROC comparisons across four datasets--ViT-B/16, BiomedCLIP, TransMIL++, and DARE-FUSE. DARE-FUSE achieves higher TPR at the same FPR. c Attention heatmaps progressively converge toward lesions as modules advance, reducing boundary leakage and false activations while better aligning with irregular tumor margins. d Radar charts summarize cross-dataset classification AUC and BraTS segmentation Dice scores, demonstrating stable gains across multiple metrics for the full model compared to baseline and ablation.Table 2Brain tumor MRI comparison across four Kaggle datasetsMethodMasoudnickparvar (4cls)Bhuvaji (4cls)Tombackert (4cls)17-ClassesAcc ↑AUC ↑Acc ↑AUC ↑Acc ↑AUC ↑Acc ↑AUC ↑Classic CNN/Transformer (supervised fine-tuning; reference)ResNet-50 (supervised)0.9550.9870.9480.9820.9520.9840.9100.964EfficientNet-B0 (supervised)0.9620.9900.9540.9850.9600.9870.9220.972ViT-B/16 (supervised)0.9640.9910.9560.9860.9620.9880.9280.975Medical zero-shot (no fine-tuning; prompt-based classification, 2024–2025)CLIP (ZS, prompt-tuned)0.4820.3440.3740.4380.4880.5460.4420.520BiomedCLIP (ZS)0.9070.9620.8980.9560.9060.9600.8660.938VPL-CLIP (ZS proxies, 2024)0.9280.9700.9180.9640.9240.9680.8860.950MedCLIP-SAM (ZS, 2024)0.9340.9740.9240.9680.9300.9720.8920.954Few-shot learning (k-shot per class, 2024–2025)MAML (2024)0.9020.9920.8940.9900.9000.9910.8620.982ProtoNet (2025)0.9140.9840.9060.9810.9120.9830.8760.972DMA-Net (meta-attentive, 2025)0.9360.9880.9260.9850.9320.9870.8960.976Few-shot multi-modal fusion (2025)0.9420.9900.9320.9870.9380.9890.9040.978Weakly supervised (image level labels; CAM/MIL-enhanced, 2024–2025)UM-CAM with classifier (2024)0.9380.9810.9280.9780.9340.9800.8960.966SCC-CAM (2024)0.9460.9840.9360.9810.9420.9830.9040.970SSC (2024)0.9480.9860.9380.9830.9440.9850.9080.972From SAM to CAMs (CVPR’24) + cls0.9520.9880.9420.9850.9500.9870.9140.975TransMIL++ (2024)0.9580.9900.9480.9880.9540.9890.9200.978Our Method**DARE-FUSE**0.968**0.995**0.960**0.992**0.966**0.994**0.9340.983Columns report, for each Accuracy (Acc ↑) and AUC (↑). Recent methods (2024–2025) cover zero-shot, few-shot, and weakly supervised families; a small block of classic CNN/Transformer baselines is included for reference. All metrics are reported as mean ± one standard deviation over three independent runs with different random seeds.Datasets. Masoudnickparvar (4 classes, 7023 images); Bhuvaji (4 classes, ~ 3264 images); Tombackert (Kaggle collection for 4-class tasks); 17-Classes (4449 images).Families. ZS = zero-shot; CAM = class activation map; MIL = multiple instance learning; k-shot as noted. Classic rows serve as reference baselines; all other families list methods from 2024–2025.
Family wise trends are consistent with expectations. Classic supervised CNN and Transformer baselines remain strong and provide high starting points across the four class datasets but drop more noticeably on the 17 Classes set^31,48,49^. Medical zero-shot models show competitive AUC yet lower accuracy, reflecting limited localization and domain-specific priors under label shift^34,50–52^. Few shot learners close part of the gap on the four class tasks but remain below supervised fine tuning on the harder taxonomy^53–56^. Weakly supervised approaches that rely on CAM or MIL improve localization driven cues and generally track the classic baselines, though they still trail when class granularity increases^35,57–59^.
The proposed design combines task-oriented alignment and evidence guidance to obtain the observed gains. DAR aligns the shared embedding toward classification discriminative subspaces without degrading segmentation features. CPG aggregates multi-scale features and produces Grad CAM++ evidence that FUSE quantifies with uncertainty attenuation, yielding ROI aware representations that reduce background confounders^60^. GLRP supplies a structural prior that complements CAM evidence in difficult cases with edema and low contrast. With SEGU frozen for ROI gating, these components translate into consistent improvements in both accuracy and AUC across datasets of different curation and label granularity.
Across six BraTS editions, the proposed framework achieves consistent improvements in Dice and HD95 for whole tumor, enhancing tumor, and tumor core^40–45^. On the external BraTS 2015 and BraTS 2016 test sets with zero training exposure, the method sustains competitive Dice and reduced surface distances, indicating generalization under protocol and site shifts^40,41^. On the official held out splits of BraTS 2018, BraTS 2019, BraTS 2020, and BraTS 2021, gains over recent supervised transformer and diffusion baselines, as well as zero-shot and weakly supervised approaches, are observed with the largest margins on cases featuring edema overlap, low contrast, and small volumes^42–45^. These trends align with the DAR provides task-oriented alignment that stabilizes shared features under mixed supervision, GLRP supplies a structural difference prior that sharpens boundaries, FUSE converts heterogeneous evidence into a calibrated continuous prior that improves localization while suppressing hallucinations through uncertainty attenuation, and SEGU yields uncertainty maps that correlate with boundary ambiguity, jointly lowering HD95 without sacrificing Dice^40,45^.
On all six BraTS benchmarks the method ranks first, with consistent gains over recent supervised transformers and diffusion models as well as over zero shot, few shot, and weakly supervised families, as shown in Table 3^23,25–27,35,40–45,59,61–66^. Compared with the strongest supervised baselines, Dice improves by 0.008 on BraTS 2018 and BraTS 2019 over 0.908 and 0.914 to 0.916 and 0.922, while HD95 drops by 0.5 mm from 8.8 and 8.4 to 8.3 and 7.9. On BraTS 2020 and BraTS 2021 Dice rises from 0.936 and 0.925 to 0.939 and 0.928 with HD95 reduced from 5.6 and 7.6 to 5.3 and 7.2. On the external BraTS 2015 and BraTS 2016 sets Dice increases by 0.008 over the best references to 0.896 and 0.904 with HD95 reduced by 0.5 to 0.9 mm. Classic U Net family baselines span Dice 0.855 to 0.914 with HD95 12.1 to 8.1, trailing more on early editions^23,63–65^. Medical zero-shot variants centered on SAM and MedSAM reach Dice 0.842 to 0.904 with HD95 14.2 to 9.7, remaining below supervised counterparts^26,27,66^. Few-shot approaches narrow the gap to Dice 0.872 to 0.912 and weakly supervised methods reach 0.864 to 0.906 with gradual HD95 reductions, but both families remain behind the proposed framework across all editions^35,59^.Table 3Brain tumor MRI comparison across six BraTS benchmarksMethodBraTS-2015BraTS-2016BraTS-2018BraTS-2019BraTS-2020BraTS-2021Dice ↑HD95 ↓Dice ↑HD95 ↓Dice ↑HD95 ↓Dice ↑HD95 ↓Dice ↑HD95 ↓Dice ↑HD95 ↓Classic U-Net family (reference baselines)U– Net (3D)0.85512.10.86211.40.87810.20.88210.00.8869.80.8909.6U– Net++ (3D)0.86211.60.86811.10.8859.80.8899.60.8929.30.8969.1Attention U– Net (3D)0.86411.30.87110.90.8889.60.8919.40.8959.10.8998.9nnU– Net (3D)0.8769.40.8849.00.8988.60.9028.30.9108.50.9148.1Medical foundation & zero-shot segmentation (SAM family, 2024–2025)SAM– Med2D (ZS, prompt)0.84214.20.85013.60.86212.80.86612.40.87212.00.87811.6MedSAM (ZS, prompt)0.84813.70.85613.10.86812.30.87211.90.87811.50.88311.1MedSAM– Adapter (ZS + adapter)0.85812.90.86512.40.87811.60.88311.10.88810.80.89310.4SAM– Med3D (ZS, volume prompts)0.86212.60.86912.10.88211.20.88710.80.89210.50.89710.1DB– SAM (MICCAI’24, ZS)0.86612.20.87411.70.88610.90.89210.40.89710.10.9029.8FastSAM3D (ZS, 3D prompts)0.86812.00.87611.50.88810.80.89410.30.89810.00.9049.7Few-shot brain tumor segmentation (2024–2025)DASA (ACCV’24) few-shot0.87211.60.88011.10.89210.20.8989.80.9049.40.9099.1MetaMed3D (2025) few-shot0.87511.40.88310.90.89410.00.9009.60.9069.20.9118.9ProtoFSS (2025) few- shot0.87611.30.88410.80.8969.90.9019.50.9079.10.9128.8Weakly supervised tumor segmentation (2024–2025)Box– sup. + MedSAM (weak)0.86812.10.87611.60.88910.70.89510.20.9009.90.9059.6Point– sup. w/ box priors (weak)0.86412.40.87211.90.88611.00.89210.50.89710.20.9029.9Scribble– prompt (weak + interactive)0.86612.20.87411.70.88810.90.89310.40.89810.10.9049.8UM– CAM + seed refine (weak)0.87012.00.87811.50.89010.60.89610.10.9019.80.9069.5Recent fully supervised SOTA (2024–2025 transformers/diffusion)SwinUNETR (MONAI’24 cfg)0.88010.80.88810.40.9009.40.9079.00.9128.70.9188.4PA– SwinUNETR (2025)0.88410.50.89210.10.9049.10.9108.80.9168.40.9228.1DMCIE (2024)0.88610.30.8949.90.9068.90.9128.60.9345.90.9237.9FCFDiff– Net (2025, diffusion)0.88810.10.8969.70.9088.80.9148.40.9365.60.9257.6ACS– ResNet (2025)0.88210.70.89010.30.9029.30.9098.90.9148.60.9208.3Our Method**DARE-FUSE**0.896**9.6**0.904**9.2**0.916**8.3**0.922**7.9**0.939**5.3**0.9287.2Columns report, for each dataset, Dice (WT, ↑) and HD95 in mm (↓). BraTS-2015 and BraTS-2016 are external test sets (no training on them); BraTS-2018/2019/2020/2021 use the official held-out test sets. All metrics are reported as mean ± one standard deviation over three independent runs with different random seeds.Notes. WT = whole tumor. ZS = zero-shot. Weak supervision rows use image/point/box/scribble-level labels and CAM-derived seeds; few-shot rows assume k-shot per class with identical splits across methods. BraTS-2015/2016 are evaluated strictly as external test sets (no training exposure); other columns use official held-out test sets. Numbers are aligned with recent reports (2024–2025) and consistent metric conventions (Dice/HD95 in mm).
Besides, we also evaluated ET across BraTS-2015 to BraTS-2021 in Table 4. DARE-FUSE attains the top Dice and consistently lower HD95 on every benchmark, with representative improvements over the strongest supervised transformer and diffusion baselines as on BraTS-2015 Dice rises from 0.852 to 0.864 and HD95 drops from 11.2 to 10.6; on BraTS-2016 Dice increases from 0.860 to 0.872 with HD95 reduced from 10.8 to 10.2; on BraTS-2018 and BraTS-2019 Dice improves from 0.872 and 0.880 to 0.884 and 0.892 while HD95 decreases from 10.1 and 9.7 to 9.5 and 9.1; on BraTS-2020 Dice moves from 0.902 to 0.910 with HD95 reduced from 6.2 to 6.0; on BraTS-2021 Dice increases from 0.892 to 0.898 and HD95 declines from 8.6 to 8.3. Classic U-Net family baselines span Dice 0.812 to 0.884 with HD95 13.8 to 9.6 and lag more on early editions. Medical zero-shot variants centered on SAM and MedSAM remain behind supervised methods with Dice 0.798 to 0.866 and HD95 15.6 to 11.2. Few-shot approaches narrow the gap to Dice 0.832 to 0.874 and weakly supervised methods reach 0.824 to 0.868 with gradual HD95 reductions, yet both families consistently trail the proposed framework across all editions.Table 4Brain tumor MRI segmentation (ET): comparison across six BraTS benchmarksMethodBraTS-2015BraTS-2016BraTS-2018BraTS-2019BraTS-2020BraTS-2021Dice ↑HD95 ↓Dice ↑HD95 ↓Dice ↑HD95 ↓Dice ↑HD95 ↓Dice ↑HD95 ↓Dice ↑HD95 ↓Classic U-Net family (reference baselines)U– Net (3D)0.81213.80.82013.20.83612.40.84212.00.84811.70.85211.4U– Net++ (3D)0.82013.20.82812.70.84412.10.85011.70.85611.30.86111.0Attention U– Net (3D)0.82213.00.83012.50.84811.90.85311.50.86011.10.86510.8nnU– Net (3D)0.83811.60.84611.20.86010.50.86810.10.87810.00.8849.6Medical foundation & zero-shot segmentation (SAM family, 2024–2025)SAM– Med2D (ZS, prompt)0.79815.60.80615.00.81814.20.82213.80.82813.40.83413.0MedSAM (ZS, prompt)0.80415.10.81214.60.82413.80.82813.40.83413.00.84012.6MedSAM– Adapter (ZS + adapter)0.81614.20.82413.70.83613.00.84212.60.84812.20.85411.8SAM– Med3D (ZS, volume prompts)0.82013.90.82813.40.84212.60.84812.20.85411.90.86011.5DB– SAM (MICCAI’24, ZS)0.82413.60.83213.10.84612.40.85212.00.85811.70.86411.3FastSAM3D (ZS, 3D prompts)0.82613.40.83412.90.84812.20.85411.80.86011.50.86611.2Few-shot brain tumor segmentation (2024–2025)DASA (ACCV’24) few-shot0.83213.00.84012.50.85211.80.85811.40.86411.10.87010.8MetaMed3D (2025) few- shot0.83412.80.84212.30.85411.60.86011.20.86610.90.87210.6ProtoFSS (2025) few- shot0.83612.70.84412.20.85611.50.86211.10.86810.80.87410.5Weakly supervised tumor segmentation (2024–2025)Box-sup. + MedSAM (weak)0.82813.50.83613.00.84812.30.85411.90.86011.60.86611.3Point– sup. w/ box priors (weak)0.82413.80.83213.30.84412.60.85012.20.85611.90.86211.6Scribble– prompt (weak + interactive)0.82613.60.83413.10.84612.40.85212.00.85811.70.86411.4UM– CAM + seed refine (weak)0.83013.40.83812.90.85012.20.85611.80.86211.50.86811.2Recent fully supervised SOTA (2024–2025 transformers/diffusion)SwinUNETR (MONAI’24 cfg)0.84211.90.85011.50.86210.70.87010.30.8769.90.8829.6PA– SwinUNETR (2025)0.84811.60.85611.20.86810.50.87610.10.8829.80.8889.4DMCIE (2024)0.85011.40.85811.00.87010.20.8789.80.9006.50.8908.9FCFDiff– Net (2025, diffusion)0.85211.20.86010.80.87210.10.8809.70.9026.20.8928.6ACS– ResNet (2025)0.84411.80.85211.30.86410.60.87210.20.8789.90.8849.5Our Method**DARE-FUSE**0.864**10.6**0.872**10.2**0.884**9.5**0.892**9.1**0.910**6.0**0.8988.3Columns report, for each dataset, Dice (ET, ↑) and HD95 in mm (↓). BraTS-2015 and BraTS-2016 are external test sets (no training on them); BraTS-2018/2019/2020/2021 use the official held-out test sets. All metrics are reported as mean ± one standard deviation over three independent runs with different random seeds.Notes. ET = enhancing tumor. ZS = zero-shot. Weak supervision rows use image/point/box/scribble-level labels and CAM-derived seeds; few-shot rows assume k-shot per class with identical splits across methods. BraTS-2015/2016 are evaluated strictly as external test sets (no training exposure); other columns use official held-out test sets. Metrics: Dice/HD95 in mm.
On TC across BraTS-2015 to BraTS-2021 (Table 5), the method attains the top Dice and lower HD95 on every benchmark relative to recent supervised transformers and diffusion models, as well as zero-shot, few-shot, and weakly supervised families. Relative to the strongest supervised references, Dice improves from 0.866 to 0.878 with HD95 reduced from 9.8 to 9.3 on BraTS-2015, from 0.874 to 0.886 with HD95 from 9.4 to 8.9 on BraTS-2016, from 0.886 to 0.898 with HD95 from 8.7 to 8.1 on BraTS-2018, and from 0.894 to 0.906 with HD95 from 8.3 to 7.7 on BraTS-2019; on BraTS-2020 Dice rises from 0.916 to 0.924 with HD95 from 5.8 to 5.6, and on BraTS-2021 from 0.906 to 0.912 with HD95 from 8.1 to 7.9. Classic U-Net family baselines span Dice 0.834 to 0.902 with HD95 12.9 to 8.7 and lag more on early editions. Medical zero-shot variants based on SAM and MedSAM remain below supervised counterparts with Dice 0.812 to 0.878 and HD95 14.8 to 10.5. Few-shot approaches narrow the gap to Dice 0.846 to 0.888 with HD95 12.3 to 9.8, and weakly supervised methods reach 0.838 to 0.882 with HD95 13.1 to 10.5, yet both families consistently trail the proposed framework across all editions.Table 5Brain tumor MRI segmentation (TC): comparison across six BraTS benchmarksMethodBraTS-2015BraTS-2016BraTS-2018BraTS-2019BraTS-2020BraTS-2021Dice *↑*HD95 *↓*Dice *↑*HD95 ↓Dice ↑HD95 ↓Dice ↑HD95 ↓Dice ↑HD95 ↓Dice ↑HD95 ↓Classic U-Net family (reference baselines)U-Net (3D)0.83412.90.84212.40.85811.60.86411.20.87010.90.87610.6U-Net++ (3D)0.84012.40.84811.90.86411.30.87010.90.87610.50.88210.2Attention U-Net (3D)0.84212.20.85011.70.86811.10.87410.70.88010.30.88610.0nnU-Net (3D)0.85610.80.86410.40.8789.70.8869.30.8969.10.9028.7Medical foundation & zero-shot segmentation (SAM family, 2024–2025)SAM- Med2D (ZS, prompt)0.81214.80.82014.20.83213.40.83613.00.84212.60.84812.2MedSAM (ZS, prompt)0.81814.30.82613.80.83813.10.84212.70.84812.30.85411.9MedSAM- Adapter (ZS + adapter)0.83013.50.83813.00.85012.30.85611.90.86211.50.86811.1SAM- Med3D (ZS, volume prompts)0.83413.20.84212.70.85412.00.86011.60.86611.30.87210.9DB-SAM (MICCAI’24, ZS)0.83812.90.84612.40.85811.70.86411.30.87011.00.87610.6FastSAM3D (ZS, 3D prompts)0.84012.70.84812.20.86011.50.86611.10.87210.80.87810.5Few-shot brain tumor segmentation (2024–2025)DASA (ACCV’24) few-shot0.84612.30.85411.80.86611.10.87210.70.87810.40.88410.1MetaMed3D (2025) few- shot0.84812.10.85611.60.86810.90.87410.50.88010.20.8869.9ProtoFSS (2025) few- shot0.85012.00.85811.50.87010.80.87610.40.88210.10.8889.8Weakly supervised tumor segmentation (2024–2025)Box– sup. + MedSAM (weak)0.84212.80.85012.30.86211.60.86811.20.87410.90.88010.6Point– sup. w/box priors (weak)0.83813.10.84612.60.85811.90.86411.50.87011.20.87610.9Scribble- prompt (weak + interactive)0.84012.90.84812.40.86011.70.86611.30.87211.00.87810.7UM-CAM + seed refine (weak)0.84412.70.85212.20.86411.50.87011.10.87610.80.88210.5Recent fully supervised SOTA (2024–2025 transformers/diffusion)SwinUNETR (MONAI’24 cfg)0.85610.50.86410.10.8769.30.8848.90.8908.60.8968.3PA- SwinUNETR (2025)0.86210.20.8709.80.8829.10.8908.70.8968.40.9028.1DMCIE (2024)0.86410.00.8729.60.8848.80.8928.40.9146.10.9048.4FCFDiff-Net (2025, diffusion)0.8669.80.8749.40.8868.70.8948.30.9165.80.9068.1ACS ResNet (2025)0.85810.40.86610.00.8789.20.8868.80.8928.50.8988.2Ours (segmentation U– SEG + DAR + FUSE + GLRP; CPG provides CAM seeds, no fine– tuning on external)Ours (U- SEG + DAR + FUSE + GLRP)0.8789.30.886**8.9**0.898**8.1**0.906**7.7**0.924**5.6**0.9127.9Columns report, for each dataset, Dice (TC, ↑) and HD95 in mm (↓). BraTS-2015 and BraTS-2016 are external test sets (no training on them); BraTS-2018/2019/2020/2021 use the official held-out test sets. All metrics are reported as mean ± one standard deviation over three independent runs with different random seeds.Notes. TC = tumor core. ZS = zero-shot. Weak supervision rows use image/point/box/scribble-level labels and CAM-derived seeds; few-shot rows assume k-shot per class with identical splits across methods. BraTS-2015/2016 are evaluated strictly as external test sets (no training exposure); other columns use official held-out test sets. Metrics: Dice/HD95 in mm.
Label scaling on BraTS 2015 (Table 6) shows that the proposed framework maintains the strongest whole tumor performance across all budgets, from Dice 0.896 with HD95 9.6 at 100% pixel masks to Dice 0.858 with HD95 12.5 at 10%^40^. Semi supervised baselines are competitive at full supervision (AdaMix MT 0.890/9.5, MC Net+ 0.889/9.4, CPS 0.888/9.5) but degrade more steeply as labels shrink, reaching Dice 0.845-0.850 and HD95 12.8-13.0 at 10%^67–73^. Weakly supervised lines remain lower throughout (for instance, UM CAM plus seed refine moves from 0.860/12.0 at 100% to 0.826/13.8 at 10%, and BoxSup with MedSAM from 0.858/12.1 to 0.824/13.9), while few-shot variants cluster around 0.850-0.854 at 100% and 0.836-0.842 at 10%^27,35,57,59,74–76^. The zero-shot foundation baseline is largely invariant by design (MedSAM 0.848/13.7 across budgets)^27^. The advantage of the proposed method widens under reduced supervision, with gaps such as +0.009 Dice over MC Net+ at 30% (0.881 versus 0.872) and +0.012 over CPS at 10% (0.858 versus 0.846), consistent with converting image-level evidence into continuous priors via CPG, GLRP, and FUSE, and stabilizing the shared embedding with DAR.Table 6Label-scaling on BraTS-2015 (WT): semi/weak/few-shot segmentation under decreasing pixel supervisionMethod100%60%30%20%10%DiceHD95DiceHD95DiceHD95DiceHD95DiceHD95Semi-supervised (labeled subset + unlabeled pool)UA-MT (uncertainty- aware mean teacher)0.8869.60.87810.30.86811.10.85612.00.84213.1CPS (cross pseudo supervision)0.8889.50.87910.20.87011.00.85811.90.84612.9MC-Net+ (mutual consistency)0.8899.40.88010.10.87210.90.86011.80.85012.8DTC (dual-task consistency)0.8849.90.87510.60.86411.50.85212.40.84413.3FixMatch-Med0.88210.10.87410.80.86411.70.85212.60.83813.5PE– MT (perturbation- enhanced MT)0.8879.60.87810.30.86811.20.85512.10.84013.2AdaMix-MT (adaptive mix for SSL)0.8909.50.88110.20.87111.10.85912.00.84513.0MeanTeacher- EMA0.88010.00.87110.70.86011.60.84812.50.83413.6Weakly– supervised (image/point/box/scribble)BoxSup + MedSAM0.85812.10.85012.40.84012.70.83213.10.82413.9PointSup w/box priors0.85412.40.84612.70.83613.00.82813.30.82014.1ScribbleVS (scribble– supervised)0.85612.20.84812.50.83812.80.83013.20.82214.0UM– CAM + seed refine0.86012.00.85212.30.84212.60.83413.00.82613.8MIL-Trans (weak MIL)0.85212.60.84412.90.83413.20.82613.60.81814.3From SAM to CAMs (weak)0.86211.90.85412.20.84412.50.83612.90.82813.7Few-shot segmentation (k-shot regimes)DASA (few-shot)0.85012.30.84612.50.84212.80.84013.00.83813.2MetaMed3D (few-shot)0.85212.10.84812.30.84412.60.84212.90.84013.1ProtoFSS (few- shot)0.85412.00.85012.20.84612.50.84412.80.84213.0AMP-Med (few- shot)0.84812.40.84412.60.84012.90.83813.10.83613.3Zero-shot foundation (no pixel masks; prompt)MedSAM (ZS, prompt)0.84813.70.84813.70.84813.70.84813.70.84813.7Our MethodDARE-FUSE0.8969.60.88910.20.88110.90.87211.70.858****12.5Columns report, for each labeled budget, Dice (↑) and HD95 in mm (↓). Unlabeled cases provide image-level labels only and are consumed as weak evidence. All metrics are reported as mean ± one standard deviation over three independent runs with different random seeds.WT denotes whole tumor. Label budgets are case-level fractions of the official BraTS-2015 training set with pixel masks; remaining cases provide image-level labels only and are leveraged as weak evidence. Zero-shot does not use pixel masks and is invariant across columns by design.
Table 7 presents a unified analysis on BraTS-2015 for segmentation and on Bhuvaji for classification. The baseline encoder-decoder with a plain global classifier attains Dice 0.860 and IoU 0.755, while the classifier reaches AUC 0.960 and F1 at Youden 0.928. Introducing DAR alone raises Dice to 0.872 and IoU to 0.772 alongside AUC 0.970 and F1 0.936, indicating that task-oriented alignment improves both pixel- and image-level discrimination by reducing negative transfer between tasks. Adding SEGU further lifts segmentation to Dice 0.885 and IoU 0.790 with negligible change in classification, consistent with SEGU’s boundary-aware decoding and uncertainty calibration being primarily beneficial to mask quality. Incorporating CPG increases classification to AUC 0.982 and F1 0.948 with minimal shifts in segmentation, aligning with CPG’s role as a discriminative head and evidence producer rather than a direct mask refiner. Enabling FUSE yields the first large segmentation jump to Dice 0.898 and IoU 0.811 and modest classification gains to AUC 0.986 and F1 0.952, reflecting that uncertainty-weighted fusion of CAMs and structural cues stabilizes localization and reduces background confounds. Finally, activating GLRP delivers the best results with Dice 0.905 and IoU 0.822 together with AUC 0.991 and F1 0.958. Relative to the baseline, the full model improves Dice by 0.045 and IoU by 0.067 and improves AUC by 0.031 and F1 by 0.030. These trends are consistent with the mechanisms of the DAR aligns the shared embedding into separable subspaces for the two tasks, SEGU converts multi-scale evidence into calibrated masks, CPG strengthens global discrimination and supplies class consistent cues, FUSE turns heterogeneous signals into a continuous prior that guides decoding under weak supervision, and GLRP contributes a generative difference prior that regularizes false positives through empty mask consistency. Visualized results are shown in Fig. 3.Fig. 3Qualitative comparison of brain tumor segmentation across methods.Each row shows a representative case; columns correspond to Original, GT, MedSAM, CPS, DTC, MIL-Trans, AMP, and DARE-FUSE (ours). Classical semi and weakly supervised baselines (CPS, DTC) often produce over or under segmentation near low contrast boundaries and perilesional edema. Prompt driven MedSAM provides broader coverage but tends to smooth irregular, lobulated margins and may leak into adjacent parenchyma. Transformer-based MIL variants (MIL-Trans) occasionally miss small satellite components and thin protrusions. The adaptive prior module in AMP improves coverage yet still truncates fine edges in ambiguous regions. In contrast, DARE-FUSE yields contours that closely adhere to the ground truth across diverse appearances, preserving subtle spiculation while avoiding spillover into ventricles or CSF spaces.Table 7Unified ablation on BraTS-2015 (segmentation) and Bhuvaji Kaggle MRI (classification)DARSEGUCPGFUSEGLRPSegmentation (BraTS-2015)Classification (Bhuvaji Kaggle)Dice (WT) ↑IoU (WT) ↑AUC ↑F1@Youden ↑×××××0.860 ± 0.0040.755 ± 0.0050.960 ± 0.0030.928 ± 0.004✓××××0.872 ± 0.0040.772 ± 0.0050.970 ± 0.0030.936 ± 0.004✓**✓×××0.885 ± 0.0030.790 ± 0.0040.971 ± 0.0030.937 ± 0.004✓✓✓××0.886 ± 0.0030.792 ± 0.0040.982 ± 0.0020.948 ± 0.003*✓✓✓✓×0.891 ± 0.0030.811 ± 0.0040.986 ± 0.0020.952 ± 0.003✓✓✓✓✓0.896 ± **0.0020.818** ± **0.0030.992 ± **0.001*0.960 ± 0.003Higher is better for Dice/IoU/AUC/F1@Youden. Mean ± std over 3 runs. A vertical bar separates task-specific metrics.Modules: DAR = Domain-Alignment Refiner; SEGU = Segmentation Head with Uncertainty; CPG = Classification and Pseudo mask Generator; FUSE = Fusion based Uncertainty-weighted Seed Estimator; GLRP = Generative Lesion Removal Prior.**Task ** SEGU/FUSE/GLRP chiefly benefit segmentation (Dice/IoU). CPG/DAR mainly benefit classification (AUC/F1@Youden). DAR also yields minor segmentation gains via alignment.Notes: WT = Whole Tumor. F1@Youden computed at the Youden-J maximizing threshold. Baseline row uses vanilla encoder-decoder and a plain global classifier (no proposed modules).
The proposed DARE-FUSE framework yielded significantly improved performance in both tumor segmentation and classification tasks. Across six multi-center BraTS benchmarks, DARE-FUSE achieved consistently higher WT Dice scores and lower HD95 distances than recent SOTA models, indicating more accurate and compact tumor delineations (Table 3). Notably, our method ranked first on each dataset, with Dice improvements on the order of 0.5–1.5 percentage points over the strongest transformer- and diffusion-based baselines, and HD95 reductions of about 0.5–1 mm, demonstrating clear advantages in boundary precision. These gains were most pronounced on cases with complex, infiltrative tumor margins and on small-volume lesions, suggesting that DARE-FUSE better captures fine-grained boundary details that other methods often miss. Specifically, our error analysis reveals that the proposed framework yields the most consistent gains in challenging scenarios characterized by small enhancing components, low-contrast boundaries, and extensive perilesional edema. In these cases, the fused priors effectively suppress scattered false positives in healthy tissue and assist in recovering thin or partially obscured tumor rims, whereas performance on large and clearly delineated lesions remains comparable to that of strong baselines. These observations align with the distribution of Dice and HD95 scores across the BraTS benchmarks reported in our results, where the improvements are most pronounced for subjects exhibiting complex boundary patterns and mixed intensity profiles. In addition, on two independent image classification cohorts, the DARE-FUSE classification branch attained top-tier performance, surpassing recently published lightweight ViT models and weakly supervised MIL classifiers by a notable margin (Table 2). These results validate that our unified approach can concurrently deliver robust pixel-level and image-level predictions, providing quantitative improvements over specialized single-task methods.
Comparison with existing DARE-FUSE not only outperforms conventional segmentation networks in accuracy, but also addresses their known shortcomings in challenging scenarios. Traditional U-shapeCNNs and ViT-based segmentation models achieve high Dice on standard benchmarks, yet tend to struggle with diffuse tumor boundaries and domain shifts. In our results, DARE-FUSE showed consistently lower HD95 and controlled over segmentation at edema–infiltration interfaces, indicating superior boundary sensitivity under complex conditions. In contrast to generalist zero-shot models like SAM, which decline on medical images, our method maintained top performance even on external test sets with no training exposure, confirming robust generalization across scanners and institutions. Moreover, unlike previous weakly supervised approaches relying only on classification cues, which often produce incomplete lesion localization, DARE-FUSE integrates both discriminative evidence from class activation and structural evidence to yield a more complete and precise tumor map. This led to markedly better small lesion detection and fewer false negatives than using CAM-based localization alone. On the classification side, pure image-level models lack pixel-level context and interpretability. By contrast, our classification branch benefits from the multi-scale features and evidential maps shared with the segmentation pipeline, resulting in higher AUC and F1 and inherently interpretable predictions. Importantly, DARE-FUSE avoids the task interference seen in naive multi-task training. Existing unified frameworks often suffer degraded segmentation when trained with additional classification data, whereas our domain-aligned design preserves segmentation fidelity even when learning from classification-only datasets. In summary, compared to the full spectrum of existing fully supervised methods, few or zero-shot, and weakly supervised DARE-FUSE demonstrates a more balanced and robust performance, thanks to its coordinated handling of both tasks and various data conditions.
Technical innovations and The performance gains can be attributed to DARE-FUSE’s novel architectural components, each addressing key challenges in tandem. First, the DAR effectively decouples the segmentation and classification representations while enforcing cross-domain consistency. This task-oriented alignment proved critical in stabilizing the shared encoder as evidenced by our ablation study, adding DAR alone boosted segmentation Dice and classification AUC notably by mitigating negative transfer. By harmonizing feature distributions across centers and training objectives, DAR enables the model to leverage additional weakly labeled data without sacrificing segmentation accuracy. Second, the evidence fusion closed loop provides a powerful mechanism to derive and utilize complementary evidence. The GLRP offers a structural cue by highlighting areas of change via healthy tissue reconstruction difference, directly targeting subtle tumor boundaries that might be overlooked. Meanwhile, the FUSE module combines this with discriminative Grad-CAM cues, via a principled intersection and union strategy under uncertainty weighting. This yields a continuous pseudo segmentation prior that is both sensitive and the intersection step ensures high confidence consensus regions, while the union with distance-based attenuation expands coverage in areas where one source is weak. The integrated uncertainty weighting further modulates this prior, down-weighting unreliable regions so that spurious evidence does not corrupt the segmentation. Consequently, our ablations show that enabling FUSE and GLRP substantially lowered over-segmentation errors and tightened the HD95 by providing the decoder with a well-calibrated foreground seed. Third, the uncertainty-aware decoder improves result reliability and model interpretability. By modeling epistemic uncertainty at the voxel level, SEGU explicitly identifies regions of low confidence, which correspond to indistinct tumor boundaries or atypical appearances. This helped reduce variance in predictions and flagged high-risk areas for clinical caution. Additionally, the uncertainty map is fed back to suppress low-confidence contributions in the loss and during evidence fusion, forming a self-reinforcing loop that lessens the impact of noisy predictions. The net effect is a segmentation output that is not only more accurate but also accompanied by an uncertainty mask delineating where the prediction is less certain, an important feature for clinical validation. Finally, the unified yet modular design of DARE-FUSE provides practical advantages in training and deployment. All components interface through consistent, resolution-aligned tensor formats and single-channel priors, avoiding intricate tuning between tasks. This modularity allowed us to train on heterogeneous data in stages without one task destabilizing the other, and it facilitates easy removal or replacement of modules. The use of lightweight ViT backbones for both encoders further ensures that the added modules do not incur prohibitive computational cost. In our implementation, the entire pipeline runs on a single GPU with reasonable memory and time, confirming the efficiency of the design. Together, these innovations work synergistically to overcome the bottlenecks of small sample learning and weak supervision, as reflected in the improved quantitative metrics and qualitative stability of the outputs.
Clinical The improvements achieved by DARE-FUSE directly address critical needs in clinical neuro oncology imaging. Precise and reliable tumor boundary delineation is essential for surgical planning, radiotherapy target definition, and disease monitoring. In our results, DARE-FUSE not only increased segmentation accuracy but also provided an uncertainty map highlighting boundary regions with low confidence. This is particularly valuable in regions of uncertainty often correspond to infiltrative tumor margins or ambiguous edema tumor interfaces that even experts find challenging. By flagging these areas, the model can alert radiologists or surgeons to interpret those regions with caution or consider additional diagnostic work up, thereby improving patient safety. In a typical radiotherapy contouring or quality-control scenario, the model would therefore operate in an assistive given a planning MRI, DARE-FUSE produces a whole-tumor mask, a normalized uncertainty map, and the ignore mask Mignore in the same image space. The mask is imported into the treatment-planning system as a candidate RT structure, while Mignore is visualized as an overlay highlighting boundary segments with elevated uncertainty. A simple rule-based policy can then map uncertainty to reviewer actions, for example, accepting low-uncertainty regions as a starting contour, requiring mandatory manual review of voxels marked by Mignore, and flagging cases with a large global fraction of high-uncertainty voxels for additional expert adjudication rather than automatic propagation. The high fidelity of segmentation also enables more dependable volumetric measurements of tumor burden. Such quantitative assessments can be used to track tumor response to treatment over time or to stratify patients by risk, ultimately supporting more informed clinical decision making. Equally important is the integration of image-level classification within the same framework. DARE-FUSE can simultaneously provide a rapid preliminary tumor identification and classification during image review. This dual capability could streamline for instance, in a pre-surgical setting, the model’s classification of tumor type might complement biopsy and molecular data, aiding in comprehensive diagnosis. In triage and screening scenarios, the ability to detect and localize tumors in one pass could help catch abnormalities earlier or prioritize urgent cases for review. Moreover, the model’s demonstrated robustness across multi-center MRI datasets indicates strong potential for deployment in diverse clinical environments. Domain alignment in DAR means the system is less likely to fail when faced with scans from a new hospital or a different scanner protocol, which is crucial for real-world adoption. The entire pipeline is designed to be lightweight and automated, making it feasible to integrate into existing radiology systems as a decision support tool. Ultimately, by unifying segmentation, quantification, and diagnosis into one AI system, DARE-FUSE can contribute to more efficient, reproducible, and interpretable brain tumor imaging assessments in clinical practice.
Limitations and future Despite its advantages, this study has several limitations that suggest avenues for future work.
A first limitation is case although the framework improves robustness across cohorts, performance still degrades for very small or highly diffuse lesions, for scans with severe motion or susceptibility artifacts, and for atypical enhancement patterns that deviate from the dominant training distribution. In such settings, the generative lesion-removal prior may either over-smooth subtle tumor components or inpaint through non-tumor abnormalities in a way that hides relevant structure, and classification-driven attention can be attracted by non-tumor findings that resemble tumor on routine MRI. Moreover, we currently do not evaluate the lesion-removal prior against a ground-truth tumor-free reference, since such annotations are absent for public datasets, and a dedicated study with paired or synthetic controls will be needed to fully quantify the behavior of this component. A second limitation is the interpretability of failure while the uncertainty map and ignore mask expose some of these effects, they do not guarantee that misleading generative edits or attention hotspots will always be flagged. Beyond uncertainty maps and ignore masks, we also do not include two-dimensional projections of intermediate features with tools such as t-SNE or UMAP that could visualize how DAR aligns domains, and we plan to investigate such feature-space analyses as part of future interpretability studies. Our evaluation is retrospective and based on public glioma cohorts and representative but not exhaustive Kaggle datasets. The current framework was developed and evaluated on retrospective public datasets, which may not capture the full heterogeneity of real-world clinical populations. These benchmarks remain relatively curated compared with routine clinical archives, so the gains observed here need to be tested prospectively on more heterogeneous and noisier hospital data before the framework can be considered for broad deployment. These limitations underline that DARE-FUSE should be viewed as an assistive system whose outputs require expert review, and motivate future work on explicit modeling of non-tumor mimics, more artifact-robust priors, and prospective, site-specific validation. The current framework was developed and evaluated on retrospective public datasets, which may not capture the full heterogeneity of real-world clinical populations. Although we included multi-center data and external test sets to mimic deployment scenarios, a prospective clinical validation will be important to assess performance in routine practice and on rarer tumor entities. Data diversity remains a for example, our classification experiments used open-source MRI collections that, while large, might contain biases and often lack consistent clinical labels. Applying DARE-FUSE to a broader spectrum of brain lesions and imaging modalities would test its generality. Second, the current model targets a binary segmentation of the whole tumor and a single category image label; it does not explicitly differentiate intra-tumoral substructures during inference. An extension to finer pathological sub-regions is warranted. In the context of gliomas, this could mean jointly segmenting edema vs. enhancing core vs. necrosis in a fully multi-class manner, or even identifying tissue sub-compartments linked to distinct pathological or molecular features. Our framework’s modular design should, in principle, accommodate multi-class outputs, but this remains to be demonstrated. Third, while the inclusion of uncertainty estimation is a strength, we currently rely on Monte Carlo dropout at test time, which can be computationally intensive and may not capture all types of uncertainty. Future research could explore more efficient uncertainty quantification to further improve speed and reliability. Another important extension will be incorporating the temporal dimension. Longitudinal analysis could allow the model to detect subtle changes in tumor extent or appearance between follow-up scans by applying the evidence fusion concept across serial MRI exams. By establishing an evidence loop in time, the framework might better characterize tumor growth dynamics and treatment response, which are critical for clinical decision making in oncology. We also plan to investigate multi-sequence joint modeling beyond concurrent multimodal MRI. This could involve integrating information from different imaging sequences or time points directly within the model, rather than treating them independently, to enrich the evidence available for both segmentation and classification. Finally, the complexity of the DARE-FUSE pipeline, with several components, is a double-edged while each module contributes to performance, it also introduces additional hyperparameters and potential failure points. A further limitation is that we keep the loss design fixed across experiments and do not explore all possible decompositions of individual objectives, and a more fine-grained analysis of the loss components may be useful in future work for simplifying the training objective while preserving the robustness properties of the framework. Simplifying the framework or distilling its knowledge into a more compact model could be beneficial for real-world deployment. In summary, addressing these limitations through larger-scale validation, extension to finer and temporal analysis, and further optimization will be key steps in translating DARE-FUSE from a research prototype to a clinically reliable tool.
Domain-Aligned Representation with Evidence-guided FUSE (DARE-FUSE) targets unified modeling of brain tumor classification and segmentation under few-shot and weak supervision. The overall system consists of five fixed modules that collaborate within a single pipeline as shown in Fig. 1. Dual encoders and a Feature-interaction bridge produce a unified multi-scale representation, Domain-Alignment Refiner (DAR) performs medical-domain alignment and outputs task-oriented features, Classification and Pseudo-mask Generator (CPG) provides image level decisions and multi-scale evidence heatmaps, Generative Lesion Removal Prior (GLRP) constructs a differential prior, Fusion based Uncertainty-weighted Seed Estimator (FUSE) fuses multi-source evidence into a continuous prior, and U-shape Decoder and Segmentation Head with Uncertainty (U-SEG with SEGU) completes pixel level prediction and uncertainty estimation. The entire pipeline achieves collaborative classification and segmentation in a single forward pass while preserving plug-in capability and interface consistency across modules.
To address the challenge of learning representations that are simultaneously boundary sensitive for segmentation and globally discriminative for classification under few-shot and weak supervision, while maintaining stability under cross-center and cross-year distribution shift, we construct a dual-view encoder subsystem using lightweight ViT backbones. It is paired with a cross-layer Feature-interaction bridge that performs bidirectional cross-attention fusion, and a Domain-Alignment Refiner (DAR) that applies statistical and adversarial rectification. Concretely, DAR operates on the shared encoder and the task-agnostic bottleneck in a deterministic and layer-wise manner. For each mini-batch, we sample volumes from up to two domains and extract feature maps from the third and fourth encoder stages, as well as a global pooled bottleneck representation. The first- and higher-order alignment losses are applied on these third- and fourth-stage features after a 1 × 1 projection that brings them to a common channel dimension, while the adversarial discriminator receives the bottleneck features passed through a gradient-reversal layer. The discriminator predicts a categorical domain label, and its gradients are used to encourage domain-invariant representations at the bottleneck, whereas the alignment losses explicitly reduce pairwise discrepancies at the two selected encoder scales. In a single forward pass, the subsystem produces multi-scale shared embeddings \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{E}_{k}^{* }}
\usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \{{E}_{k}^{seg}\} $$\end{document}{Ekseg} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \{{E}_{k}^{cls}\} $$\end{document}{Ekcls} for consumption by U-SEG with SEGU and CPG, respectively. In practice, DAR operates on the last two encoder scales and the shared first- and higher-order alignment losses are applied to rectified feature maps at scales *k* ∈ {3, 4}, while the adversarial discriminator consumes the global pooled bottleneck embedding passed through a gradient-reversal layer. This restricts domain regularization to high-level representations that are shared across tasks, leaving shallow feature extraction unchanged. To retain long-range dependencies of self-attention while ensuring efficient inference, the segmentation view encoder (Seg Enc) adopts a lightweight ViT backbone at the TinyViT 21M level with a shallow multi-scale neck to produce a feature pyramid^77^. The classification view encoder (Cls Enc) adopts a lightweight ViT backbone at the DeiT III Small (ViT S-16) level to emphasize globally discriminative representations^78^. Both backbones apply patch embedding to split *X* into a sequence of length *N*, add learnable positional encodings, and stack standard Transformer blocks consisting of Multi Head Self Attention (MSA) and a feed forward Multi Layer Perceptron (MLP) with GELU activation and Layer Normalization (LN). For layer *ℓ*,1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {z}^{(0)}=[{x}_{{\mathrm{cls}}};{x}_{1};\ldots ;{x}_{N}]+{E}_{{\mathrm{pos}}}. $$\end{document}z(0)=[xcls;x1;…;xN]+Epos.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {z}^{(\ell )}={\mathrm{MSA}}({\mathrm{LN}}({z}^{(\ell -1)}))+{z}^{(\ell -1)}. $$\end{document}z(ℓ)=MSA(LN(z(ℓ-1)))+z(ℓ-1).3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {y}^{(\ell )}={\mathrm{MLP}}({\mathrm{LN}}({z}^{(\ell )}))+{z}^{(\ell )}. $$\end{document}y(ℓ)=MLP(LN(z(ℓ)))+z(ℓ). Because segmentation requires fine grained spatial cues whereas classification requires discriminative global semantics, the Feature-interaction bridge establishes complementarity via bidirectional cross-attention at each scale *k*. Let the projection-aligned features from Seg-Enc and Cls-Enc be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\widetilde{F}}_{k},{\widetilde{G}}_{k}\in {{\mathbb{R}}}^{{h}_{k}{w}_{k}\times d} $$\end{document}F~k,G~k∈Rhkwk×d. As shown in Fig. 4a, we define queries, keys, and values 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {Q}_{k}^{(s)}={\widetilde{F}}_{k}{W}_{Q}^{(s)},\,{K}_{k}^{(s)}={\widetilde{G}}_{k}{W}_{K}^{(s)},\,{V}_{k}^{(s)}={\widetilde{G}}_{k}{W}_{V}^{(s)}, $$\end{document}Qk(s)=F~kWQ(s),Kk(s)=G~kWK(s),Vk(s)=G~kWV(s),5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {Q}_{k}^{(c)}={\widetilde{G}}_{k}{W}_{Q}^{(c)},\,{K}_{k}^{(c)}={\widetilde{F}}_{k}{W}_{K}^{(c)},\,{V}_{k}^{(c)}={\widetilde{F}}_{k}{W}_{V}^{(c)}. $$\end{document}Qk(c)=G~kWQ(c),Kk(c)=F~kWK(c),Vk(c)=F~kWV(c).Cross-view information injection via scaled dot-product attention 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {Z}_{k}^{(s\leftarrow c)}={\mathrm{softmax}}\left(\frac{{Q}_{k}^{(s)}{({K}_{k}^{(c)})}^{\top }}{\sqrt{d}}\right){V}_{k}^{(c)},\,{Z}_{k}^{(c\leftarrow s)}={\mathrm{softmax}}\left(\frac{{Q}_{k}^{(c)}{({K}_{k}^{(s)})}^{\top }}{\sqrt{d}}\right){V}_{k}^{(s)}. $$\end{document}Zk(s←c)=softmaxQk(s)(Kk(c))⊤dVk(c),Zk(c←s)=softmaxQk(c)(Kk(s))⊤dVk(s).To preserve each view’s prior advantages during fusion and suppress noise propagation, we apply learnable gating and residual concatenation to obtain the shared 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {E}_{k}^{* }=\phi \left([{\widetilde{F}}_{k}+{\gamma }_{k}^{(s)}{Z}_{k}^{(s\leftarrow c)}\,\parallel \,{\widetilde{G}}_{k}+{\gamma }_{k}^{(c)}{Z}_{k}^{(c\leftarrow s)}]{W}_{{\mathrm{fuse}}}\right), $$\end{document}Ek*=ϕ[F~k+γk(s)Zk(s←c)∥G~k+γk(c)Zk(c←s)]Wfuse,where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\gamma }_{k}^{(\cdot )} $$\end{document}γk(⋅) are scalar gates, [ ⋅ ∥ ⋅ ] denotes channel concatenation, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {W}_{fuse}\in {{\mathbb{R}}}^{2d\times d} $$\end{document}Wfuse∈R2d×d projects back to dimension *d*, and *ϕ* denotes LayerNorm followed by GELU. The bridge outputs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\{{E}_{k}^{* }\}}_{k=1}^{K} $$\end{document}{Ek*}k=1K across scales, ensuring that structural geometry from the segmentation view and category semantics from the classification view interact in a unified space. Attention and feed-forward definitions follow standard Transformer practice.Fig. 4Feature bridge design and uncertainty aware evidence fusion for gated prior construction.**a** Mechanism of the feature bridge based on the cross-attention architecture. **b** Grad-CAM evidence and structural difference are split by Core Gate (intersection) and Ext Gate (union with distance-based attenuation), while an uncertainty map applies pixel-wise down-weighting. The outputs are a continuous prior and an ignore mask for gated injection and weak supervision. To keep the shared embeddings usable, separable, and stable for downstream tasks with different requirements, the Domain-Alignment Refiner (DAR) performs task-oriented statistical and adversarial alignment, then produces \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \{{E}_{k}^{seg}\} $$\end{document}{Ekseg} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \{{E}_{k}^{cls}\} $$\end{document}{Ekcls} via two lightweight projection heads serving as the sole upstream interfaces for segmentation and 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {E}_{k}^{seg}={\psi }_{{\mathrm{seg}}}({T}_{{\mathrm{DAR}}}({E}_{k}^{* })),\,{E}_{k}^{cls}={\psi }_{{\mathrm{cls}}}({T}_{{\mathrm{DAR}}}({E}_{k}^{* })), $$\end{document}Ekseg=ψseg(TDAR(Ek*)),Ekcls=ψcls(TDAR(Ek*)),where *T*~DAR~ is a composite rectifier and *ψ*~seg~, *ψ*~cls~ are linear-normalization-activation projections aligned in dimension and resolution with the downstream tasks. First-order statistical rectification re-centers channel-wise means and variances via learnable normalization so that shared representations at the same scale converge to task-specific target statistics. We adopt Adaptive Instance Normalization (AdaIN):9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\mathrm{AdaIN}}(x;{\mu }_{t},{\sigma }_{t})={\sigma }_{t}\cdot \frac{x-\mu (x)}{\sigma (x)}+{\mu }_{t}, $$\end{document}AdaIN(x;μt,σt)=σt⋅x-μ(x)σ(x)+μt,where (*μ*~*t*~, *σ*~*t*~) are predicted by an affine layer to align channel statistics with task-side targets (classification and segmentation differ). We also mix Instance and Batch normalization at shallow layers to balance appearance invariance and discrimination, and employ Switchable Whitening when stronger decorrelation is needed. For distribution-level alignment, DAR further compresses statistical discrepancies to covariance and kernel mean embedding (MMD) levels. Second-order alignment uses the Deep CORAL energy to match covariances between source representations and task-side 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{E}}}_{\mathrm{CORAL}}({E}_{k,s}^{* },{E}_{k,t}^{* })=\frac{1}{4{d}^{2}}\parallel C({E}_{k,s}^{* })-C({E}_{k,t}^{* }){\parallel }_{F}^{2}, $$\end{document}ECORAL(Ek,s*,Ek,t*)=14d2∥C(Ek,s*)-C(Ek,t*)∥F2,11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \,C(Z)=\frac{1}{n-1}\left({Z}^{\top }Z-\frac{1}{n}{({{\bf{1}}}^{\top }Z)}^{\top }({{\bf{1}}}^{\top }Z)\right). $$\end{document}C(Z)=1n-1Z⊤Z-1n(1⊤Z)⊤(1⊤Z).Higher-order alignment adopts the kernel Maximum Mean Discrepancy (MMD):12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \begin{array}{l}{\varepsilon }_{{\mathrm{MMD}}}({E}_{k,s}^{* },{E}_{k,t}^{* })=\frac{1}{{n}_{s}^{2}}\mathop{\sum }\limits_{i,{i}^{{\prime} }}k({e}_{k,i}^{(s)},{e}_{k,{i}^{{\prime} }}^{(s)})+\frac{1}{{n}_{t}^{2}}\mathop{\sum }\limits_{j,j{\prime} }k({e}_{k,j}^{(t)},{e}_{k,j{\prime} }^{(t)})\\ -\frac{2}{{n}_{s}{n}_{t}}\mathop{\sum }\limits_{i,j}k({e}_{k,i}^{(s)},{e}_{k,j}^{(t)}),\end{array} $$\end{document}εMMD(Ek,s*,Ek,t*)=1ns2∑i,i′k(ek,i(s),ek,i′(s))+1nt2∑j,j′k(ek,j(t),ek,j′(t))-2nsnt∑i,jk(ek,i(s),ek,j(t)),where *k*( ⋅ , ⋅ ) is a positive definite kernel and (*s*, *t*) correspond in our framework to the two sides of the mapping from shared embeddings to task-oriented representations. At the non-separability regularization level, DAR places a lightweight domain discriminator at each scale and applies adversarial constraints via a Gradient Reversal Layer (GRL), making the rectified shared features separable for task decisions yet non-separable to the discriminator. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\widetilde{E}}_{k} $$\end{document}E~k denote rectified features at scale *k* and *D*~*k*~ the discriminator. The objective minimizes the discriminator cross-entropy:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{L}}}_{{\mathrm{dom}}}^{(k)}=\frac{1}{n}\mathop{\sum }\limits_{i}{\mathrm{CE}}({D}_{k}({\widetilde{e}}_{k,i}),{d}_{i}), $$\end{document}Ldom(k)=1n∑iCE(Dk(e~k,i),di),while the feature extractor receives reversed gradients through GRL to promote non-separability. To suppress instability from low-confidence regions, we weight dense features by pixel level uncertainty from the segmentation side, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {w}_{k,i}=1-{\widehat{U}}_{k,i} $$\end{document}wk,i=1-U^k,i, and use weighted statistics and adversarial 14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{L}}}_{{\mathrm{dom}},{\mathrm{w}}}^{(k)}=\frac{1}{{\sum }_{i}{w}_{k,i}}\mathop{\sum }\limits_{i}{w}_{k,i}\,{\mathrm{CE}}({D}_{k}({\widetilde{e}}_{k,i}),{d}_{i}). $$\end{document}Ldom,w(k)=1∑iwk,i∑iwk,iCE(Dk(e~k,i),di).The same weights are applied to define weighted covariance and kernel means, reducing the influence of unreliable locations and preventing them from dominating the alignment direction. ### Evidence-guided image level classification To obtain stable image level discrimination under few-shot and weak supervision while providing localization evidence for downstream prior fusion, the classification branch is centered on the CPG, which directly consumes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\{{E}_{k}^{cls}\}}_{k=1}^{K} $$\end{document}{Ekcls}k=1K. As described in the previous section, to let the classification representation inherit structural geometry from the segmentation side and remain semantically complementary, the Feature-interaction bridge performs bidirectional cross-attention at each scale and then the Domain-Alignment Refiner (DAR) applies task-oriented statistical and adversarial rectification to output classification-side embeddings. Let the channel-aligned features of the segmentation and classification paths at scale *k* be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\widetilde{F}}_{k},{\widetilde{G}}_{k}\in {{\mathbb{R}}}^{{h}_{k}{w}_{k}\times d} $$\end{document}F~k,G~k∈Rhkwk×d. Scaled dot-product cross-attention is implemented as15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {Z}_{k}^{(c\leftarrow s)}={\mathrm{softmax}}\left(\frac{{{\rm{Q}}}_{{\rm{k}}}^{({\rm{c}})}{({{\rm{K}}}_{{\rm{k}}}^{({\rm{s}})})}^{{\rm{\top }}}}{\sqrt{{\rm{d}}}}\right){V}_{k}^{(s)},\,{Z}_{k}^{(s\leftarrow c)}={\mathrm{softmax}}\left(\frac{{{\rm{Q}}}_{{\rm{k}}}^{({\rm{s}})}{({{\rm{K}}}_{{\rm{k}}}^{({\rm{c}})})}^{{\rm{\top }}}}{\sqrt{{\rm{d}}}}\right){V}_{k}^{(c)}. $$\end{document}Zk(c←s)=softmaxQk(c)(Kk(s))⊤dVk(s),Zk(s←c)=softmaxQk(s)(Kk(c))⊤dVk(c).The shared embedding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {E}_{k}^{* } $$\end{document}Ek* is formed by gated residual fusion, and DAR uses the composite rectifier *T*~DAR~ with lightweight heads to produce the classification-side 16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {E}_{k}^{cls}={\psi }_{{\rm{c}}{\rm{l}}{\rm{s}}}({T}_{{\mathrm{DAR}}}({E}_{k}^{* })), $$\end{document}Ekcls=ψcls(TDAR(Ek*)),where *T*~DAR~ combines first order rectification (AdaIN, IBN and Whitening) and second or higher order alignment (Deep CORAL and MMD) with adversarial constraints via GRL, ensuring consistency between classification embeddings and task decision boundaries in the shared semantic space. To achieve stable image level prediction while retaining interpretable spatial evidence, CPG aggregates multi-scale representations and applies linear discrimination, then computes Grad-CAM++ on mid-to-high-level spatial activations. Let the aggregator be17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ g={\mathcal{A}}(\{{E}_{k}^{cls}\}), $$\end{document}g=A ({Ekcls}),and the logits and probabilities be18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ z={W}_{{\mathrm{cls}}}\,g+b,\,{p}_{{\mathrm{cls}}}={\mathrm{softmax}}(z)\in {{\rm{{\mathbb{R}}}}}^{C}. $$\end{document}z=Wclsg+b,pcls=softmax(z)∈ℝC.Let the selected activation tensor be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ A\in {{\mathbb{R}}}^{H\times W\times K} $$\end{document}A∈RH×W×K. For class *c*, Grad-CAM++ defines the channel weights and heatmap as19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {w}_{k}^{c}=\mathop{\sum }\limits_{i,j}{\alpha }_{k,ij}^{c}\,{\mathrm{ReLU}}\left(\frac{\partial {y}^{c}}{\partial {A}_{k,ij}}\right),\,{{\mathcal{L}}}^{c}={\mathrm{ReLU}}\left(\mathop{\sum }\limits_{k}{w}_{k}^{c}\,{A}_{k}\right), $$\end{document}wkc=∑i,jαk,ijcReLU∂yc∂Ak,ij,Lc=ReLU∑kwkcAk,and over a set of scales {*A*^(*s*)^}, we independently compute, upsample, and normalize to obtain20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\mathrm{CAM}}_{ms}=\{{\widehat{{\mathcal{L}}}}^{c,(s)}\}. $$\end{document}CAMms={L^c,(s)}.Grad-CAM++ offers more stable localization in multi-target and fine-structure scenarios and thus serves as a suitable source of weak supervisory evidence. To suppress spurious activations in difficult regions and to maintain a consistent uncertainty convention with the segmentation side, classification evidence is down-weighted at the pixel level before entering prior fusion. Given the segmentation-side uncertainty map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \widehat{U}\in {[0,1]}^{H\times W} $$\end{document}U^∈[0,1]H×W,21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\mathop{{\mathcal{L}}}\limits^{ \sim }}^{c,(s)}=(1-\widehat{U})\odot {\hat{{\mathcal{L}}}}^{c,(s)},\,{{\mathrm{CAM}}}_{ms}\leftarrow {\{{\mathop{{\mathcal{L}}}\limits^{ \sim }}^{c,(s)}\}}_{s}. $$\end{document}L~c,(s)=(1-U^)⊙L^c,(s),CAMms←{L~c,(s)}s.The down-weighted multi-scale heatmaps are used in two first, as interpretable image level evidence and visualization artifacts; second, as weak supervisory inputs for the evidence pathway of Fusion based Uncertainty-weighted Seed Estimator (FUSE) and Generative Lesion Removal Prior (GLRP). GLRP generates a tumor-free reconstruction *X*~no − tumor~ and a differential prior Δ = *X* − *X*~no − tumor~. FUSE fuses CAM~*m**s*~ and Δ over core and extended regions and outputs a continuous prior *P* ∈ [0, 1] and an ignore mask \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {M}_{\mathrm{ignore}} $$\end{document}Mignore. This pathway supplies only weak supervision and priors without altering the discrimination structure of CPG. The explicit outputs of the classification branch are *p*~cls~ and CAM~*m**s*~, strictly aligned in interface and resolution with the segmentation branch, thereby achieving consistent coordination between discrimination and evidence in a single forward pass. ### Uncertainty-aware segmentation with generative prior fusion To obtain pixel-level predictions that are sensitive to both boundaries and volumes under scarce annotations and weak supervision, this branch centers on U-SEG with SEGU. Its main input is the segmentation-side multi-scale representations \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\{{E}_{k}^{seg}\}}_{k=1}^{K} $$\end{document}{Ekseg}k=1K rectified by the Feature-interaction bridge and the Domain-Alignment Refiner (DAR). The decoder adopts a U-shape top-down pyramid (upsampling with lateral connections) to aggregate context and local structures as resolution increases, a design that preserves precise localization with limited annotations and is well established in medical imaging. The overall pipeline is shown in Fig. 4b. Let the top-down features at scale *k* be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {D}_{k}\in {{\mathbb{R}}}^{{h}_{k}\times {w}_{k}\times d} $$\end{document}Dk∈Rhk×wk×d. Define upsampling and lateral mappings as22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{U}}}_{2}(\cdot )={\mathrm{bilinear}}\_{\mathrm{up}}(\cdot ,\times 2),\,{{\mathcal{L}}}_{k}(\cdot )={{\mathrm{Conv}}}_{1\times 1}(\cdot ). $$\end{document}U2(⋅)=bilinear_up(⋅,×2),Lk(⋅)=Conv1×1(⋅).Then the multi-scale fusion proceeds as23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\mathop{D}\limits^{ \sim }}_{k}={{\mathcal{U}}}_{2}({D}_{k+1})+{{\mathcal{L}}}_{k}({E}_{k}^{seg}),\,k=K-1,\ldots ,1. $$\end{document}D~k=U2(Dk+1)+Lk(Ekseg),k=K-1,…,1.24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {D}_{k}={\mathrm{GELU}}({\rm{C}}{\rm{o}}{\rm{n}}{{\rm{v}}}_{3\times 3}({\mathop{D}\limits^{ \sim }}_{k})). $$\end{document}Dk=GELU(Conv3×3(D~k)). For conventional tumor subregion segmentation, we employ a 3-class softmax head to predict WT, ET and TC simultaneously. Denote the logits by *S* and the per-pixel class probabilities by *p*:25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ S={W}_{o}* {D}_{1}+{b}_{o},\,p={\mathrm{softmax}}(S)\in {{\rm{{\mathbb{R}}}}}^{H\times W\times 3}. $$\end{document}S=Wo*D1+bo,p=softmax(S)∈ℝH×W×3.For uncertainty estimation, Monte Carlo Dropout is enabled in the last two decoder layers. We aggregate foreground probability as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\widehat{Y}}_{mask}(x)={\sum }_{c\in \{WT,ET,TC\}}{p}_{c}(x) $$\end{document}Y^mask(x)=∑c∈{WT,ET,TC}pc(x) and perform *T* stochastic forward passes to obtain mean and 26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \mu (x)=\frac{1}{T}\mathop{\sum }\limits_{t=1}^{T}{\widehat{Y}}_{{\mathrm{mask}}}^{(t)}(x),\,{U}_{\mathrm{map}}(x)=\frac{1}{T-1}\mathop{\sum }\limits_{t=1}^{T}{({\widehat{Y}}_{\mathrm{mask}}^{{(}t{)}}(x)-\mu (x))}^{2}. $$\end{document}μ(x)=1T∑t=1TY^mask(t)(x),Umap(x)=1T-1∑t=1T(Y^mask(t)(x)-μ(x))2.The map *U*~map~ serves as a unified down-weighting signal across modules. To enhance foreground response and boundary sensitivity while supplying consumable priors under weak supervision, two external evidences are injected with lightweight gating. Generative Lesion Removal Prior (GLRP) reconstructs a tumor-free image and forms a differential prior. Fusion-based Uncertainty-weighted Seed Estimator (FUSE) fuses multi-scale class activation evidence CAM~*m**s*~ from CPG with Δ, applies uncertainty weighting \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {(1-\widehat{U})}^{\gamma } $$\end{document}(1-U^)γ, and outputs a continuous prior *P* ∈ [0, 1] and an ignore mask \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {M}_{{\mathrm{ignore}}} $$\end{document}Mignore. For scale-wise injection, the continuous prior *P* and the differential map Δ are injected at all decoder stages after the lateral skip fusion. To ensure strictly aligned feature integration, both maps are resized to the corresponding resolution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{R}}}_{k}(\cdot ) $$\end{document}Rk(⋅) and passed through 1 × 1 linear projections \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{A}}}_{k} $$\end{document}Ak and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{B}}}_{k} $$\end{document}Bk to align the channel dimension and dynamic range with the decoder features. We then modulate the decoder features via gated additive 27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {D}_{k}^{+}={D}_{k}\odot \sigma (1+{\alpha }_{k}\,{{\mathcal{A}}}_{k}({P}_{k})+{\beta }_{k}\,{{\mathcal{B}}}_{k}({\Delta }_{k})), $$\end{document}Dk+=Dk⊙σ (1+αkAk(Pk)+βkBk(Δk)),where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{A}}}_{k},{{\mathcal{B}}}_{k} $$\end{document}Ak,Bk are 1 × 1 linear mappings for channel and range alignment, *α*~*k*~, *β*~*k*~ are learnable gates, *σ* is the sigmoid, and ⊙ denotes the Hadamard product. This unified parameterization strengthens foreground and highlights boundary transitions at each scale. The training objective combines class-balanced segmentation losses, a narrow-band boundary term, and a prior-guided weak supervision 28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{L}}}_{seg}=\mathop{\sum }\limits_{c\in \{WT,ET,TC\}}{\lambda }_{c}({{\mathcal{L}}}_{Dice}^{(c)}+{{\mathcal{L}}}_{CE}^{(c)})+{\lambda }_{bd}\,{{\mathcal{L}}}_{boundary}+{\lambda }_{seed}\,{{\mathcal{L}}}_{seed}(P,\Delta ). $$\end{document}Lseg=∑c∈{WT,ET,TC}λc (LDice(c)+LCE(c))+λbdLboundary+λseedLseed(P,Δ).During training, *U*~map~ is used for loss masking and prior down-weighting; during inference, it supports risk exposure and post-processing. We adopt a 3-class softmax segmentation head and define \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\widehat{Y}}_{mask} $$\end{document}Y^mask as the aggregated foreground probability for uncertainty estimation. If any earlier single-channel mask description appears, it should be replaced by this multi-class formulation to keep the notation and the uncertainty pathway consistent across modules. To provide dense evidence that is in phase with structural changes and directly consumable by the decoder under scarce annotations and weak supervision, we construct the Generative Lesion Removal Prior (GLRP). The idea is to map the input MRI *X* to a tumor-free counterpart and take their difference to obtain a differential prior that is more sensitive to foreground and boundaries. The motivation is that the segmentation branch requires precise boundaries and volumetric cues, whereas class-activation-based discriminative maps are often insufficiently responsive to morphological details. Introducing a generative reconstruction-based structural difference yields stable geometric evidence that is decoupled from appearance variations and is suitable for subsequent prior fusion and boundary enhancement. Let the generator be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{G}}}_{\theta }:{{\mathbb{R}}}^{H\times W\times C}\to {{\mathbb{R}}}^{H\times W\times C} $$\end{document}Gθ:RH×W×C→RH×W×C, implemented with large-receptive-field Fourier and dilated convolutions and a skip-connected decoder for context-consistent high-resolution reconstruction. Given *X*, the tumor-free reconstruction and the differential prior are defined as29\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {X}_{{\mathrm{no}}-{\mathrm{tumor}}}={{\mathcal{G}}}_{\theta }(X),\,\,\,\,\,\,\,\,\,\,\,\Delta ={\mathcal{N}}(\,{||X-{X}_{{\mathrm{no}}-{\mathrm{tumor}}}||}_{1})\in [0,1], $$\end{document}Xno-tumor=Gθ(X),Δ=N(X-Xno-tumor1)∈[0,1],where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\mathcal{N}} $$\end{document}N denotes per-image linear normalization to [0, 1] (mean-variance normalization followed by clamping). The quantity Δ measures the minimal structural modification needed to explain the input while keeping the context consistent; it typically shows high responses within lesions and along boundaries, and low responses in background and normal tissues. For decoder-scale alignment, define30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\Delta }_{k}={{\mathcal{R}}}_{k}(\Delta ),\,{{\mathcal{R}}}_{k}:[H\times W]\to [{h}_{k}\times {w}_{k}], $$\end{document}Δk=Rk(Δ),Rk:[H×W]→[hk×wk],so that {Δ~*k*~} can be injected as gated structural enhancement at each level of the U-shape decoder (see the gating formula in the previous subsection). To make \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{G}}}_{\theta } $$\end{document}Gθ learn a mapping that removes lesions while preserving anatomical structure and background consistency, GLRP enforces reconstruction consistency jointly in the image, perceptual, and frequency domains. Let the random irregular mask be *M* ∈ {0, 1}^*H*×*W*^ with known region \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \overline{M}=1-M $$\end{document}M¯=1-M, and define the masked reconstruction \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \widetilde{X}={{\mathcal{G}}}_{\theta }(X\odot \overline{M}) $$\end{document}X~=Gθ(X⊙M¯). The reconstruction and consistency objectives are31\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{L}}}_{{\mathrm{rec}}}=\parallel \widetilde{X}-X{\parallel }_{1,M}+{\lambda }_{{\mathrm{id}}}\parallel \widetilde{X}-X{\parallel }_{1,\bar{M}}, $$\end{document}Lrec=∥X~-X∥1,M+λid∥X~-X∥1,M¯,32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{L}}}_{{\mathrm{perc}}}=\mathop{\sum }\limits_{\ell }\parallel {\phi }_{\ell }(\widetilde{X})-{\phi }_{\ell }(X){\parallel }_{1},\,{{\mathcal{L}}}_{\mathrm{ff}{\rm{t}}}=\parallel {\mathcal{F}}(\widetilde{X})-{\mathcal{F}}(X){\parallel }_{1}, $$\end{document}Lperc=∑ℓ∥ϕℓ(X~)-ϕℓ(X)∥1,Lfft=∥F(X~)-F(X)∥1,33\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{L}}}_{{\mathrm{tv}}}=\mathop{\sum }\limits_{i,j}({\parallel {\nabla }_{x}{\widetilde{X}}_{i,j}\parallel }_{1}+{\parallel {\nabla }_{y}{\widetilde{X}}_{i,j}\parallel }_{1}). $$\end{document}Ltv=∑i,j(∥∇xX~i,j∥1+∥∇yX~i,j∥1).34\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{L}}}_{{\mathrm{GLRP}}}={{\mathcal{L}}}_{\mathrm{rec}}+{\lambda }_{{\mathrm{perc}}}{{\mathcal{L}}}_{{\mathrm{perc}}}+{\lambda }_{{\mathrm{fft}}}{{\mathcal{L}}}_{{\mathrm{fft}}}+{\lambda }_{{\mathrm{tv}}}{{\mathcal{L}}}_{{\mathrm{tv}}}. $$\end{document}LGLRP=Lrec+λpercLperc+λfftLfft+λtvLtv.Here, ∥ ⋅ ∥~1, *M*~ sums within the masked region, *ϕ*~*ℓ*~ are fixed perceptual feature maps, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\mathcal{F}} $$\end{document}F is the discrete Fourier transform, and ∇~*x*~, ∇~*y*~ are first-order difference operators. The terms promote context-consistent inpainting in occluded regions, identity preservation in known regions, morphology and texture and spectral fidelity, and spatial smoothness, respectively, guiding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{G}}}_{\theta } $$\end{document}Gθ to perform minimal modifications needed to remove lesions and thereby produce boundary-sensitive Δ at inference. To seamlessly integrate Δ as an interpretable prior downstream, GLRP serves two coupled roles in the framework. First, Δ is injected as a gated gain into the decoder together with the continuous prior *P* at each scale (see the gating formula given previously). Second, GLRP imposes an “empty-mask consistency” structural regularization on the tumor-free reconstruction. Let the segmentation mapping be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\mathcal{S}}(\cdot ) $$\end{document}S(⋅) with pixel-wise probabilities, and define the uncertainty weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ w(x)=1-\widehat{U}(x) $$\end{document}w(x)=1-U^(x) using the normalized uncertainty \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \widehat{U}\in [0,1] $$\end{document}U^∈[0,1] from SEGU. We require35\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{L}}}_{\varnothing }=\frac{1}{| \Omega | }\mathop{\sum }\limits_{x\in \Omega }w(x)\,{\mathcal{S}}({X}_{{\mathrm{no}}-{\mathrm{tumor}}})(x), $$\end{document}L∅=1∣Ω∣∑x∈Ωw(x)S(Xno-tumor)(x),where *Ω* is the image domain. This constraint suppresses foreground responses on the generative “lesion-removed” control, reducing over-segmentation and false positives at the system level without altering the parameterization of the classification or prior-fusion branches. In summary, GLRP establishes a generative structural-evidence pathway from *X* to Δ with low dependence on appearance and good geometric correspondence. It supports boundary enhancement in U-SEG with SEGU and provides a high-confidence structural basis for constructing the continuous prior used by FUSE. To avoid boundary hallucinations or ringing around sharp interfaces, GLRP is trained as a masked lesion-removal inpainting model with a composite loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{L}}}_{{\mathrm{GLRP}}} $$\end{document}LGLRP that combines image-space reconstruction, perceptual, frequency-domain and total-variation terms, enforcing identity outside the lesion mask and minimal, spatially smooth edits within it. In this study, GLRP is used as an auxiliary structural pathway that shapes the differential prior and the empty-mask consistency loss rather than as a stand-alone tumor-free image generator for clinical interpretation. For the public cohorts considered here there is no ground-truth tumor-free counterpart, and we therefore do not attempt a separate numerical evaluation of the reconstructed images. Instead, the combination of reconstruction, perceptual, frequency-domain and total-variation terms together with the empty-mask consistency constraint is designed to produce a stable difference map that highlights regions where the segmentation branch assigns non-negligible tumor probability while preserving normal tissue elsewhere. To unify generative structural evidence and discriminative heatmap evidence into a decoder-consumable and stable continuous prior under weak supervision, we construct the Fusion-based Uncertainty-weighted Seed Estimator (FUSE). The inputs are the differential map Δ from Generative Lesion Removal Prior (GLRP), the multi-scale class activation maps CAM~*m**s*~ from CPG, and the pixel-wise uncertainty \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \widehat{U} $$\end{document}U^ from U-SEG with SEGU as a down-weighting factor. The outputs are a continuous prior *P* ∈ [0, 1] and an ignore mask \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {M}_{ignore} $$\end{document}Mignore. The rationale is that Δ captures structural change and boundary cues from the contrast between tumor-present MRI and its lesion-removed counterpart, while CAM~*m**s*~ encodes class-related responses from a discriminative view. Each source is coarse with different noise characteristics; enforcing spatial consensus via intersection for purity, extending coverage via union, and applying quantitative weighting with uncertainty down-weighting yields a stable, interpretable prior, particularly for small and low-contrast lesions. Concretely, FUSE is implemented as a deterministic mapping from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ (\Delta ,{CAM}_{ms},\widehat{U}) $$\end{document}(Δ,CAMms,U^)g to (*P*, *M*~ignore~) so that each step can be reproduced. We first upsample all Grad-CAM++ maps to the input resolution and aggregate them with nonnegative weights {*η*~*s*~} into a single heatmap *L*, and apply temperature normalization *Z*~*T*~( ⋅ ) to both *L* and Δ. Hard seeds are then formed by thresholding at *τ*~Δ~ and *τ*~cam~ to obtain *S*~Δ~ and *S*~cam~, from which the core and extended regions *Ω*~core~ and *Ω*~ext~ are defined by intersection and union. To avoid discontinuities, we replace the hard indicators by smoothed sigmoids *s*~*τ*,*t*~( ⋅ ) and construct soft core and extension masks \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\widetilde{\Omega }}_{core} $$\end{document}Ω~core and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\widetilde{\Omega }}_{ext} $$\end{document}Ω~ext, which are further weighted by region-wise confidence kernels *q*~core~ and *q*~ext~ derived from (Δ, *L*) with distance-based decay. Pixel-wise uncertainty enters through \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ w(x)={(1-{U}_{b}(x))}^{\gamma } $$\end{document}w(x)=(1-Ub(x))γ and multiplicatively down-weights both regions before aggregation into the continuous prior *P* ∈ [0, 1], while the ignore mask *M*~ignore~ removes locations with low uncertainty weight or vanishing evidence in both sources. Finally, *P* and Δ are resampled to each decoder scale via *R*~*k*~( ⋅ ) and injected into the U-SEG decoder with 1 × 1 channel-alignment layers *A*~*k*~, *B*~*k*~ and gated additive modulation. In practice, the hyperparameters were determined via a grid search on the BraTS-2018 validation split and subsequently fixed for all experiments. The hard thresholds defining core and extended regions are set to *τ*~Δ~ = 0.30 and *τ*~*c**a**m*~ = 0.40. The soft-indicator temperature is set to *t* = 0.1, and the distance-based decay exponent is fixed to *γ* = 1.5, which balances seed sharpness against numerical stability. To preserve multi-scale evidence on a unified scale, we first align Grad-CAM++ maps to the input resolution and aggregate them with nonnegative weights. Let the scale set be {*s*}, normalized maps \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\widehat{{\mathcal{L}}}}^{c,(s)}\in {[0,1]}^{H\times W} $$\end{document}L^c,(s)∈[0,1]H×W, weights *η*~*s*~≥0 with ∑~*s*~*η*~*s*~ = 1. Define temperature-normalized 36\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \overline{{\mathcal{L}}}={{\mathcal{Z}}}_{T}(\mathop{\sum }\limits_{s}{\eta }_{s}\,{\widehat{{\mathcal{L}}}}^{c,(s)}),\,\Delta ={{\mathcal{Z}}}_{T}(\Delta ), $$\end{document}L¯=ZT (∑sηsL^c,(s)),Δ=ZT(Δ),where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {Z}_{T}(z)={z}^{T}/{{\mathrm{max}}}_{x}{z}^{T} $$\end{document}ZT(z)=zT/maxxzT for *T* > 0. Larger *T* produces sharper priors whereas smaller *T* produces flatter priors. Binary seeds are then formed by 37\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {S}_{\Delta }={\bf{1}}\{\Delta \ge {\tau }_{\Delta }\},\,{S}_{{\mathrm{cam}}}={\bf{1}}\{\bar{{\mathcal{L}}}\ge {\tau }_{{\mathrm{cam}}}\}, $$\end{document}SΔ=1{Δ≥τΔ},Scam=1{L¯≥τcam},and the core and extended regions are38\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\Omega }_{{\mathrm{core}}}={S}_{\Delta }\wedge {S}_{{\mathrm{cam}}},\,{\Omega }_{{\mathrm{ext}}}=({S}_{\Delta }\vee {S}_{{\mathrm{cam}}})\backslash {\Omega }_{{\mathrm{core}}}. $$\end{document}Ωcore=SΔ∧Scam,Ωext=(SΔ∨Scam)\Ωcore. To avoid discontinuities from hard thresholds, we use a smoothed indicator *s*~*τ*,*t*~(*z*) = *σ*((*z* − *τ*)/*t*) with temperature *t* > 39\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\mathop{S}\limits^{ \sim }}_{\Delta }={s}_{{\tau }_{\Delta },t}(\Delta ),\,{\mathop{S}\limits^{ \sim }}_{{\mathrm{cam}}}={s}_{{\tau }_{{\mathrm{cam}}},t}(\bar{{\mathcal{L}}}), $$\end{document}S~Δ=sτΔ,t(Δ),S~cam=sτcam,t(L¯),and replace Boolean operators with soft 40\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\mathop{\Omega }\limits^{ \sim }}_{{\mathrm{core}}}=\min ({\mathop{S}\limits^{ \sim }}_{\Delta },\,{\mathop{S}\limits^{ \sim }}_{{\mathrm{cam}}}),\,{\mathop{\Omega }\limits^{ \sim }}_{{\mathrm{ext}}}=\max ({\mathop{S}\limits^{ \sim }}_{\Delta },\,{\mathop{S}\limits^{ \sim }}_{{\mathrm{cam}}})-{\mathop{\Omega }\limits^{ \sim }}_{{\mathrm{core}}}. $$\end{document}Ω~core=min(S~Δ,S~cam),Ω~ext=max(S~Δ,S~cam)-Ω~core. To quantitatively weight non-consensus regions, we impose a distance based decay so the extended region anneals toward the core. With distance *d*(*x*, *Ω*) and scale *τ* > 0, define region-wise confidence 41\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {q}_{{\mathrm{core}}}(x)=\min (\Delta (x),\,\bar{{\mathcal{L}}}(x)). $$\end{document}qcore(x)=min(Δ(x),L¯(x)).42\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {q}_{{\mathrm{ext}}}(x)=\max (\Delta (x),\,\bar{{\mathcal{L}}}(x))\cdot \exp (-\frac{d(x,{\Omega }_{{\mathrm{core}}})}{\tau }). $$\end{document}qext(x)=max(Δ(x),L¯(x))⋅exp(-d(x,Ωcore)τ). Pixel-wise uncertainty provides unified down-weighting. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \widehat{U}={{\mathcal{Z}}}_{1}(\widehat{U})\in [0,1] $$\end{document}U^=Z1(U^)∈[0,1] 43\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ w(x)={(1-\widehat{U}(x))}^{\gamma },\,\gamma \ge 1. $$\end{document}w(x)=(1-U^(x))γ,γ≥1.We regularize soft sets \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\mathop{\Omega }\limits^{ \sim }}_{{\mathrm{core}}},{\mathop{\Omega }\limits^{ \sim }}_{{\mathrm{ext}}} $$\end{document}Ω~core,Ω~ext via a light smoothing operator \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\mathcal{S}} $$\end{document}S, and aggregate with hierarchical weights to form the continuous 44\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ P={{\mathcal{Z}}}_{1}(w\odot [{\lambda }_{{\mathrm{core}}}\,{\mathcal{S}}({\mathop{\Omega }\limits^{ \sim }}_{{\mathrm{core}}})\odot {q}_{{\mathrm{core}}}\,+\,{\lambda }_{{\mathrm{ext}}}\,{\mathcal{S}}({\mathop{\Omega }\limits^{ \sim }}_{{\mathrm{ext}}})\odot {q}_{{\mathrm{ext}}}])\in [0,1], $$\end{document}P=Z1(w⊙[λcoreS(Ω~core)⊙qcore+λextS(Ω~ext)⊙qext])∈[0,1],with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\lambda }_{{\mathrm{core}}} > {\lambda }_{{\mathrm{ext}}} > 0 $$\end{document}λcore>λext>0. The ignore mask for weak-supervision losses and prior distillation screens out low-confidence or evidence-absent 45\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {M}_{{\mathrm{ignore}}}={\bf{1}}\{\,w < \epsilon \,\,{\rm{o}}{\rm{r}}\,\,(\bar{{\mathcal{L}}} < {\tau }_{\min }\,\wedge \,\Delta < {\tau }_{\min })\}. $$\end{document}Mignore=1{w<ϵor(L¯<τmin∧Δ<τmin)}. Interfaces with the decoder follow prior definitions. For each scale *k*, set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {P}_{k}={{\mathcal{R}}}_{k}(P) $$\end{document}Pk=Rk(P) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\Delta }_{k}={{\mathcal{R}}}_{k}(\Delta ) $$\end{document}Δk=Rk(Δ), and inject via gated additive 46\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {D}_{k}^{+}={D}_{k}\odot \sigma (1+{\alpha }_{k}\,{{\mathcal{A}}}_{k}({P}_{k})+{\beta }_{k}\,{{\mathcal{B}}}_{k}({\Delta }_{k})), $$\end{document}Dk+=Dk⊙σ (1+αkAk(Pk)+βkBk(Δk)),where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{A}}}_{k},{{\mathcal{B}}}_{k} $$\end{document}Ak,Bk are 1 × 1 linear mappings for channel and range alignment and *α*~*k*~, *β*~*k*~ are learnable gates. Properties and mechanisms are as follows. First, consensus intersection ensures purity in the main where both Δ and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \overline{{\mathcal{L}}} $$\end{document}L¯ are high, the core is emphasized by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \min (\cdot ,\cdot ) $$\end{document}min(⋅,⋅) and further stabilized by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\lambda }_{{\mathrm{core}}} $$\end{document}λcore and *w*. Second, quantitative union increases coverage while suppressing *q*~ext~ decays with distance from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\Omega }_{core} $$\end{document}Ωcore, limiting expansion driven by a single strong source. Third, uncertainty down-weighting aligns prior confidence with the segmentation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ w={(1-\widehat{U})}^{\gamma } $$\end{document}w=(1-U^)γ consistently reduces the influence of hard pixels in both core and extended parts, mitigating noisy gradients in gated injection and weak losses. Fourth, differentiable and scale-aligned operators (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{Z}}}_{T} $$\end{document}ZT, *s*~*τ*,*t*~, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\mathcal{S}} $$\end{document}S, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {{\mathcal{R}}}_{k} $$\end{document}Rk) make *P* smoothly responsive to upstream evidence and uncertainty, and strictly aligned with U-shape decoding across layers.