Empowering photodynamic therapy with artificial current trends and future directions

Authors: Avijit Paul, Marvin Xavierselvan, David Aebisher, Tomasz Kubrak, Dorota Bartusik-Aebisher, Srivalleesha Mallidi

Abstract

The evolution of photodynamic therapy (PDT), from ancient photomedicine practices to modern clinical applications, reflects its remarkable versatility in oncology and beyond. PDT relies on the interaction between photosensitizers, light, and tissue oxygen to generate reactive oxygen species that selectively destroy diseased cells. While the therapy has proven effective across various cancers and non-malignant conditions, tailoring treatment to individual patients remains challenging due to patient-specific variations in tissue optical properties, photosensitizer pharmacokinetics, and tumor heterogeneity. The rapid advancement of artificial intelligence (AI), including machine learning and deep learning, offers transformative opportunities to address these challenges through data-driven optimization and personalization. In this review, we examine how AI is being integrated across the PDT pipeline. We analyze AI-driven approaches for photosensitizer development, including quantitative structure-activity relationship modeling, graph neural networks for property prediction, and generative models for de novo molecular design. We examine machine learning applications in nanoparticle-based drug delivery systems, encompassing synthesis optimization, nano-bio interaction prediction, and stimuli-responsive release modeling. The review further explores AI integration in treatment planning through real-time tissue optical property estimation, and in clinical decision-making through treatment response monitoring and outcome prediction using multimodal imaging data. We critically assess current limitations, including small dataset challenges, model interpretability concerns, and the gap between preclinical research and clinical translation. Finally, we outline future directions, including federated learning, explainable AI, and regulatory considerations. This review aims to bridge the AI and PDT communities, providing a roadmap for improved patient outcomes.

Introduction

Photodynamic therapy (PDT) is a minimally invasive treatment modality that combines a photosensitizing drug, light, and tissue oxygen to selectively destroy malignant and diseased cells (1–3). Since the first FDA approval in 1995 (4, 5), PDT has demonstrated efficacy across multiple cancer types and non-malignant conditions (6–8). Among established cancer treatments, PDT offers distinct spatial selectivity through localized light delivery, minimal systemic toxicity, and repeatability without cumulative dose limitations (2, 6). Yet clinical adoption has not matched its therapeutic potential. The fundamental barrier is not the therapy itself but our inability to tailor treatments to individual patients, a challenge that artificial intelligence (AI) is uniquely positioned to address.

Although PDT has ancient roots where civilizations like Egypt, Greece, and India used plant extracts and sunlight to treat skin ailments (9, 10), its modern scientific foundation dates to 1900 when Raab first observed light-activated cytotoxicity. Clinical development accelerated after Dougherty’s trials in the 1970s established therapeutic efficacy (4, 11–14). FDA approval of Photofrin in 1995 marked a milestone but also established standardized treatment protocols that persist today (5). Subsequent generations of photosensitizers (PS), including phthalocyanines, chlorins, and improved porphyrin derivatives, addressed early limitations such as prolonged photosensitivity while expanding clinical applications (15). Recent advances include integration with imaging modalities for real-time monitoring, nanoparticle-based delivery systems, and combination approaches with immunotherapy and chemotherapy (16–22). Yet despite these technological advances, PDT protocols remain largely standardized rather than individualized.

Despite this progress, the gap between PDT’s potential and its clinical reality stems from biological variability that current protocols cannot accommodate. Tissue optical properties, including absorption and scattering coefficients, vary substantially between patients and even within the same tumor, with resulting variations in delivered fluence that can exceed 100% in some cases (23). A light dose that achieves a therapeutic effect in one patient may be insufficient or excessive in another. PS pharmacokinetics add another layer of uptake rates, biodistribution patterns, and clearance kinetics differ based on tumor type, vasculature, and individual patient factors (24). Tumor heterogeneity compounds these challenges. Oxygen concentration, critical for the photodynamic reaction, varies spatially within tumors and changes during treatment as the photodynamic process itself consumes oxygen (18). Hypoxic regions may receive subtherapeutic doses while well-oxygenated areas experience adequate treatment. Current clinical protocols rely largely on standardized parameters, such as fixed light doses and standard drug-light intervals, that ignore this patient-specific variability (25). The consequence is highly variable response rates depending on the indication and patient selection. Improving outcomes requires moving beyond one-size-fits-all protocols toward adaptive, individualized treatment planning.

Machine learning (ML) is a subset of AI in which algorithms learn patterns from data rather than following explicitly programmed rules (26–28). Traditional ML approaches, such as support vector machines (SVMs), K-nearest neighbors (KNN), principal component analysis (PCA), Naive Bayes, random forests, and gradient boosting, excel in structured data analysis but require domain experts to manually define relevant features (29–36) (Figure 1). Deep learning (DL) overcomes this limitation through neural networks with multiple hierarchical layers that automatically extract features from raw inputs. Convolutional neural networks have revolutionized medical image analysis, while recurrent architectures and transformers capture temporal and sequential patterns (37–41). These methods have driven advances in medical imaging, drug discovery, and clinical decision support (42–47), with particular success in radiation therapy treatment planning (48–51). PDT presents a high-dimensional parameter space spanning tissue optical properties, PS concentration, oxygen levels, and light dose (25), creating optimization challenges ideally suited to ML approaches. Pattern recognition from multimodal imaging enables real-time treatment adaptation, while structure-property prediction accelerates PS discovery. The computational tractability of PDT physics, combined with growing imaging datasets, positions this modality for significant AI-driven advances.

AI showed transformative potential in radiotherapy treatment planning (52) and accelerated drug discovery pipelines (46), yet PDT scarcely appears in the broader AI-oncology literature (53, 54). Conversely, PDT reviews address computational modeling such as Monte Carlo (MC) simulations and finite element methods (55), but rarely engage with modern ML techniques. This gap is PDT’s high-dimensional parameter space, encompassing tissue optics, PS pharmacokinetics, oxygen dynamics, and light dosimetry, is precisely the type of complex optimization challenge where AI excels. Moreover, critical areas remain transfer learning for small datasets, federated approaches for multi-institutional collaboration, and regulatory pathways for clinical implementation. This review bridges the AI and PDT communities by providing the first comprehensive survey of ML and DL applications across the complete PDT pipeline, from PS discovery and nanoparticle design through treatment planning, response monitoring, and outcome prediction. Beyond cataloging existing work, we critically assess translational barriers, including small dataset challenges, model interpretability, and regulatory pathways that must be addressed for clinical implementation. We conclude with future research directions for realizing AI-empowered personalized PDT.

Photodynamic mechanisms and applications

PDT: photophysics and mechanisms

PDT, at its core, operates on a deceptively simple light activates a drug to destroy diseased tissue. Yet beneath this simplicity lies a precisely orchestrated sequence of photophysical events, oxygen dynamics, and biological responses that collectively determine therapeutic success. Understanding these mechanisms is essential, not only for optimizing treatment protocols but also for identifying the parameters that machine learning models can leverage to predict and improve outcomes. The process begins when photons excite PS molecules from their ground state to an excited state, initiating reactions that generate reactive oxygen species (ROS) (56). The Jablonski diagram (Figure 2) illustrates the quantum transitions underlying this process. In biological systems, these ROS interact with cellular membranes, proteins, and DNA, causing oxidative damage that results in cell death—the therapeutic goal of PDT. Upon absorbing a photon with energy matching the gap between ground state ( PS0) and excited singlet state ( PS1*), the PS undergoes

The excited PS then undergoes intersystem crossing (ISC) to a longer-lived triplet state (^3^ PS*) (19). ISC efficiency is critical for PDT effectiveness, as higher ISC yields mean more PS molecules reach the triplet state, leading to more interactions with molecular oxygen and greater ROS production. From the triplet state, the PS can generate cytotoxic species through two competing pathways (19, 57, 58).

Type I: free radical pathway. The triplet-state PS interacts directly with cellular substrates, transferring electrons or hydrogen atoms to form radicals that react with oxygen to produce superoxide anion ( O2−), hydroxyl radical ( OH•), or hydrogen peroxide ( H2O2) (59–61).

Type II: singlet oxygen pathway. The triplet-state PS transfers energy directly to ground-state molecular oxygen (^3^ O2), converting it to highly reactive singlet oxygen (^1^ O2) (59, 60). Singlet oxygen causes extensive oxidative damage to cellular membranes, proteins, and DNA, leading to cell death.

Both Type I and Type II mechanisms play crucial roles in PDT, with each pathway offering unique advantages depending on the specific clinical context (57, 58, 61–63). The relative efficiency depends on the PS used and the local tissue microenvironment, including the presence of substrates and the concentration of molecular oxygen. Type I PDT may be more effective in hypoxic (low oxygen) environments for certain tumor microenvironments, as it can still generate ROS through alternative pathways. Type II PDT is typically more efficient in well-oxygenated tissues where singlet oxygen production can occur readily.

Cornerstones of PDT

Three interdependent components majorly determine PDT the PS, light parameters, and oxygen availability. PSs are light-activated compounds that generate ROS upon excitation (59, 60, 64). Effective PSs must absorb light in the red to near-infrared region (NIR) (600–800 nm) to achieve adequate tissue penetration. They must also resist photodegradation during prolonged exposure, exhibit high quantum yield for ISC, and accumulate selectively in target tissue. Additionally, they should show minimal dark toxicity to ensure cytotoxic effects occur only upon illumination (65–68). PSs have evolved across several generations. Porphyrins such as Photofrin absorb around 630 nm and offer moderate tissue penetration; they were the first clinically approved agents (64, 69). Chlorins like Chlorin e6 absorb at longer wavelengths (~660 nm), providing improved penetration and higher singlet oxygen quantum yields (64, 70). Phthalocyanines (~670–700 nm) offer excellent photostability, while bacteriochlorins (~735–740 nm) achieve the deepest tissue penetration (71, 72). These agents can be administered topically for superficial lesions, intravenously for deep-seated tumors, or conjugated to antibodies and peptides for targeted delivery. Light characteristics directly influence PDT outcomes (73, 74). The wavelength must match the PS’s absorption spectrum, with red and NIR light preferred for their ability to penetrate several millimeters to centimeters into tissue. For superficial lesions, light is typically delivered externally using lasers or light-emitting diodes, while deep-seated tumors require interstitial optical fibers positioned within the tissue (73, 74). The total energy delivered (fluence, J/cm²) and rate of delivery (irradiance, mW/cm²) both require optimization. Too little energy fails to generate sufficient ROS, while excessive irradiance depletes local oxygen faster than replenishment, paradoxically reducing efficacy. Fractionated protocols address this limitation by interspersing dark intervals between exposures, allowing oxygen replenishment and often improving outcomes over continuous illumination (73). Molecular oxygen is essential for generating the ROS responsible for PDT’s cytotoxic effects, and its availability influences both Type I and Type II reaction efficiency, though Type I shows reduced oxygen dependence (75–77). Because PDT consumes oxygen during ROS generation, a continuous supply is necessary to sustain the photodynamic effect, yet this demand creates a challenge. Prolonged illumination leads to localized oxygen depletion in the treated area, and many solid tumors already exhibit hypoxic regions due to inadequate blood supply (78). Hypoxia particularly impairs Type II PDT, which relies heavily on oxygen for singlet oxygen production. Several strategies have been developed to address this limitation and are reviewed in detail elsewhere (79–81). Understanding and managing the role of molecular oxygen is essential for optimizing PDT protocols and achieving consistent clinical outcomes.

Biological effects of PDT

PDT destroys tumors through three interconnected direct cell killing, vascular damage, and immune activation (82, 83). The balance among these effects depends on PS localization, light dose, and tissue characteristics. At the cellular level, ROS-mediated damage can trigger either apoptosis or necrosis. When ROS damage is moderate and localized—particularly to mitochondria—cells undergo apoptosis. This controlled form of death is characterized by caspase activation, chromatin condensation, and fragmentation into apoptotic bodies that phagocytes clear without provoking inflammation (Figure 3a) (84, 85). The intrinsic pathway, which predominates in PDT-induced apoptosis, begins when mitochondrial damage releases cytochrome C, initiating the caspase cascade. The extrinsic pathway involves ROS-induced upregulation of surface death receptors that activate caspase-8 (86). When oxidative stress is overwhelming, however, cells die by necrosis—an uncontrolled process in which extensive lipid peroxidation destroys membrane integrity, causing cells to swell and rupture (Figure 3b) (87–90). Necrotic cells release their contents, including damage-associated molecular patterns (DAMPs), which activate immune cells and trigger inflammation. This inflammatory response can amplify tumor destruction but may also affect surrounding healthy tissue. Beyond direct cytotoxicity, PDT damages tumor vasculature. ROS injure endothelial cells lining tumor blood vessels, disrupting vessel integrity and activating platelets to form microthrombi. The resulting vascular occlusion starves the tumor of oxygen and nutrients, causing widespread ischemic cell death (91, 92). PDT also disrupts angiogenic signaling pathways, inhibiting the formation of new blood vessels and contributing to sustained tumor control. The relative contribution of direct cell killing versus vascular shutdown varies with treatment parameters and can be tuned for specific clinical objectives.

Applications of PDT

The ability to destroy diseased tissue while sparing healthy structures has made PDT valuable across multiple medical specialties. From early applications in esophageal and skin cancers, PDT has expanded to include ophthalmologic disorders, localized infections, gynecologic conditions, and precancerous lesions in multiple organ systems. Its minimally invasive nature and ability to be repeated without cumulative toxicity make it attractive where surgery poses a significant risk or where tissue preservation is paramount. Table 1 summarizes established and investigational applications.

Despite proven efficacy across these diverse applications, PDT outcomes remain inconsistent. Treatment response varies with tumor oxygenation, PS biodistribution, light penetration through irregular tissue geometries, and patient-specific factors that are difficult to assess or predict clinically. This complexity arising from the interplay of photophysical, biological, and anatomical variables creates both a challenge for standardized protocols and an opportunity for computational approaches. The following section introduces machine learning and deep learning techniques that can extract patterns from high-dimensional PDT data and enable the predictive, personalized treatment strategies that conventional methods cannot achieve.

Learning techniques and advances

Machine learning fundamentals

ML encompasses algorithms that identify patterns and make predictions from data without explicit programming for each specific task (26–28). Rather than encoding rules manually, these methods learn relationships directly from examples. This approach is well-suited to the multivariable complexity where interactions among PS properties, light parameters, tissue characteristics, and biological responses resist simple analytical models. The standard workflow partitions data into training sets for model development, validation sets for parameter tuning, and held-out test sets for unbiased performance evaluation. This separation is models that perform well on training data but poorly on new examples have “overfit” to noise rather than learning generalizable patterns. Performance metrics vary by accuracy, precision, recall, and F1-score for classification; mean squared error and correlation coefficients for regression (125). Several learning paradigms prove relevant for PDT applications. Supervised learning trains models on labeled data where each input is paired with a known outcome, enabling tasks like predicting treatment response from pretreatment measurements or classifying tissue types from spectroscopic signatures. Unsupervised learning discovers structure in unlabeled data, identifying patient subgroups with similar characteristics or segmenting tissue regions based on optical properties without predefined categories. Reinforcement learning (RL) optimizes sequential decisions through interaction with an environment, receiving feedback that guides future actions. This paradigm is well-suited to adaptive treatment protocols that adjust light delivery based on real-time tissue response. Self-supervised learning, increasingly prominent in modern AI, generates training signals from the data itself, enabling pretraining on large unlabeled datasets before fine-tuning on limited labeled examples. This approach holds particular promise for PDT, where annotated clinical datasets remain scarce, but imaging data accumulates steadily. The choice among these paradigms depends on data availability and clinical goals. Supervised approaches require outcome labels (whether a patient responded, whether tissue was malignant), limiting training to completed cases with documented results. Unsupervised methods can leverage larger imaging datasets without outcome annotation, potentially discovering predictive patterns that supervised approaches constrained by existing labels might miss.

Classical machine learning methods

Since most PDT studies involve only tens to hundreds of labeled cases, algorithm selection must prioritize methods that learn effectively from limited data while providing interpretable predictions. Regression, tree-based, and distance-based approaches meet these criteria. Regression methods model relationships between input variables and continuous outcomes. Linear regression establishes baseline associations between treatment parameters and results, while logistic regression handles binary classification by modeling the probability of outcomes such as treatment success or tumor recurrence (126, 127). These interpretable methods remain valuable not only for prediction but for identifying which input features strongly influence outcomes, providing insight that can guide clinical decision-making and experimental design. Tree-based methods partition data through hierarchical decision rules learned from training examples. Decision trees offer transparency that facilitates clinical physicians can trace the logic from patient characteristics through branching criteria to predicted outcomes (128). Random forests aggregate predictions from hundreds of trees, each trained on random data subsets, improving accuracy while reducing the overfitting that plagues individual trees (31). Gradient boosting methods such as XGBoost and LightGBM build trees sequentially, with each new tree correcting errors from previous ones (129). These ensemble approaches frequently achieve state-of-the-art performance on tabular clinical data and provide built-in metrics for feature importance. SVMs find optimal decision boundaries by maximizing the margin between classes (130). Kernel functions project data into higher-dimensional spaces where linear separation becomes possible, enabling SVMs to capture complex nonlinear relationships in tissue classification tasks. KNN takes a simpler approach, classifying new samples based on similarity to labeled examples in feature space (131). While computationally straightforward, KNN can capture local patterns useful for patient matching applications. Clustering algorithms discover natural groupings without predefined labels (132). K-means partitions data into clusters based on distance to learned centroids, while hierarchical clustering builds tree-structured relationships revealing multi-scale organization. Both approaches help identify patient phenotypes or tissue subtypes that may respond differently to treatment, enabling stratified protocols tailored to subgroup characteristics. For the small-to-moderate datasets characteristic of single-institution PDT studies (typically tens to hundreds of patients), these classical methods often outperform DL approaches that require orders of magnitude more training examples. Their interpretability also addresses a practical clinicians understandably hesitate to act on predictions from opaque models they cannot interrogate or explain to patients.

Deep learning architectures

DL extends ML through multi-layer neural networks that automatically discover hierarchical feature representations from raw data (133–137). This eliminates the manual feature engineering required by classical methods, a substantial advantage when relevant features are unknown or difficult to specify mathematically. The tradeoff is data deep networks typically require thousands to millions of training examples to learn robust representations without overfitting. The source of this representational power lies in the network architecture itself. Neural networks consist of interconnected computational units organized in layers. Each unit applies learned weights to its inputs, sums the weighted values, and passes the result through a nonlinear activation function such as rectified linear unit or sigmoid (138). Training adjusts weights throughout the network via backpropagation, an algorithm that efficiently computes how each weight contributes to prediction errors and updates them accordingly (139). Stacking many layers enables the network to learn progressively abstract representations, from simple patterns in early layers to complex concepts in deeper layers (Figure 4). Convolutional neural networks (CNNs) revolutionized image analysis by learning spatial filters that detect meaningful patterns regardless of their location in an image (140). Early convolutional layers learn to detect edges and textures; deeper layers combine these into higher-level features like tumor margins or necrotic regions. Pooling operations between layers reduce spatial dimensions while preserving important information, making CNNs robust to small shifts in feature position. These properties make CNNs particularly valuable for PDT imaging segmenting tumor boundaries from fluorescence images, classifying tissue types from OCT scans, or detecting treatment response from serial imaging. Recurrent neural networks (RNNs) and their long short-term memory (LSTM) variants process sequential data by maintaining hidden states that carry information across time steps (141). Standard RNNs struggle with long sequences because gradients vanish or explode during backpropagation through many time steps. LSTMs address this limitation through gating mechanisms that control information flow, deciding what to remember, what to forget, and what to output at each step (142). These architectures suit PDT applications involving temporal monitoring photobleaching kinetics, tracking oxygenation changes during treatment, or analyzing time-series spectroscopic measurements. Generative models learn to create new data samples resembling their training distribution. Generative adversarial networks (GANs) pit two networks against each a generator that creates synthetic samples and a discriminator that distinguishes real from generated data (143). Through adversarial training, generators learn to produce increasingly realistic outputs. Diffusion models take a different approach, learning to reverse a gradual noise-addition process to generate samples from pure noise (144). Both architectures address PDT’s data scarcity challenge by generating synthetic training examples and augmenting limited clinical datasets with plausible variations. Transformers and attention mechanisms, originally developed for natural language processing, increasingly impact biomedical imaging and multimodal data integration (41). Self-attention enables each element in a sequence to attend to all other elements, capturing long-range dependencies without the sequential processing constraints of RNNs. Vision transformers apply these principles to images by treating image patches as sequence elements. For PDT, transformers offer potential for integrating diverse data types, combining imaging, spectroscopy, and clinical variables in unified models that learn cross-modal relationships.

With these algorithmic foundations established, attention turns to their practical deployment. The diversity of methods reflects the diversity of PDT image segmentation favors CNNs, temporal monitoring suits LSTMs, molecular design leverages generative models, and emerging multimodal problems may benefit from transformer architectures. Section 4 examines specific applications across the PDT pipeline. Table 2 summarizes representative ML/DL methods applied across five key domains of PDT: tissue optical characterization, PS design, nanoparticle optimization, treatment monitoring, and outcome prediction.

Applications of learning algorithms in PDT

Before examining specific applications, it is important to contextualize the current state of AI-PDT research. The studies reviewed here span a spectrum of validation stages, from computational simulations and phantom experiments through preclinical animal models to retrospective analyses of clinical data. Notably, no AI-PDT application has yet achieved prospective clinical validation or regulatory approval. Table 2 categorizes each study by validation stage to help readers distinguish between proof-of-concept demonstrations and approaches closer to clinical translation. This distinction is critical for assessing the maturity and near-term clinical relevance of different AI applications in PDT.

Tissue optical characterization and light dosimetry

Light propagation through tissue depends on absorption and scattering coefficients that vary between patients, tumor types, and even within individual lesions. Traditional PDT protocols assumed nominal optical properties, but these patient-specific variations significantly affect light dosimetry and treatment outcomes (145). ML offers approaches to recover optical properties in real time and adjust light delivery to individual tissue characteristics, with methods maturing over the past decade from proof-of-concept toward clinical deployment. Among the first applications of ML to PDT instrumentation, Xie et al. demonstrated spectroscopic probes enhanced by SVM regression for intraoperative guidance (146). Fiber optic systems combining fluorescence and reflectance measurements generate complex signals influenced by PS concentration, tissue optical properties, and ambient light contamination. Strong ambient light in the operating room poses particular challenges for accurate measurements. The team addressed this through a hand-held fiber optic probe system with fast pulse modulation and frequency-specific detection to suppress background interference. Their nonlinear least-squares SVM (LS-SVM) regression model quantified protoporphyrin IX concentrations in tissue phantoms while compensating for large variations in tissue scattering and absorption. This approach ensured that fluorescence signals correlated linearly with true fluorophore concentrations despite optical property variations. Applied to clinical skin lesion data, the approach achieved 100% diagnostic precision for tissue classification (Figure 5a). This work established that classical ML methods could handle the nonlinear signal dependencies inherent in real surgical settings, but the approach remained limited to pre-treatment tissue characterization with site-specific probe configurations. The question of whether ML could enable adaptation during treatment, rather than just before it, was addressed by Yassine et al. in their work on interstitial PDT for deep-seated tumors (145). When light must propagate through centimeters of tissue, uncertainties in optical properties become critical since small errors compound over longer path lengths. The team employed cylindrical diffusers that serve as both light sources and detectors, using photon measurements combined with treatment geometry and source locations as input features for neural networks and gradient boosting regression trees (GBRT). These models predicted the absorption coefficient ( μa) and reduced scattering coefficient (µs′) of tumor tissues in real time. The key advance was integration with treatment planning software PDT-SPACE, enabling immediate adjustment of optical power allocations based on recovered properties rather than assumed nominal values (Figure 5b). Validation on 3D virtual glioblastoma models constructed from real MRI images demonstrated significantly reduced prediction errors in critical volumes compared to conventional planning. This represented a shift from static pre-treatment characterization toward dynamic intra-treatment adaptation. However, purely data-driven neural networks, while fast and flexible, can operate as black boxes that occasionally produce physically unrealistic outputs such as negative optical property values.

Chong and Pramanik addressed this limitation through physics-guided neural networks (PGNN) that embed domain knowledge directly into network architecture (147). Traditional methods for estimating optical properties relied on physics-based models like the diffusion approximation (DA) or MC simulations. These approaches simulate light transport accurately but are computationally intensive, limiting their utility for real-time applications. Pure data-driven networks offer speed but sacrifice interpretability and physical consistency. The physics-guided approach bridges this gap by incorporating DA results as network inputs and applying physics-based constraints during training (Figure 5c). Evaluation on simulated datasets of single-layer tissue samples, with separate in-domain and out-domain test sets, demonstrated that these hybrid networks outperformed both traditional ML models and unconstrained neural networks in prediction accuracy, generalizability, and consistency with physical laws. By ensuring predictions remain physically plausible, this approach builds the reliability necessary for clinical trust. Yet even robust, physics-informed models face a practical barrier to spectroscopic probes vary in geometry and optical characteristics across clinical sites, and models trained on one configuration may not transfer directly to another. Transfer learning offers a solution to this deployment challenge. Hannan and Baran developed an artificial neural network (ANN) to predict optical properties from diffuse reflectance spectra collected by a specific fiber optic probe. The network used 8 input nodes corresponding to different source-detector separations and 2 output nodes for absorption and reduced scattering coefficients (148). The network combined two ensembles, one mapping experimental spectra to simulated reflectance and another mapping simulated reflectance to optical properties (Figure 5d). To adapt this model to probes with different source-detector configurations, they applied transfer learning by freezing the feature-extraction layers and retraining only output layers with small datasets from target probes. This approach significantly reduced prediction errors compared to applying the original model directly, as measured by root mean squared error and mean absolute percent error. The practical implication is rather than building and validating new models for each clinical site, institutions can adapt proven models to their equipment with minimal additional data collection. Collectively, these studies trace maturation from demonstrating feasibility to enabling practical deployment. Early work proved ML could handle clinical signal complexity. Subsequent advances enabled real-time treatment adaptation, physics-informed predictions that clinicians can trust, and transfer learning that reduces barriers to multi-site adoption. To summarize the translational status of these Xie et al. represents the only study with validation on human patients, achieving 100% diagnostic accuracy for skin tumor classification during clinical PDT. Yassine et al. and Chong and Pramanik developed their methods using computational simulations, while Hannan and Baran validated their methods using tissue-mimicking phantoms. This distribution, predominantly simulation and phantom studies with limited clinical validation, reflects the broader landscape of AI in PDT dosimetry and underscores the need for prospective clinical trials before implementation.

AI-driven photosensitizer development and design

The discovery and optimization of PSs represent one of the most promising applications of artificial intelligence in PDT. Traditional PS development has relied on iterative synthesis and characterization, a process that is time-consuming, resource-intensive, and often guided by chemical intuition rather than systematic exploration of the vast molecular space (46). The integration of ML and DL approaches into PS design offers opportunities to accelerate this process while achieving optimized photophysical and biological properties tailored for therapeutic applications. Advances have progressed from predicting properties of known compounds to generating novel structures and screening for clinical viability. Figure 6 illustrates this integrated pipeline, showing how molecular representations feed into property prediction, generative design, mechanism screening, and experimental validation through an active learning feedback loop.

Predictive modeling of photosensitizer properties

A fundamental challenge in PS development is accurately predicting photophysical properties that determine PDT efficacy, particularly the singlet oxygen quantum yield (ΦΔ), absorption wavelength, and ISC efficiency. These properties are critical for Type II PDT, where singlet oxygen generation directly correlates with therapeutic outcomes (56). Traditional quantum chemical calculations, such as time-dependent density functional theory (TD-DFT), while accurate, are computationally expensive and impractical for high-throughput screening of large molecular libraries. Quantitative structure-property relationship (QSPR) models emerged as early tools for predicting PS photophysical properties. Foundational work established approaches using linear regression and SVMs to correlate molecular descriptors with singlet oxygen quantum yields for porphyrins and BODIPYs (149, 150). These models demonstrated that PS efficacy is influenced by multiple molecular descriptors, including lipophilicity, electrostatic parameters, and quantum chemical descriptors such as HOMO-LUMO gaps (151). While limited in accuracy, this work established that computational prediction of PS properties was feasible. Recent advances have substantially improved prediction accuracy through more sophisticated ML approaches. He et al. developed ML models using Morgan fingerprints and combined molecular/quantum chemical descriptors to predict singlet oxygen quantum yields across diverse PS structures (152). Their XGBoost model achieved R^2^ values of 0.86, demonstrating superior performance compared to traditional approaches. Importantly, these models incorporated experimental conditions, including solvent type and excitation wavelength, that significantly influence measured ΦΔ values. This approach provided more universally applicable predictions rather than predictions tied to specific measurement conditions. Specialized approaches address particular PS classes. For heavy-atom-free BODIPY PSs that undergo spin-orbit charge transfer ISC (SOCT-ISC), Buglak et al. established QSPR models using over 5,000 calculated molecular descriptors (149). Their multiple linear regression models achieved correlation coefficients of R = 0.88–0.91 for predicting ΦΔ values across solvents of varying polarity, enabling pre-synthetic screening of environment-activatable PS candidates. This approach is particularly valuable for designing theranostic agents that combine fluorescence imaging with PDT capabilities, where balancing fluorescence quantum yield and singlet oxygen generation presents a fundamental design challenge. For transition metal complex PSs, Wang et al. proposed a hybrid DFT-ML framework where excited-state quantum chemistry descriptors were integrated with molecular descriptors to predict ΦΔ for Ru-, Ir-, and Re-based complexes (153). This physics-informed approach addresses the limitation that traditional SMILES-based representations cannot adequately capture the structural complexity of metal-ligand interactions.

Graph neural networks (GNNs) have emerged as particularly powerful architectures for molecular property prediction due to their ability to naturally represent molecules as graphs where atoms correspond to nodes and bonds to edges (154). Unlike descriptor-based methods requiring manual feature engineering, GNNs automatically learn hierarchical molecular representations capturing both local and global structural features. Joung et al. employed graph convolutional networks trained on over 30,000 chromophore-solvent combinations to predict peak absorption and emission wavelengths (155). Their approach accounts for chromophore-solvent interactions through concatenated feature vectors, achieving a root mean square error of 26.6 nm for absorption wavelength prediction across diverse molecular scaffolds. The integration of molecular fingerprints with graph representations has further enhanced prediction capabilities. Hybrid GNN architectures combining molecular graphs with Morgan fingerprints capture both topological and electronic features, achieving R² values up to 0.92 for maximum absorption wavelength prediction (156). These advances are particularly relevant for PS design, where absorption in the red to NIR is essential for therapeutic tissue penetration. Graph transformers incorporating attention mechanisms represent the current frontier. The AAPSI (AI-Accelerated PhotoSensitizer Innovation) workflow employs graph transformers trained on a curated database of over 102,000 PS-solvent pairs to predict both ΦΔ and λmax simultaneously (157). This multi-objective approach enables identification of PS candidates that optimize the inherent trade-off between singlet oxygen generation efficiency and absorption wavelength, a trade-off that single-property optimization cannot address.

Generative models for de novo PS design

Property prediction enables screening of existing compounds, but generative AI models offer the more revolutionary capability to design novel PS structures with desired properties from scratch. This represents a paradigm shift from traditional approaches that rely on modifying known scaffolds based on chemical intuition. Deep generative models, including variational autoencoders (VAEs), GANs, and transformer-based architectures, can learn the underlying distribution of molecular structures and generate novel candidates satisfying multiple property constraints (158). RL approaches enable goal-directed molecular optimization where generated structures are iteratively refined based on predicted property scores. The ReLeaSE (Reinforcement Learning for Structural Evolution) framework combines generative and predictive neural networks, training them jointly to bias molecular generation toward desired properties (159). For PS optimization, RL agents can maximize singlet oxygen quantum yield while simultaneously optimizing for red-shifted absorption, low dark toxicity, and synthetic accessibility. Multi-objective RL frameworks such as MolDQN (Molecule Deep Q-Networks) directly define modifications on molecules while ensuring 100% chemical validity, avoiding the issue of generating synthetically infeasible structures (160). Diffusion models, which learn to reverse a noise-adding process, represent an emerging class of generative architectures showing promising results for molecular design (161). Their application to PS design, while nascent, holds potential for generating structurally novel candidates beyond the chemical space of known PSs. The first experimentally validated AI-driven PS design workflow was demonstrated through the AAPSI platform (157). This closed-loop framework integrates expert knowledge, scaffold-based molecule generation, and Bayesian optimization. From 23 expert-curated scaffolds, the system generated 6,148 synthetically accessible candidates screened using graph transformers. The hypocrellin-based candidate HB4Ph, identified at the Pareto frontier of high ΦΔ and long λmax, was subsequently synthesized and experimentally validated, demonstrating exceptional photodynamic performance. This validation establishes that AI-generated PS candidates can exhibit competitive or superior performance compared to empirically discovered PSs, moving the field from theoretical promise to demonstrated capability.

Mechanism and safety screening

Beyond photophysical efficacy, clinical viability requires control over photochemical mechanism and safety profile. The distinction between Type I and Type II PDT mechanisms has significant implications for treating hypoxic tumors, where Type I PSs may offer advantages due to reduced oxygen dependence (61). AI models capable of classifying PS mechanism preference are therefore clinically valuable. Chen et al. developed a multi-stage screening workflow combining a graph convolutional network for Type I PS identification with subsequent screening for NIR absorption and RNA-targeting capabilities (162). From an initial pool of 2,768 molecules, the model identified 782 Type I PSs, demonstrating the ability to navigate chemical space toward mechanistically distinct classes. Chemical space visualization using t-SNE dimensionality reduction revealed that Type I and non-Type I PSs exhibit distinct structural distributions, reflecting underlying electronic features governing photochemical mechanism. For Type II PSs, models targeting the singlet-triplet energy gap and ISC efficiency are critical. Chen et al. developed fragment-based and character-based models integrating conditional transformers, recurrent neural networks, and RL (163). These models significantly outperformed traditional baselines with prediction accuracies of 73% versus 4% for identifying high-efficiency triplet PSs. An ideal PS should exhibit minimal dark toxicity while maximizing phototoxicity upon illumination. Predicting these properties is essential for identifying clinically viable candidates during early development. ML models for phototoxicity prediction have leveraged data from standardized assays such as the 3T3 Neutral Red Uptake (NRU) test. Schmidt et al. developed Random Forest and deep neural network models using pharmacophoric fingerprints and quantum chemical descriptors, achieving accuracies of 83–85% with sensitivity reaching 86–90% (164). Graph convolutional approaches have further improved prediction; Igarashi et al. developed a model achieving F1 scores of 0.857 for predicting phototoxicity (165). Importantly, integrated gradient methods enable visualization of substructures contributing to predictions, providing interpretable insights for structural optimization rather than black-box classification. Multi-task deep learning models trained on in vitro, in vivo, and clinical toxicity data offer comprehensive screening by simultaneously predicting multiple toxicity endpoints (166). The integration of mechanism classification and safety prediction into virtual screening workflows enables comprehensive candidate evaluation before synthesis. Combined with active learning strategies that iteratively select the most informative compounds for experimental evaluation, these approaches can accelerate PS discovery while minimizing experimental costs (167). The AAPSI validation of HB4Ph demonstrates that this integrated pipeline, from property prediction through generative design to mechanism and safety screening, can yield experimentally validated PSs with superior performance (157).

AI in nanoparticle design and drug delivery systems

Nanoparticle-based delivery systems have emerged as essential platforms for enhancing PDT efficacy by addressing fundamental limitations of conventional PS administration. Traditional PSs often suffer from poor water solubility, nonspecific biodistribution, suboptimal tumor accumulation, and inadequate tissue penetration depth (181, 182). Nanocarriers, including liposomes, polymeric nanoparticles, mesoporous silica, metal-organic frameworks, and upconversion nanoparticles, offer solutions to these limitations. They provide enhanced permeability and retention effects, active targeting capabilities, and the ability to co-deliver multiple therapeutic agents (183, 184). However, the vast parameter space governing nanoparticle design, encompassing size, shape, surface chemistry, composition, and drug loading, has traditionally required extensive trial-and-error experimentation. The integration of AI and ML into nanoparticle design enables rational, data-driven optimization that accelerates development timelines while improving therapeutic outcomes (185, 186).

Machine learning for nanoparticle synthesis optimization

The synthesis of nanoparticles for PDT applications involves numerous interdependent variables that collectively determine physicochemical properties and biological performance. ML algorithms have demonstrated remarkable capability in navigating this complex parameter space to optimize synthesis conditions and predict nanoparticle characteristics. Tao et al. provided a comprehensive framework demonstrating how ML can guide the synthesis of colloidal nanocrystals by learning structure-property relationships from experimental data, significantly reducing the number of experiments required to achieve target specifications (187). For lipid-based nanoparticles, which serve as important carriers for PS delivery, AI-driven approaches have revolutionized formulation development. Li et al. developed an accelerated ionizable lipid discovery platform that combines ML with combinatorial chemistry (172). By screening 584 ionizable lipids and training models on the resulting data, they predicted transfection efficiency across a virtual library of 40,000 compounds. The identified candidates outperformed established benchmarks in cellular delivery, demonstrating the power of computational screening to identify optimal formulations without exhaustive experimentation. Similarly, the AI-Guided Ionizable Lipid Engineering platform extends this approach by employing GNNs to screen thousands of lipid variants, achieving enhanced delivery efficiency while dramatically reducing experimental burden (188). ANNs and gradient boosting algorithms have proven particularly effective for predicting nanoparticle size, polydispersity, and encapsulation efficiency based on synthesis parameters. Chaurawal et al. demonstrated the integration of ML with Design of Experiments approaches to optimize sorafenib-loaded nanoparticles, achieving precise control over particle characteristics while minimizing experimental iterations (189). The combination of automated optimization algorithms with high-throughput synthesis platforms enables rapid exploration of formulation space that would be impractical through conventional approaches. Physics-informed ML models represent an emerging strategy incorporating domain knowledge into predictive frameworks. By constraining model predictions to obey fundamental physical principles governing nanoparticle formation and behavior, these hybrid approaches achieve improved generalizability and interpretability compared to purely data-driven methods (190). Such models can predict not only synthesis outcomes but also the relationship between synthesis conditions and downstream biological performance, enabling end-to-end optimization of nanoparticle formulations.

Predicting nano-bio interactions and biodistribution

Understanding how nanoparticles interact with biological systems is critical for designing effective PDT delivery platforms. Upon introduction into biological fluids, nanoparticles rapidly acquire a protein corona that fundamentally alters their biological identity, affecting biodistribution, cellular uptake, and therapeutic efficacy (191). Predicting protein corona composition and its functional consequences has traditionally been challenging due to the complex interplay of physicochemical factors governing protein adsorption. ML approaches have demonstrated significant progress in addressing this challenge. Ban et al. developed ML models that predict the functional composition of the protein corona and subsequent cellular recognition of nanoparticles, achieving R^2^ values exceeding 0.75 for functional protein categories including immune proteins, complement proteins, and apolipoproteins (168). Their analysis identified nanoparticle surface modification and core material as the most important factors determining corona composition, providing actionable insights for rational nanocarrier design. Fu et al. extended this work by employing ensemble methods, including Extremely Randomized Trees and GBRT, to predict relative protein abundance on protein coronas, enabling comprehensive prediction of multiple proteins simultaneously (169). The Protein Corona Database represents a significant resource for the field, compiling data from over 80 studies spanning 2000–2024 and integrating quantitative profiles of nearly 2,500 adsorbed proteins across 817 nanoparticle formulations (192). Meta-analysis revealed that silica, polystyrene, and lipid-based nanoparticles smaller than 100 nm with moderately negative to neutral zeta potentials preferentially bind apolipoproteins APOE and APOB-100, proteins linked to receptor-mediated uptake and enhanced delivery efficiency. Such database-driven insights enable prospective design of nanoparticles with favorable corona compositions for enhanced tumor targeting. Predicting cellular uptake represents another critical application. Alafeef et al. applied ML to estimate internalization behavior of carbon nanoparticles across different cancer cell lines, demonstrating that computational models can accurately predict uptake efficiency based on physicochemical properties (170). For biodistribution prediction, Chou et al. developed an AI-assisted physiologically-based pharmacokinetic model integrating ML with mechanistic modeling to predict nanoparticle delivery to tumors in mice (171). Mi et al. further developed ML models predicting tissue distribution and tumor delivery efficiency, enabling prospective selection of formulations likely to achieve therapeutic concentrations in target tissues (193). Digital twin technology represents an emerging application enabling real-time simulation and optimization of drug delivery processes. By creating computational replicas of physical systems that update based on sensor data, digital twins can guide personalized treatment optimization and predict individual patient responses to nanoparticle-mediated therapy (194). Recent applications have reduced formulation optimization timelines from months to weeks while achieving enhanced stability and performance.

Optimizing nanocarriers for PDT-specific applications

The unique requirements of PDT, including efficient PS delivery, appropriate subcellular localization, and adequate light activation, create specialized demands for nanocarrier optimization. Several recent studies have demonstrated AI-driven optimization of nanoparticles specifically for phototherapeutic applications. Upconversion nanoparticles (UCNPs) address the fundamental limitation of light penetration depth in PDT by enabling NIR excitation and visible light emission for PS activation in deep tissues (195). These lanthanide-doped nanostructures can absorb NIR light (800–1000 nm) and emit UV/visible light capable of activating conventional PS, extending PDT applicability to tumors beyond the reach of direct visible light irradiation (196). Computational approaches can optimize UCNP composition and surface functionalization to maximize energy transfer efficiency to PSs while maintaining biocompatibility and tumor targeting capabilities. Smart nanoplatforms responding to tumor microenvironment stimuli represent another area of active development. pH-responsive, redox-responsive, and enzyme-responsive nanocarriers enable triggered release of PSs specifically within tumor tissues, minimizing off-target photosensitivity (197). ML models can predict stimulus-response characteristics based on nanocarrier composition, enabling rational design of platforms with optimal release kinetics for specific tumor types. Zhou et al. reviewed how nanogenerator strategies can overcome barriers in PDT, highlighting opportunities for AI-driven optimization of these emerging platforms (184). The integration of multiple therapeutic modalities within single nanoplatforms, combining PDT with chemotherapy, photothermal therapy (PTT), or immunotherapy, creates additional optimization challenges that benefit from AI approaches. Varon et al. developed predictive models for nanotechnology-based PDT combined with PTT, using ML to optimize laser radiation parameters for enhanced treatment efficiency (174). Their analytical function fitting (AFF) models identified optimal laser intensity and duration settings, establishing foundations for personalized combination therapy optimization where multiple treatment parameters must be simultaneously tuned.

Generative AI for novel nanocarrier design

Generative AI models, including GANs and VAEs, represent an emerging frontier in nanocarrier design by enabling de novo generation of novel molecular structures with desired properties (158). While extensively developed for small molecule drug design, these approaches are increasingly applied to nanomedicine formulation development with promising results. GANs have been employed to generate novel ionizable lipid structures for nanoparticle-mediated delivery, producing lipids with optimized pKa values and lipophilicity for enhanced endosomal escape (198). Out of over 1200 generated lipids, approximately 90% were synthesizable within three synthetic steps, and the top candidates demonstrated superior mRNA binding affinity compared to clinically approved lipids. This generative approach accelerates the exploration of chemical space far beyond what traditional screening methods can achieve. The directed evolution framework introduced by Shan et al. combines virtual and physical compound libraries with ML-driven analysis to iteratively refine nanoparticle designs (199). This approach merges combinatorial synthesis, DNA/peptide barcoding for high-throughput in vivo screening, and computational prediction to identify structure-activity relationships guiding optimization. Such integrated workflows represent the future of nanocarrier development, where AI continuously learns from experimental feedback to propose improved candidates. RL offers complementary capabilities by iteratively refining nanoparticle designs through feedback loops that reward configurations meeting performance criteria such as enhanced tumor targeting, reduced immune clearance, or optimal drug release kinetics (200). By learning optimal design strategies through trial-and-error interaction with simulated or real experimental environments, RL agents can navigate complex multi-objective optimization landscapes that challenge conventional approaches. The convergence of generative design, high-throughput experimentation, and autonomous synthesis platforms promises to accelerate nanocarrier development for PDT applications substantially.

Monitoring treatment responses

Monitoring treatment responses during PDT enables clinicians to assess therapeutic efficacy and adjust parameters based on observed tissue or cellular changes. Individual variations in PS uptake, tumor microenvironment, and biological response affect treatment efficacy, making real-time and post-treatment monitoring essential for personalized therapy. DL approaches now enable automated segmentation of vascular structures, quantification of cellular responses, and tracking of morphological dynamics across treatment time points. Vascular-targeted PDT (V-PDT) requires precise assessment of blood vessel responses to evaluate therapeutic efficacy. Xu et al. introduced the Global Attention-Xnet (GA-Xnet) model, a multi-step deep neural network designed for segmentation of subfascial blood vessels in the dorsal skinfold window chamber model (173). Accurate vessel segmentation in this context is challenging due to complex tissue architecture and imaging artifacts. The GA-Xnet pipeline employs three sequential stages. First, a Hough transform combined with a U-Net model extracts circular regions of interest from images. Second, an Attention U-Net learns global features and performs coarse segmentation of blood vessels (the GA step). Third, the coarse segmentation results combined with retinal images from the DRIVE dataset are input to a UNet++ model that learns multiscale features and produces fine segmentation maps (the Xnet step). This multi-step approach combining global feature learning, attention mechanisms, and transfer learning from established retinal vessel datasets, addresses the challenge of limited V-PDT training data (Figure 7a). The model achieved high accuracy, sensitivity, and specificity for subfascial vessel segmentation, demonstrating potential for clinical application in evaluating V-PDT responses. Further studies will be needed to assess the feasibility of long-term response evaluation, but the approach establishes a foundation for quantitative vascular monitoring. While vascular segmentation captures tissue-level effects, cellular response monitoring provides direct insight into treatment efficacy. Varon et al. addressed the challenge of optimizing combined PDT and PTT by systematically monitoring cellular responses across laser conditions (174). They presented a methodology combining nanotechnology-based PDT and PTT with ML to optimize cancer treatment parameters. In vitro cytotoxicity assays gathered data on cell death induced by PDT and PTT using a single nanocomplex, with measurements of cell death after light radiation divided into training and test sets. Three predictive models were developed to determine optimal laser radiation intensity and duration that maximize treatment efficiency (Figure 7b). The regression model predicted relationships between input parameters (laser intensity and duration) and treatment outcome. The interpolation model estimated values within the training data range, providing continuous predictions based on existing data points. The AFF model approximated the response surface with a low-degree function to predict treatment outcomes. The models were used to identify optimal laser radiation intensity and duration settings that maximize treatment efficiency. Comparing prediction errors across models, the AFF approach offered the best balance of simplicity and accuracy, making it particularly suitable for clinical applicability. This model was subsequently used for sensitivity analysis, examining how treatment performance responds to parameter variations, aiding in the refinement of treatment protocols. The work establishes a foundation for personalized cancer treatment optimization where monitored cellular responses directly inform parameter selection.

Understanding dynamic morphological changes in cancer cells during PDT offers insights for treatment personalization. Rahman et al. demonstrated the potential of DL, particularly the Cellpose algorithm, to enhance understanding of PDT’s dynamic effects on cancer cells (175). The advanced instance segmentation provided by Cellpose offers detailed insights into cellular responses, enabling personalized treatment strategies (Figure 7c). They analyzed fluorescence microscopic time series images of hepatocellular carcinoma (HCC) cells exposed to different laser intensities (6%-12.5%). These images included 16,100 cells segmented using the pre-trained Cellpose model, allowing for comprehensive morphological characterization, including dimensions and geometric properties. The authors employed Cellpose, a deep learning model based on U-Net architecture, for robust instance segmentation of cancer cells. Cellpose is known for its ability to handle diverse cell morphologies and provides superior accuracy compared to traditional methods. It was chosen due to its resilience to noise and capability to handle a wide range of cell shapes and sizes, which is crucial for accurately monitoring PDT effects on HCC cells. The authors used a pre-trained version, leveraging transfer learning to adapt the model to their specific dataset without extensive retraining. This approach saved time and computational resources while maintaining high segmentation accuracy. Following segmentation, various morphological features, including length, width, area, perimeter, and diameter, were extracted using the scikit-image library. This detailed morphological characterization was essential for understanding cellular responses to PDT. A human-in-the-loop approach identified and rectified instances of missing cells during segmentation, with iterative manual correction and retraining ensuring high prediction accuracy and reliability. To predict cellular behavior post-PDT, logistic growth modeling was applied, providing insights into dynamic responses of cancer cells to different laser intensities and enhancing understanding of PDT efficacy. The study included extensive measurement and analysis of morphological attributes of all segmented cells, offering insights into physical dimensions and geometric properties. The combination of Cellpose-based segmentation and logistic growth modeling revealed significant morphological changes in HCC cells after PDT, such as changes in cell size, shape, and proliferation rates. These findings were critical for optimizing PDT parameters and improving treatment efficacy. The study highlights the challenges of measuring onset times of anti-cancer drugs and the limitations of traditional methods. Their findings underscore the potential of DL to improve cancer cell analysis and contribute to the development of more effective cancer therapies. These approaches address complementary aspects of PDT monitoring. GA-Xnet enables precise vascular segmentation for evaluating V-PDT responses. Systematic cytotoxicity measurement across laser conditions informs ML-driven parameter optimization for combined therapies. Cellpose-based morphological tracking reveals dynamic cellular changes that guide treatment personalization. Together, these capabilities advance the real-time and post-treatment assessment needed for adaptive PDT.

Predicting treatment outcomes

Predicting treatment outcomes before or after PDT enables tailored protocols, efficient resource allocation, and informed clinical decision-making. Unlike real-time monitoring that guides immediate parameter adjustments, outcome prediction addresses longer-term which patients will respond to treatment, what factors influence recurrence, and how therapy can be personalized based on individual characteristics. AI models integrating imaging data with clinical variables have demonstrated significant predictive capability across diverse PDT applications.

Central serous chorioretinopathy (CSC) treatment with PDT presents a prediction challenge where identifying treatable versus refractory cases could guide therapeutic decisions. Yoo et al. developed DeepPDT-Net, a two-stage deep learning model predicting 1-year PDT outcomes using initial clinical data (176). Their dataset included 166 eyes with chronic CSC and 745 healthy control eyes. Clinical data encompassing demographic details, medical history, corticosteroid use, stress conditions, and ophthalmologic measurements, including fundus photographs (FPs), OCT measurements, and PDT protocol parameters. Data augmentation through flipping, translation, rotation, zooming, and shearing prevented overfitting, given the limited CSC dataset size. The first stage employed ResNet50 pretrained on ImageNet and fine-tuned to distinguish CSC from normal eyes using the larger dataset containing both patient and control FPs. The model was further fine-tuned on CSC eyes alone to differentiate treatable from refractory cases, adapting specifically to CSC-related features. For the combined approach, deep features extracted from FPs by the pretrained ResNet50 were combined with clinical variables using XGBoost, creating the DeepPDT-Net model that predicts the likelihood of complete subretinal fluid absorption after PDT (Figure 8a). Performance was evaluated using the area under the curve (AUC) and Youden’s index for optimal threshold setting. Shapley Additive Explanations (SHAP) analyzed the contribution of each input variable to model decisions, while Gradient-weighted Class Activation Mapping (Grad-CAM) generated heatmaps highlighting FP regions influencing predictions. The authors noted that while Grad-CAM provided insights into regions of interest, the heatmaps remained challenging to interpret and limited in clinical actionability, highlighting the ongoing gap between model explainability and clinical utility. While DeepPDT-Net focused on predicting treatment response, understanding the underlying choroidal changes in CSC requires detailed structural analysis. Previous studies relied primarily on 2D choroidal thickness measurements from swept-source OCT (SS-OCT), which may not capture the complex 3D structure involved in disease pathogenesis. Hara et al. introduced a novel AI-based method for 3D volumetric analysis of the choroid to better understand CSC and treatment response (178). Patients underwent SS-OCT before treatment and at 1 and 3 months post-treatment. A custom deep learning noise reduction algorithm utilizing U-Net was applied to original volumetric scans, followed by shadow reduction techniques addressing artifacts in deep choroidal layers. Attenuation compensation and contrast enhancement with a local Laplacian filter improved visibility of choroidal vessels. The preprocessed scans were segmented using the Topcon Advanced Boundary Segmentation algorithm with manual corrections as needed, and a fully automated composite method combining local and global thresholding accurately segmented choroidal vessels from stroma (Figure 8b). The segmented structures were visualized in 3D using VisIt software, enabling quantification of total choroidal volume, vessel volume, and stromal volume. This detailed 3D analysis provided insights into choroidal architecture changes with treatment that 2D measurements cannot capture, suggesting pathogenic mechanisms and enabling more precise response assessment.

![Figure 8: AI-based prediction of PDT treatment outcomes across clinical applications. (a) DeepPDT-Net predicts 1-year PDT outcomes for central serous chorioretinopathy (CSC) using a two-stage ResNet50 extracts deep features from fundus photographs through stepwise transfer learning, while XGBoost integrates clinical variables including demographic, medical, and ophthalmologic data to classify treatable versus refractory cases. Adapted from Choi, J.Y., Kim, H., Kim, J.K. et al. (176)]. (b) 3D volumetric analysis of choroidal structure in CSC using SS-OCT with U-Net denoising, shadow reduction, local Laplacian filtering, and Topcon Advanced Boundary Segmentation (TABS); VisIt software enables 3D visualization of choroidal vessel and stromal volumes. Adapted from (178). (c) Low-dose PDT (LDPDT) assessment for diabetic wound healing using Raman spectroscopy with 5-aminolevulinic acid (5-ALA) and methylene blue (MB) PSs; PCA with Mahalanobis distance quantifies wound healing progression by analyzing spectral changes in amide I and CH2 bands. Adapted from (177). (d) Recurrence prediction for oral leukoplakia (OLK) following ALA-PDT: whole slide images (WSIs) are segmented and processed through an autoencoder for feature extraction, K-means clustering assigns impact scores for recurrence risk, and COX regression combined with clinical variables achieves high AUC for 12-month prediction. Adapted from (179).](fonc-16-1771804-g008.jpg)

Beyond ophthalmology, PDT outcome prediction extends to other clinical applications. Zuhayri et al. investigated low-dose PDT effects on diabetic wounds in mice using two PSs, 5-aminolevulinic acid and methylene blue, with laser doses of 1 J/cm² and 4 J/cm² (177). Raman spectroscopy provided non-invasive biochemical and structural analysis of tissue state. The study developed a quantitative assessment method applying PCA to Raman spectroscopy data to reduce dimensionality and identify patterns explaining variance across wound healing states (Figure 8c). PCA transformed potentially correlated spectral features into linearly uncorrelated principal components, enabling visualization of chemical content changes distinguishing wound states across study groups. The analysis revealed that specific principal components (PC2 and PC3) in the spectral range of 2800–3000 cm^-1^, corresponding to amide I and CH2 band intensities, were particularly effective for monitoring the healing process. Mahalanobis distance quantified differences between wound states and a healthy skin reference, providing a metric for assessing how closely treated wounds approached normal tissue characteristics. This approach enables objective, non-invasive tracking of wound healing progression that could guide treatment decisions. Oral leukoplakia (OLK) treated with aminolaevulinic acid PDT (ALA-PDT) presents a recurrence prediction challenge where identifying high-risk patients could enable intensified follow-up or adjuvant therapy. Wang et al. applied ML and DL to enhance the prediction of short-term efficacy and recurrence (179). Whole slide images (WSIs) of pathological sections were segmented into 128×128 pixel patches using Python’s Openslide, with filtering criteria removing contaminated or irrelevant patches. An autoencoder, trained for 50 epochs to minimize reconstruction loss, extracted meaningful features from the image patches in an unsupervised manner. K-means clustering grouped the extracted feature vectors, and each patch was labeled based on whether the patient experienced OLK recurrence after PDT (Figure 8d). Impact scores calculated for each cluster determined the likelihood of promoting or preventing recurrence, with scores above 0.5 indicating recurrence-associated features and below 0.5 indicating protective features. Pathologists assessed patches within each cluster to categorize them by histomorphological characteristics, providing interpretable connections between learned features and tissue biology. Predictive models combining deep learning-generated features with clinical variables were constructed using logistic regression, COX regression, and SVM, with 10-fold cross-validation for evaluation. Comparison of models using only WSIs, only clinical variables, or both demonstrated that combining data types significantly improved predictive performance, with the combined model achieving high accuracy and AUC values. The COX regression model showed good predictive performance for recurrence up to 12 months after PDT, demonstrating that integrated pathological and clinical features can inform recurrence risk assessment. These prediction studies span diverse clinical contexts but share common themes. Integration of imaging features with clinical variables consistently improves performance over either data type alone. DL enables the extraction of features from complex medical images that complement traditional clinical predictors. Extending these approaches to incorporate multimodal data and enable real-time adaptive treatment represents the next frontier.

Future research directions

Integration of multimodal imaging and data fusion

Current AI applications in PDT typically analyze single data streams, whether optical spectra, fluorescence images, or clinical variables, yet the complexity of tumor biology and treatment response demands more comprehensive characterization. Integrating data from multiple imaging modalities, including MRI for anatomical context, PET for metabolic activity, OCT for microstructural detail, and fluorescence imaging for PS distribution, would provide holistic views of tumor characteristics and the dynamic microenvironment. Deep learning architectures capable of processing and fusing these diverse data types, whether through late fusion of modality-specific predictions or attention-based integration of learned representations, represent a critical research frontier (201). The challenge extends beyond technical architecture to practical implementation. Multimodal approaches complicate training and require careful handling of missing data when not all modalities are available for every patient. Research must develop robust fusion strategies that degrade gracefully when inputs are incomplete, maintaining clinical utility even with partial information. Additionally, combining imaging data with molecular markers, including genetic, proteomic, and metabolomic profiles, would enable truly integrative tumor characterization, moving PDT planning beyond morphological assessment toward biologically-informed treatment design. Federated learning offers a pathway to realize these ambitions despite data fragmentation across institutions (202). By training models locally and aggregating only weight updates rather than sharing sensitive patient data, federated approaches enable multi-institutional collaboration while preserving privacy (203). This paradigm is particularly valuable for PDT, where individual centers may have limited patient volumes, but collective datasets could support sophisticated multimodal models. Research priorities include developing federated architectures optimized for heterogeneous imaging protocols and establishing standardized data representations that facilitate cross-institutional learning (204).

Real-time adaptive PDT systems

The static nature of current PDT protocols, where treatment parameters are determined pre-operatively and applied uniformly, fails to account for the dynamic tissue responses occurring during light delivery. Real-time adaptive systems that continuously monitor tissue state and adjust treatment parameters accordingly represent a paradigm shift toward truly responsive therapy. RL algorithms that learn optimal control policies from real-time feedback, adjusting light dosimetry and irradiance patterns based on continuous physiological monitoring, could maximize therapeutic effect while minimizing collateral damage (205, 206). Such approaches have shown promising results in retrospective dosimetric studies of analogous adaptive radiotherapy applications, where RL-based systems dynamically optimize dose fractionation based on tumor response (207). Implementation requires integration of real-time sensing technologies including optical, thermal, and biochemical sensors into PDT delivery systems, providing immediate feedback on tissue oxygenation, PS photobleaching, and cellular stress responses. The computational challenge lies in processing these multivariate signals rapidly enough to inform treatment adjustments within clinically relevant timeframes. Edge computing architectures that perform inference locally rather than relying on cloud connectivity may prove essential for achieving the latency requirements of intraoperative adaptation. Synthetic data generation through GANs addresses a fundamental barrier to developing adaptive the limited availability of real-time monitoring data from diverse patient populations. GANs trained on existing tissue optical property datasets can generate realistic synthetic samples spanning parameter spaces underrepresented in clinical collections, enabling robust algorithm development prior to prospective validation. Research should focus on ensuring synthetic data captures the physiological variability and measurement noise characteristic of real clinical environments.

Personalized PDT protocol optimization

Moving beyond population-level treatment guidelines toward individualized protocols tailored to each patient’s tumor biology and systemic characteristics represents perhaps the most consequential opportunity for AI in PDT. Machine learning models analyzing genomic and proteomic data alongside imaging and clinical variables can identify patient subgroups with differential treatment responses, enabling precision medicine approaches to PDT planning. Pharmacogenomic insights may reveal genetic variants affecting PS metabolism, cellular susceptibility to oxidative stress, or immune response to PDT-induced immunogenic cell death. Foundation models pre-trained on massive chemical and biological datasets offer transformative potential for personalized PS selection and design (208, 209). These large-scale models learn universal molecular representations transferable to PDT-specific prediction tasks even with limited domain data. Incorporating three-dimensional molecular geometry and conformational dynamics into predictive frameworks, potentially through physics-informed neural networks embedding quantum mechanical principles for excited-state property prediction, would enhance accuracy for novel PS candidates (210). Multi-scale models integrating molecular-level predictions with cellular uptake, subcellular localization, and tissue distribution could optimize the entire PS-to-therapeutic-effect pathway. The vision of closed-loop automated laboratories, where robotic synthesis systems guided by machine learning iteratively design, synthesize, characterize, and refine PSs, is transitioning from speculation to early implementation (211, 212). Self-driving laboratories combining high-throughput experimentation with active learning algorithms that prioritize the most informative experiments could dramatically accelerate PS optimization cycles (213). Research should address integration challenges, including standardized characterization protocols, automated quality control, and feedback mechanisms linking in vitro results to computational model refinement.

AI-driven clinical decision support

Translating AI advances into clinical impact ultimately requires decision support tools that enhance rather than disrupt existing practice workflows. Developing interpretable models that provide not only predictions but also explanations accessible to clinicians represents a critical research priority (214). Techniques beyond gradient-weighted class activation mapping, including concept-based explanations, counterfactual reasoning, and uncertainty quantification, could help clinicians understand model recommendations and appropriately calibrate their trust (215, 216). Validation across diverse patient populations and clinical settings is essential before widespread deployment. Models developed at single institutions may fail to generalize when confronted with different imaging equipment, patient demographics, or clinical protocols. Research must establish robust frameworks for external validation, domain adaptation, and continuous model monitoring to detect performance degradation over time. Prospective clinical trials comparing AI-guided PDT protocols against standard-of-care will ultimately determine whether computational advances translate to improved patient outcomes. Integration pathways must consider clinical workflow constraints, regulatory requirements, and liability considerations. AI tools designed for seamless incorporation into existing electronic health records and treatment planning systems, with clear documentation of intended use and performance limitations, will facilitate adoption (217). Collaboration between AI researchers, PDT clinicians, and regulatory scientists is essential to navigate the evolving landscape of AI medical device oversight while maintaining the pace of innovation.

Research priorities and translational pathways

Unlike radiology, where over 1,200 FDA-authorized AI devices now operate within established infrastructure (218, 219), PDT lacks the foundational elements for AI translation. Most clinical PDT treatments are still administered with little more than application of a prescribed drug dose and timed light delivery, with no standardized dosimetry protocols across treatment centers (220). No benchmark datasets enable algorithm comparison, and no dedicated research consortium coordinates multi-institutional efforts. Addressing these infrastructure gaps represents the most urgent near-term priority. Realizing AI-empowered PDT requires coordinated progress across defined horizons informed by successful translations in analogous fields. Near-term efforts (1–3 years) should focus on establishing imaging protocol standardization through dedicated task groups within professional societies, building upon existing dosimetry guidelines (23), creating pilot multi-center data sharing agreements with standardized annotations, and curating initial retrospective datasets targeting 500-1,000 cases with harmonized acquisition parameters. The Medical Imaging and Data Resource Center, a joint ACR/RSNA/AAPM consortium that has accumulated over 150,000 imaging studies using FAIR data principles (221), provides a directly applicable governance template for PDT-specific infrastructure. Mid-term goals (3–5 years) include completing external multi-center retrospective validation across three or more institutions following established reporting guidelines (222), conducting prospective validation studies across diverse clinical sites, and engaging FDA through pre-submission meetings to establish evidence requirements for specific indications. The validation pathway proven for IDx-DR—the first autonomous AI diagnostic device—combined retrospective algorithm development with a prospective pivotal trial across 10 primary care sites, achieving FDA authorization within 89 days and demonstrating 87.2% sensitivity and 90.7% specificity for diabetic retinopathy detection (223). This hybrid retrospective-prospective model offers a directly applicable template for AI-PDT translation. Long-term objectives (5–10 years) encompass regulatory authorization through 510(k) or De Novo pathways, establishment of post-market surveillance programs, pursuit of reimbursement pathways (224), and deployment with continuous performance monitoring. Translational success requires academic-industry hybrid partnerships that maintain scientific rigor while enabling regulatory-grade development. Funding mechanisms including NIH SBIR/STTR programs, NIBIB Trailblazer Awards, and the NSF-NIH Smart Health Program support early-stage development, while ARPA-H’s PRECISE-AI initiative addresses the critical challenge of detecting and correcting AI algorithm drift post-deployment (225). Federated learning approaches enable multi-institutional collaboration while preserving data privacy (202, 204). Figure 8 illustrates the development pipeline and current translational barriers; the priorities outlined here provide a temporal roadmap for addressing these challenges.

Challenges and limitations

Small dataset challenges

Despite the remarkable progress detailed in the preceding sections, significant barriers remain between algorithmic development and clinical implementation. One of the most persistent obstacles confronting AI applications in PDT is the scarcity of high-quality, annotated datasets. Unlike general computer vision tasks, where millions of labeled images are readily available (226, 227), PDT-specific datasets remain remarkably limited. This scarcity stems from the specialized nature of the treatment modality, heterogeneous imaging protocols across institutions, and resource-intensive expert annotation requirements. This scarcity is compounded by the diversity of PDT datasets collected for ophthalmic PDT may have limited relevance for dermatologic or oncologic applications, fragmenting an already small data landscape. Models developed with limited data frequently exhibit overfitting, wherein they memorize training examples rather than learning generalizable patterns (228). This manifests as artificially inflated performance metrics during development that fail to translate to real-world clinical settings. In PDT, where treatment decisions carry significant implications for patient outcomes, such performance degradation poses unacceptable risks (229). Several strategies address these challenges with varying applicability. Transfer learning leverages models pre-trained on large general-purpose datasets, enabling reasonable performance even with limited PDT data (230). Studies have demonstrated that this approach can reduce prediction errors and enhance generalizability across different imaging configurations (148). However, fundamental differences between natural images and PDT-relevant data, including fluorescence imaging, OCT, and spectroscopic measurements, may limit feature transferability from non-medical domains (231). Data augmentation through geometric transformations and GANs offers another avenue for artificially expanding datasets, though concerns regarding the clinical validity of synthetic data persist (232, 233). The heterogeneity of PDT data presents additional complexity beyond mere quantity. Treatment protocols vary across institutions, PS differ in their optical properties and accumulation patterns, and light delivery systems span a range of configurations (3). This variability means that even datasets of nominally adequate size may not capture the full spectrum of clinical scenarios, potentially limiting model generalizability. Multi-center collaborations and standardized data collection protocols represent potential solutions. However, coordinating such efforts across institutions with different equipment, expertise, and regulatory environments remains challenging (234).

Interpretability and the black box problem

DL architectures frequently achieve predictive capabilities at the expense of interpretability. Neural networks with many layers and millions of parameters function as effective black boxes, producing outputs without transparent explanations of the reasoning process (235). In clinical settings where treatment decisions must be justified to patients, clinicians, and regulatory bodies, this opacity presents fundamental challenges (236). When an AI system recommends a specific PDT protocol or predicts treatment response, clinicians require an understanding of the factors driving these recommendations (237). Without such insight, practitioners cannot evaluate whether the model’s reasoning aligns with established clinical knowledge, identify potential errors, or appropriately calibrate their trust in the system’s outputs (238). This interpretability gap undermines AI integration into clinical workflows, regardless of their demonstrated accuracy on benchmark datasets. Explainable AI (xAI) methodologies have emerged in response to these concerns, offering techniques to elucidate model decision-making. Gradient-weighted Class Activation Mapping generates visual heatmaps highlighting influential image regions (239), revealing whether models attend to clinically relevant features such as subretinal fluid or retinal pigment epithelium changes in ophthalmic PDT applications (176). Similarly, SHAP values quantify individual feature contributions, enabling identification of influential clinical variables (240). However, current techniques remain limited in generating clinically actionable explanations. Heatmaps indicate which regions influenced predictions without explaining their pathophysiological relevance (128), and post-hoc explanations may not reflect true computational processes within the model (241). PGNNs represent a promising middle ground, incorporating domain knowledge of light-tissue interactions and photodynamic mechanisms into model architectures to constrain predictions within physically plausible bounds while maintaining interpretability (147). Such approaches have demonstrated improved accuracy in predicting tissue optical properties, a critical parameter for PDT planning. Beyond technical solutions, explanations must be tailored to end-user needs; what satisfies a computer scientist may differ substantially from what a treating clinician requires (242). Research on effective human-AI collaboration in medical decision-making remains nascent, and significant work is needed to understand how clinicians interpret, trust, and act upon AI-generated insights in PDT contexts (243).

Clinical translation gap

A striking disparity characterizes AI in PDT: while preclinical and computational research flourish, validated clinical implementations remain conspicuously absent (224). This translational gap reflects fundamental barriers beyond typical research-to-practice lag. The preponderance of AI-PDT research has relied on retrospective designs utilizing existing clinical data or phantoms and animal models. While essential for method development, these approaches cannot substitute for prospective clinical evaluation and are susceptible to selection bias that may not reflect contemporary practice (244, 245). The retrospective studies that have examined AI in PDT applications, primarily in ophthalmologic conditions, represent encouraging initial steps but lack prospective validation (180). Clinical translation faces interconnected obstacles. First, PDT’s diversity across oncology, dermatology, and ophthalmology means models developed for one indication may not generalize to others (246). Unlike radiotherapy, where relatively standardized workflows enabled broader AI adoption (52), PDT protocol heterogeneity across disease sites, PSs, and light delivery systems fragments development efforts and prevents the accumulation of experience with any single platform. Second, real-time treatment monitoring applications require computational infrastructure for rapid inference, seamless integration with existing imaging and light delivery systems, and user interfaces accessible to clinicians without specialized AI expertise (247). Developing such integrated systems demands interdisciplinary collaboration among AI researchers, biomedical engineers, clinical physicists, and treating physicians that remains challenging to coordinate (248). Third, evidence standards for clinical adoption exceed those typically reported in technical publications. Demonstrations of accuracy on held-out test sets are insufficient to establish clinical utility (249). Clinicians appropriately require evidence that AI tools improve patient outcomes, fit within existing care pathways, and offer acceptable risk-benefit profiles (250). Generating such evidence requires randomized controlled trials or well-designed prospective studies demanding substantial time, resources, and clinical infrastructure. The lack of active AI-PDT clinical trials perpetuates a cycle where insufficient clinical experience generates inadequate data to train and validate AI tools, providing little impetus for adoption (251). Addressing this gap requires academic medical centers as research incubators, industry partnerships for integrated platforms, and professional consensus on evidence standards. Most importantly, it requires demonstration of clear clinical benefit in well-designed studies (252, 253). Figure 9 synthesizes these multifaceted challenges into a comprehensive translational roadmap for AI-empowered PDT. The figure illustrates the development pipeline from AI-driven PS discovery through clinical implementation, mapping key barriers to corresponding enabling solutions. Data scarcity, interpretability limitations, clinical validation gaps, regulatory uncertainties, and biological constraints are each paired with targeted strategies. Technical approaches, including transfer learning, federated learning, and generative models, address small dataset challenges, while xAI methods such as PGNN, Grad-CAM, and SHAP offer pathways toward clinician trust. The roadmap also highlights deployment risks, including domain shift between preclinical and clinical settings, hidden confounders from imaging equipment variability, and bias amplification in small PDT cohorts. This integrated view underscores that successful clinical translation requires simultaneous progress across technical, regulatory, and validation domains.

Regulatory and ethical frameworks

PDT occupies an unusual regulatory position that presents distinct challenges for AI integration. Unlike most medical devices or drugs reviewed by a single FDA center, PDT systems are classified as cross-labeled combination products under 21 CFR 3.2(e), where the PS and light delivery device are separately packaged but specifically labeled for combined use (254). The PS’s pharmacological action constitutes the primary mode of action, placing the Center for Drug Evaluation and Research (CDER) as lead review center, with device components reviewed in consultation with CDRH (255). This bifurcated structure, established when porfimer sodium became the first approved PS in 1995, creates jurisdictional complexity that would be compounded by AI components requiring additional software oversight. Despite three decades of clinical experience, only six PSs have received FDA approval for limited indications, primarily in dermatology and ophthalmology (256). International regulators have authorized substantially broader applications; the EMA and Japan’s PMDA have approved agents for head and neck, lung, and esophageal cancers that remain unavailable in the United States (257, 258). This regulatory heterogeneity has direct implications for AI systems trained on European or Japanese clinical data incorporating these agents cannot be assumed valid for US submissions, while limited domestic indications constrain available on-label training data. Light delivery devices add further regulatory fragmentation. Laser-based PDT systems (FDA Product Code MVF) require Class III Premarket Approval (259). LED and lamp-based illuminators designed specifically for use with photosensitizing drugs such as the BLU-U, Aktilite CL128, and BF-RhodoLED, similarly require Class III authorization, typically as components of drug-device combination products (260). Non-coherent light devices for general phototherapy applications without PSs may receive Class II clearance through 510(k) under different product codes. The absence of FDA guidance specifically addressing PDT dosimetry compounds this fragmentation. Treatment parameters including light dose, irradiance, PS concentration, and drug-light interval vary substantially across institutions, with no standardized protocols analogous to those in radiotherapy (23). This variability is compounded by extensive off-label use documented across dermatology practices for indications including acne, photorejuvenation, and viral warts (261). For AI systems, training data inevitably incorporates these heterogeneous protocols, non-standardized dosimetry, and mixed on-label and off-label applications. An AI-enabled PDT platform integrating treatment planning, real-time dosimetry optimization, or outcome prediction would constitute an unprecedented three-component combination product spanning drug, device, and software jurisdictions. Such systems would likely require coordination between CDER, CDRH, and the Office of Combination Products, with formal Request for Designation advisable prior to development to establish the appropriate review pathway (262). The regulatory route would depend on how AI functionality integrates with existing whether as embedded dosimetry optimization within a light delivery device, standalone treatment planning software, or an integral system affecting drug dosing. The IMDRF Software as Medical Device framework provides risk categorization guidance, with AI controlling PDT parameters likely falling into Category III or IV given serious clinical consequences of dosimetry errors (263). Unlike radiology, where over 1,200 FDA-authorized AI devices, predominantly in radiology, with oncology and ophthalmology also represented, benefit from standardized imaging protocols and large annotated datasets (219, 264), PDT lacks this foundational infrastructure. The absence of benchmark datasets and obvious predicate devices combining PDT with AI suggests the De Novo pathway may be necessary for most AI-PDT applications (265).

Traditional regulatory frameworks designed for static medical devices are poorly suited to AI systems that must learn from variable real-world data while adapting over time (266). The FDA’s January 2025 draft guidance addresses this mismatch by establishing a Total Product Lifecycle approach, recognizing that AI devices require oversight not only at initial authorization but throughout operational lifespan, requiring documentation of model architecture, training data provenance, validation methodology, and performance monitoring throughout operational lifespan (267). For AI-PDT applications, compliance demands systematic characterization of training data sources, including explicit documentation of what proportion derives from off-label use, which PSs and light sources are represented, and how dosimetry parameters were recorded. Most authorized AI devices have received clearance through the 510(k) pathway based on substantial equivalence to predicate devices (265), while novel technologies have utilized the De Novo pathway (268). Understanding which pathway is appropriate for a given AI-PDT application and what evidence will be required is essential for developers planning regulatory strategies. The Predetermined Change Control Plan framework offers pathways for AI devices requiring post-market updates (269). Manufacturers can specify in advance allowable algorithm modifications and governing validation procedures. For AI-PDT tools incorporating real-time learning from treatment outcomes or periodic retraining as protocols evolve, this pathway may prove essential for maintaining compliance while enabling improvement. However, interaction between PCCP provisions for AI components and existing regulatory status of underlying PSs and devices remains subject to evolving interpretation (268). Labeling requirements have become increasingly specific, with implications for how AI-PDT developers must communicate inherent limitations. The FDA’s 2025 guidance recommends that labeling clearly indicate AI use, provide plain-language algorithm descriptions, detail model inputs and outputs, describe training data characteristics, report performance metrics, and disclose known limitations and bias sources (219). For AI-PDT devices trained on data from multiple PSs, varied light sources, or mixed indications, these requirements necessitate clear communication about which clinical scenarios have been validated and where generalizability may be limited. International harmonization remains incomplete, presenting additional challenges. The EU framework operates under the Medical Device Regulation and AI Act, with divergent approved PSs meaning systems trained on European data may not translate directly to US submissions (270). Differences in classification criteria, evidence requirements, and post-market surveillance obligations between jurisdictions require developers to navigate multiple regulatory pathways, potentially delaying the availability of beneficial technologies.

Beyond regulatory compliance, several ethical considerations warrant attention. Accountability for adverse outcomes from AI-assisted PDT involves multiple clinician, institution, PS manufacturer, device manufacturer, and AI developer. Current malpractice frameworks designed around human decision-makers may inadequately address this distributed responsibility (271, 272). Algorithmic bias presents particular concern. Analysis reveals less than one-third of FDA-authorized AI device evaluations provide sex-specific performance data, and only one-fourth address age-related subgroups (265). For AI-PDT, where skin pigmentation influences fluorescence imaging quality and PS visualization, ensuring equitable performance across diverse populations is both an ethical imperative and a practical necessity (273). The historical concentration of PDT trials in fair-skinned populations for conditions like actinic keratoses may introduce systematic bias that developers must actively address. Data privacy further intersects with development needs, as training robust AI-PDT models requires large multi-institutional datasets subject to HIPAA and GDPR protections (274). Federated learning enables collaboration without sharing raw patient data, potentially allowing institutions with different protocols and PS preferences to contribute while preserving privacy (275, 276). Effective clinical translation requires attention to human-AI interaction. Clinical decision support systems must foster appropriate reliance, where clinicians neither ignore beneficial recommendations nor follow them blindly (277). For PDT, where treatment decisions integrate patient factors, lesion characteristics, and dosimetry considerations, designing interfaces that support rather than supplant clinical judgment requires careful attention to how recommendations are presented and how clinicians are trained to interpret outputs (278). In this context, xAI transitions from a desirable feature to a necessity (279–284). A clinician must understand whether an AI’s recommended dose is driven primarily by lesion size, fluorescence intensity, or historical outcome patterns from similar patients. Saliency maps highlighting which image regions most influenced a recommendation, or counterfactual explanations showing how different inputs would change the output, can foster appropriate trust and facilitate error detection (285–287). The path to clinical translation for AI-PDT systems requires a proactive, parallel development strategy. Future efforts should focus on creating consortia to establish benchmark datasets with standardized annotation protocols for imaging, dosimetry parameters, and outcomes across diverse skin types. Digital phantoms and simulation platforms can augment scarce clinical data for initial algorithm training. Engaging with regulators early via the FDA’s Pre-Submission program offers opportunities to gain alignment on validation strategies for hybrid combination products. Adaptive clinical trial designs that can evaluate both the AI component and its interaction with the PDT combination product in iterative phases may prove essential for navigating the unique regulatory landscape these systems present. As AI-PDT applications mature from prototypes to clinical tools, institutional quality management systems addressing validation, monitoring, and bias detection must develop in parallel (288).

Conclusion

The historical evolution of PDT from ancient practices to modern clinical treatment highlights its potential and adaptability in oncology and beyond. PDT’s continued advancement, driven by innovations in PS design, imaging technologies, and delivery systems, underscores its growing importance in the medical field. Concurrently, the rapid progression of ML and DL offers unprecedented opportunities to enhance PDT’s efficacy and precision. Integrating AI with PDT can significantly improve tissue characterization, PS discovery, treatment monitoring, and outcome prediction, leading to personalized and optimized therapeutic approaches. Given the explosion of research in these two dynamic fields, this review aimed to synthesize existing knowledge, identify gaps, and provide guidance for future research that would be valuable for researchers, clinicians, and policymakers. However, realizing this potential requires addressing the challenges outlined data scarcity, interpretability limitations, translational barriers, and evolving regulatory frameworks. We hope that highlighting the synergy between AI and PDT will promote collaborative research efforts. Researchers in computer science, oncology, radiology, and bioinformatics must work together to bridge the gap between algorithmic innovation and clinical implementation. By addressing these technical challenges and exploring synergistic advancements, we can pave the way for more effective and individualized PDT treatments, ultimately improving patient outcomes and expanding the therapeutic potential of this versatile modality.