Authors: Eugene Shkolyar (1Department of Urology, Stanford University School of Medicine, Stanford, CA, USA; 2Veterans Affairs Palo Alto Health Care System, Palo Alto, CA, USA), Steve R. Zhou (1Department of Urology, Stanford University School of Medicine, Stanford, CA, USA), Camella J. Carlson (3Department of Biomedical Engineering, Vanderbilt University, Nashville, TN, USA), Shuang Chang (3Department of Biomedical Engineering, Vanderbilt University, Nashville, TN, USA), Mark A. Laurie (1Department of Urology, Stanford University School of Medicine, Stanford, CA, USA; 4Department of Radiation Oncology, Stanford University School of Medicine, Stanford, CA, USA), Lei Xing (4Department of Radiation Oncology, Stanford University School of Medicine, Stanford, CA, USA), Audrey K. Bowden (3Department of Biomedical Engineering, Vanderbilt University, Nashville, TN, USA; 5Department of Electrical and Computer Engineering, Vanderbilt University, Nashville, TN, USA), Joseph C. Liao (1Department of Urology, Stanford University School of Medicine, Stanford, CA, USA; 2Veterans Affairs Palo Alto Health Care System, Palo Alto, CA, USA)
Categories: Article
Source: Nature reviews. Urology
Authors: Eugene Shkolyar, Steve R. Zhou, Camella J. Carlson, Shuang Chang, Mark A. Laurie, Lei Xing, Audrey K. Bowden, Joseph C. Liao
Diagnostic cystoscopy in combination with transurethral resection of the bladder tumour are the standard for the diagnosis, surgical treatment and surveillance of bladder cancer. The ability to inspect the bladder in its current form stems from a long chain of advances in imaging science and endoscopy. Despite these advances, bladder cancer recurrence and progression rates remain high after endoscopic resection. This stagnation is a result of the heterogeneity of cancer biology as well as limitations in surgical techniques and tools, as incomplete resection and provider-specific differences affect cancer persistence and early recurrence. An unmet clinical need remains for solutions that can improve tumour delineation and resection. Translational advances in enhanced cystoscopy technologies and artificial intelligence offer promising avenues to overcoming the progress plateau.
The entry point to localized bladder cancer management is endoscopic visualization and resection of malignancy. White light cystoscopy (WLC) and transurethral resection of the bladder tumour (TURBT) provide diagnosis and — when performed thoroughly — reliable local staging and cancer control^1–4^. Despite considerable technological advances in the past few decades, such as fibre optics, digital endoscopy and enhanced cystoscopy^5,6^, the overall treatment paradigm remains largely unchanged, and bladder cancer recurrence and progression rates remain high after endoscopic resection^1–4^. This situation is not only a result of the aggressiveness of the disease state but also a product of persistent limitations in our techniques and tools; incomplete resection accounts for a substantial portion of early recurrences, with restaging TURBTs finding residual tumour in up to 40–50% of patients and upstaging in up to 15%^1–3^. Evidence also shows that TURBT performed by experienced surgeons can reduce tumour recurrence by up to five times when compared with less-experienced practitioners^4^. For this reason, an unmet clinical need remains for solutions that can enable improved tumour delineation and resection.
Translational advances in optical imaging and artificial intelligence (AI) — such as blue light cystoscopy (BLC) and deep learning — offer promising avenues to overcoming the progress plateau^7,8^. In this Review, we summarize the current landscape of enhanced imaging and AI research in endoscopic bladder cancer management. First, we highlight major milestones in imaging and endoscopy that have led to cystoscopy and TURBT in its current form^5,6^. Second, we discuss the limitations of the existing technology and evidence. Finally, we outline key translational applications of augmented cystoscopy and offer a vision for its future.
Cystoscopy in its current form evolved from a singular to produce the best possible image of the bladder. Thus, the actual differentiation of neoplasm from benign tissue depended not just on the technology but on the discerning eyes of the clinician. The first functional cystoscope was developed in 1877 by Maximilian Carl-Friedrich Nitze, who provided rudimentary visualization of the bladder through a telescopic lens under the illumination of a heated wire^6^. Major subsequent advances from numerous contributors focused on improving illumination and irrigation; the introduction of irrigation enabled hydrodistension rather than air insufflation, whereas the magnifying lenses enlarged the field of view^5,6^. The introduction of fibre optics in the 1950s brought about a major advance in cystoscope illumination, and ultimately led to the commercialization of the modern rigid cystoscope by Karl Storz in 1967 (refs. 5,6). Iterative improvements since then have included the development of flexible instruments, digital endoscopes, and high-definition cameras^5,6^. These improvements have enabled high-quality, minimally invasive management of numerous urological conditions — including bladder cancer.
Thorough endoscopic resection under WLC is the mainstay of non-muscle-invasive bladder cancer (NMIBC) management, followed by adjuvant intravesical therapies for patients with high-risk disease^9^. Emerging evidence also suggests that muscle-invasive bladder cancer, which has traditionally been managed with extirpative surgery, can be treated with TURBT combined with systemic chemotherapy and radiation if the endoscopic resection is thorough^10^. However, the quality of TURBT under WLC varies widely, with direct effects on cancer recurrence and progression. Results of numerous studies support the original observation that provider-specific differences affect tumour recurrence^4,9,11^. Incomplete resection accounts for a substantial portion of early recurrences, and residual tumour is found in up to 50% of patients with high-grade Ta disease, with upstaging occurring in 15%^1–3^. Among patients with T1 disease, restaging TURBT results in upstaging 40% of the time when muscularis propria is absent from the specimen, and 15% of the time when it is present^1^. For this reason, restaging TURBT is recommended in selected patients with NMIBC, particularly patients with T1 disease and recognition of incomplete resection from initial TURBT^12^. These restaging resections have been shown to improve recurrence-free and overall survival, especially in patients with high-risk disease^13,14^. In one randomized controlled trial, 210 patients were randomized to either restaging TURBT or not^15^. In the restaging-TURBT arm, residual tumour was found in 33% of patients, and 8% experienced upstaging to muscle-invasive disease. Recurrence-free survival, progression-free survival and cancer-specific survival were all improved in the restaging arm. Clinicians have taken some measures to improve TURBT with existing cystoscopy technology. En bloc endoscopic resection — the removal of the tumour in one piece to preserve the architecture and orientation of the sample — has been shown to improve pathological staging after TURBT, although ongoing work is aimed at determining whether a benefit in recurrence-free or progression-free survival exists^16^. The use of quality control metrics such as surgical checklists at the time of TURBT has been associated with improvement in recurrence-free survival^17,18^. However, these measures have yet to replace the need for restaging TURBTs owing to incomplete resections.
Standard TURBT under WLC guidance has limitations in detecting multifocal disease and differentiating cancer from surrounding benign or non-malignant tissue. Several enhanced cystoscopy technologies have been developed to assist clinicians to detect, enumerate and characterize bladder tumours (Fig. 1). These approaches either induce and/or detect malignant tissue change at the cellular and molecular levels or apply image processing and filtering to enhance existing visual differences.
Perhaps the most notable and widely adopted of these technologies is BLC. BLC leverages the phenomenon in which tumour cells of epithelial origin preferentially accumulate fluorescent endogenous porphyrins — such as protoporphyrin IX — when exposed to the haem biosynthesis precursor 5-aminolevulinic acid^19^. Hexaminolevulinate (HAL), a lipophilic derivative of 5-aminolevulinic acid with increased potency, is the approved imaging agent that is instilled in the bladder before endoscopic examination under blue light, thereby enabling the selective red fluorescence of neoplastic cells that might otherwise go undetected under WLC^20^.
The advantage of BLC over WLC for detecting bladder cancer was first reported in 1996 in a pilot study including 106 patients with suspected primary or recurrent malignancy, showing that BLC was more sensitive than WLC (96.9% versus 72.7%, P < 0.0001)^21^. Shortly thereafter, numerous prospective trials employing HAL as the imaging agent refined the clinical workflow of BLC and confirmed superior sensitivity in bladder cancer detection, particularly for carcinoma in situ (CIS) and Ta disease^22–24^. In 2013, a meta-analysis (total n = 831) showed that 24.9% of patients had at least one Ta or T1 tumour seen on BLC and missed on WLC. Additionally, BLC enabled identification of CIS not seen on WLC in 26.7% of patients^22^. In a large, multi-institutional prospective series in which BLC was evaluated in 1,632 tissue samples from 533 patients, BLC improved detection of papillary lesions by 12% and CIS by 43%^23^. BLC also helps detection of recurrences after intravesical bacillus Calmette–Guérin therapy^24^. Efforts are also ongoing to evaluate the role of in-office BLC for surveillance; results of a multicentre comparative study following surveillance of 304 patients showed that 13 of 63 participants (20.6%) had recurrent cancer undetected on WLC but identified on BLC, including 5 instances of CIS^25^.
Despite the demonstrated benefit of additional tumour detection, the effect of BLC on relevant cancer outcomes remains unclear; multiple meta-analyses support a reduction in disease recurrence, but the evidence is less conclusive regarding progression^26–28^. A 2013 meta-analysis of 12 randomized controlled studies including 2,258 patients demonstrated a reduced recurrence rate (OR 0.5, P < 0.00001) and increased time to recurrence by 7.39 weeks (P < 0.0001) when BLC was used to augment TURBT^26^. The benefit persisted at 2 years of follow-up monitoring (HR 0.65, P = 0.0004). In terms of progression to muscle-invasive bladder cancer, BLC did not demonstrate a benefit (OR 0.85, P = 0.39). Results of a 2022 Cochrane review of 16 randomized trials (4,325 patients) showed that TURBT with BLC reduced recurrence rates (HR 0.66, 95% CI 0.54–0.81), but also showed that progression rates were reduced as well (HR 0.65, 95% CI 0.50–0.84)^27^. Of note, based on a Grading of Recommendations, Assessment, Development, and Evaluations analysis, the Cochrane review determined the existing body of evidence to be of low certainty owing to heterogeneity in protocols, including variable use of intravesical BCG, repeat resection, and modern resection tool. To address these shortcomings, a new multicentre pragmatic randomized trial was conducted using up-to-date treatment pathways and endoscopic tools^29^. Results showed no difference in 3-year recurrence-free survival between WLC and BLC (HR 0.94, 95% CI 0.69–1.28), although the trial was ultimately still affected by similar issues to those in the Cochrane review; only 86 of 374 (23%) patients in the intermediaterisk cohort were exposed to BCG, many of whom received incomplete induction and/or maintenance courses, and an imbalance of treatment adherence between trial arms could affect results. Patients with CIS were also under-represented in the trial population, a subgroup that has previously demonstrated exceptional benefit from BLC^30,31^. Together, these limitations decrease the external validity of the findings in this contemporary trial.
Based on existing evidence, the American Urological Association (AUA) and Society of Urologic Oncology joint guidelines for the management of NMIBC (updated 2020) state that “a clinician should offer blue light cystoscopy at the time of TURBT, if available, to increase detection and decrease recurrence (Grade B)”. Based on expert opinion, they further recommend that “in a patient with a history of NMIBC with normal cystoscopy and positive cytology, a clinician should consider…enhanced cystoscopic techniques (blue light cystoscopy, when available)”^32^. The AUA guidelines notably mention that BLC has not been shown to decrease progression rates and does not mandate the widespread adoption of BLC as the standard. Similarly, the European Association of Urology only recommends BLC-guided biopsies in the setting of patients with positive cytology but negative WLC^33^. The National Comprehensive Cancer Network states that BLC can improve staging and detection, but “data are still limited … and WLC remains the mainstay of bladder cancer staging”^34^. The equivocal phrasing and lack of clear consensus between guidelines reflects the ambiguity of existing evidence for BLC; thus, the adoption of BLC in practice might ultimately depend upon cost, availability and local care pathways^34^.
Despite high initial and usage costs, incorporation of BLC is an economical option in the US health-care system. Evidence shows that use of BLC with initial TURBT introduces an upfront cost of 1,400 more than WLC; other longitudinal cost analyses show that the overall reduction in cancer recurrence over 5 years after a single initial instillation of HAL results in a net financial advantage (30,581 WLC)^35,36^. In clinical surveillance settings, results have shown that the addition of BLC to flexible cystoscopy introduces a net difference of $0.76 per cystoscopy over 2 years^37^. Overall, use of BLC might be cost effective in some health-care settings owing to a reduced cumulative recurrence rate and frequency of invasive procedures.
Narrow-band imaging (NBI) is an endoscopic technology (Olympus) that increases the contrast of vascularized tissue to improve the detection of bladder cancer^38^. In NBI wavelengths of 415 ± 30 nm and 540 ± 30 nm are used, which are absorbed by haemoglobin, enhancing capillary patterns in the urothelium, which might preferentially highlight the hypervascularity of malignant tissue. In contrast to BLC, NBI does not require intravesical instillation of an imaging agent to enhance the contrast of the malignant tissue. In a meta-analysis including 1,040 patients in 7 prospective studies comparing NBI with WLC, use of NBI helped the identification of cancer in an additional 17% of patients (95% CI 10–25%) and an additional 24% of tumours (95% CI 17–31%). The false-positive rate was notably increased with NBI in multiple individual studies, but this finding was not statistically significant on pooled analysis^39^.
In terms of oncological control, the data are not as clear. In 2012, the results of the first randomized study comparing the cancer recurrence rate 1 year after TURBT guided by NBI versus WLC (n = 148) were reported^40^. The NBI group had a lower recurrence rate than the WLC group (32.9% versus 51.4%, P = 0.01), but a subsequent larger multicentre randomized controlled trial including 965 patients reported no difference between the two modalities (25.4% NBI versus 27.1% WLC, P = 0.585). Sub-analysis did demonstrate an isolated reduction in recurrence rate in the NBI group for patients with low-risk cancer only^41^. Other meta-analyses do report that TURBT with NBI reduces recurrence rates, although the majority of the studies were non-randomized prospective trials^42,43^. Based on these conflicting results, the European Association of Urology and National Comprehensive Cancer Network both conclude that, currently, evidence supporting the benefit of NBI is insufficient^33,34^. The AUA and Society of Urologic Oncology currently provide a conditional recommendation (Grade C evidence strength) that clinicians “may consider use of NBI to increase detection and decrease recurrence”^12^.
CHROMA (Karl Storz) is another endoscope-based technology that does not require an exogenous imaging agent and achieves contrast enhancement of vascular lesions by image processing^44^. An early study of 165 tumours in 47 patients showed that CHROMA contrast enhancement helped the detection of 25 additional tumours not seen with WLC^7^. Further study is needed to fully characterize the benefit of this technology.
Several nascent technologies that offer high spatial resolutions beyond the macroscopic endoscopic view have been investigated for bladder cancer, specifically aimed at improving characterization of the suspected lesions such as CIS and providing cancer grading. Confocal laser endomicroscopy is a fibre optic imaging probe that can be inserted into the standard cystoscope working channel to provide an optical biopsy of suspicious bladder lesions with cellular resolution^45^. Results of meta-analyses have shown that confocal laser endomicroscopy offers between 54.6 and 93.6% correspondence with histopathological examination, but evidence is currently limited to a small number of prospective non-randomized studies^46^. Optical coherence tomography similarly enables real-time identification of cancerous tissue. Low-coherence light from a fibre optic interferometer illuminates target tissue, whereas high-speed longitudinal scanning captures the unique two-dimensional optical scattering pattern to reveal the internal microscopic biological tissue structure^47^. Early evidence shows that optical coherence tomography has excellent sensitivity (94.9%) and specificity (85.6%) for distinguishing bladder cancer from non-malignant tissue, but comparative and randomized data are lacking^48^. Pilot studies have also explored whether Raman spectroscopy can be used to provide an in vivo endoscopic bladder cancer diagnosis^49^. A very small portion of light reflected by matter will demonstrate inelastic scattering, and the consequent shift in frequency of these photons owing to the energy exchange — the Raman shift — is specific to the molecular bonds in the sample tissue^50^. Plotting this Raman spectrum produces a photonic vibrational fingerprint for a respective sample and can be used to detect dysplastic tissue change at the molecular level. These enhanced cystoscopy techniques have not gained widespread traction or spurred large-scale investigation, probably owing to the high costs of capital equipment, as well as the need for additional training for image interpretation.
Computer-aided diagnosis (CAD) uses software tools to identify suspicious features and aid diagnosis^51^. In the past two decades, CAD has been introduced into the workflow for medical imaging and employed to assist physicians in multiple tasks, including early detection of cancer^52^. AI and computer vision have been explored for cystoscopy^53^. AI offers several advantages over existing enhanced-cystoscopy technologies for bladder cancer diagnosis, including the potential to process high-density information with speed and consistency, and to be deployed at a lower cost, as no specialized endoscope is needed^54^.
Harnessing largely untapped cystoscopy video recordings, computer vision techniques have been investigated to generate panoramas^55,56^ to full 3D reconstructions^57–63^ of bladder cancer in the context of surrounding non-cancerous urothelium. Incorporating reconstructions into patient records could ease longitudinal comparisons of indeterminant bladder lesions, simplify multifocal tumour evaluation and facilitate intuitive data sharing; furthermore, it could enable a pipeline in which technical staff or a relatively inexperienced clinician conducts the cystoscopy for asynchronous review by an experienced urologist. Published 3D bladder reconstruction algorithms are structure-from-motion-based pipelines that involve frame selection and preprocessing, feature extraction and matching, geometric reconstruction, and texture mapping^57–63^ (Fig. 2). Several reconstructions have been successfully created from patient procedures, demonstrating feasibility and potential utility^55,57,58,62^. Phantom and simulation studies, which provide realistic organ-mimicking environments, have proved important for testing new pipeline iterations, evaluating the limits of algorithms, and refining user protocols^56,64,65^. Pipelines that integrate exogenous technologies such as structured light^60^ or sensors^63^ with the cystoscope can improve position estimation, but with the trade-off of additional cost and increased complexity for clinical deployment. Time-intensive processing renders most current algorithms only suitable for post hoc review. However, real-time 3D localization after an initial 3D reconstruction in the surveillance space has been developed by taking advantage of previous recordings of a patient’s bladder to generate a data map^61^ or when facilitated by precise robotic control^64^, potentially improving training and real-time intraoperative navigation.
Importantly, reconstructions are limited by the quality of the underlying source data; hence, some rely on a standardized scan technique to ensure feature visualization^55^. Cystoscopy video acquisition might require a prescribed path^57^ or, at minimum, ensure loops and the capturing of frames with overlapping features^58^. Factors that hinder feature visualization include active haematuria, debris and surgical instrumentation^55,57,58,66^. Current work is aimed at addressing these problems by providing real-time assessment of frame quality to the clinician to ensure the completeness and quality of the collected dataset^56,67^. Notably, full-bladder reconstruction is challenging with rigid cystoscopy, particularly near the bladder neck region^58^, and flexible cystoscopy with the capability for retroflexion might be better suited to full reconstruction^61^.
Last, cystoscopy data are feature rich, and textual documentation in standard electronic health records is not adequate to capture and distil the key imaging findings that are essential for longitudinal surveillance. To this end, an algorithm to automatically generate short video sequences from screenshot-captured bladder regions of interest has been devised^68^ and a systematic framework to standardize cystoscopy data acquisition, storage and retrieval defined^69^. Computer-aided tools can potentially address current shortcomings in cystoscopy, including variability in provider experience, incomplete visualization of the bladder, text-based documentation of cystoscopic findings and challenges in serial evaluation of bladder cancer patients undergoing surveillance.
Unlike enhanced cystoscopy, augmented cystoscopy through integration of AI requires no specialized endoscopic instrumentation or imaging agents. Given that such systems can potentially be deployed at a lower cost than existing enhanced cystoscopy technologies such as BLC, these advantages can maximize efficiency in high-resource settings and improve access to urology care in remote or low-resource settings^70^. In particular, AI has been investigated to improve the sensitivity of WLC^8^ and demonstrates great potential to objectively facilitate diagnosis in real time^53,71,72^. Current applications of AI in bladder cancer detection include frame-level classification^73–83^, tumour localization^8,84^, tumour segmentation^8,73,85–90^ and digital staining^91,92^ (Fig. 3).
Frame-level classification has been developed to differentiate images containing tumour from those with only benign tissue and urothelium (Fig. 3), and has been demonstrated with machine-learning classifiers^74^, artificial neural networks^76^, and CNNs^75,81,83^. For example, in an ablation study in which the performance of classical multilayer perceptron models for bladder tumour classification were compared, the area under the receiver operating characteristic curve score was 0.99 (ref. 76). To overcome the general scarcity of cystoscopy data sets required to train robust AI models, transfer learning has been employed in which a CNN pre-trained on a gastrointestinal dataset was fine-tuned using images of bladder tumours and non-malignant urothelium^77^. The fine-tuned CNN classified cystoscopy images with 89.7% sensitivity and 94.0% specificity. Efforts have also been undertaken to use video-based models — that is, models that consider longitudinal sequences of cystoscopy videos rather than individual frames — for frame-level classification^83^. A direct comparison of image classification trained using video-based models and ones trained with individual frames remains to be performed. Overall, frame-level classification is useful for the initial evaluation of cystoscopy video, particularly when the frame contains a solitary region of interest rather than multifocal disease.
Tumour localization extends frame-level classification by dynamically highlighting the general location of tumours in cystoscopy videos for frames that contain them (Fig. 3). A CNN-based frame analysis platform, CystoNet, has been constructed for detection of bladder cancer, particularly the papillary form^8^. In a validation cohort that consisted of 54 cystoscopies and 44 tumours, CystoNet accurately detected bladder tumours with 90.9% sensitivity and 98.6% specificity^8^. To understand how CystoNet performs in real-time clinical settings, the platform was evaluated in the operating room during TURBT and clinic cystoscopy^54^. Here, CystoNet obtained a per-frame specificity of 98.8% and 95.4% for cystoscopy and TURBT, respectively. With dynamic overlays of tumour-bounding boxes on cystoscopy videos, CystoNet shows promise in aiding clinicians during cystoscopy and TURBT in real-time^54^.
Tumour segmentation extends localization by outlining precise tumour borders (Fig. 3). Segmentation could help both diagnosis and treatment, enabling improved tumour resection during TURBT. A wide variety of classical and state-of-art segmentation deep-learning models have been applied to outline tumour borders^87,89,90^. For example, the effectiveness of an augmented version of the classical U-net model, in which image frames were flipped and randomly rotated to increase image variability, was validated using a small cystoscopy imaging dataset collected from 120 patients. Pixelwise sensitivity, specificity, positive predictive value and dice similarity coefficient were 84.9%, 88.5%, 86.7% and 83.0%, respectively, using this method^88^. These models have also been validated on large, multicentre datasets, containing over 10,000 patients, in which a masked residual CNN was used to segment tumours with 95.4% sensitivity and 90.2% accuracy^85^. The utility of tumour segmentation remains to be demonstrated; however, this technique could serve as a basis for evaluating the quality and thoroughness of tumour resection or guiding surgery.
Efforts have also been made to perform digital staining of WLC to obtain BLC-like frames using deep learning (Fig. 3) techniques called generative adversarial networks for style-transfer. A software-based alternative to blue light would not require additional hardware or photosensitizer, so it could serve as a cost-effective option towards improving the detection of papillary and flat lesions in the office and operating room^91,92^. Further investigation and validation are needed for digital staining to assess labelling accuracy and colour consistency, as published studies to date remain at an early stage and limited by a small sample size.
Incorporation of AI into modalities other than WLC has been investigated. For example, in a multicentre clinical study of 216 images from 216 patients, CNNs trained with BLC images demonstrated performance comparable with, if not superior to, expert urologists on a variety of clinical tasks^80^. End points included the assessment of malignancy, stage and grade; sensitivity and specificity of the CNNs for these end points were 95.77% and 87.84%, 88% and 95.56%, and 92.07% and 96.0%, respectively. When compared with the performance of expert urologists, these results are a 15% improvement for the classification of malignancy, 50% improvement for the classification of stage, and 40% improvement for the classification of grade. The use of a generative adversarial network model using WLC and NBI data for bladder tumour detection has also been investigated in a dataset consisting of 6,523 white light images and 1,636 NBI images. Using a definition of accurate detection as a >50% overlap between the AI estimated lesion area and the annotated lesion area, average sensitivity, specificity and F1 scores were 83.7%, 82.3% and 0.637, respectively^82^.
AI has demonstrated tremendous technical potential in accurate image analysis across modalities when focused on traditional metrics such as classification sensitivity and specificity or pixel area overlap. How this performance, measured under static conditions, translates to real-time use and what influence this performance has on the ultimately important end points of additional cancer detection, disease recurrence and usability by clinicians remains to be seen.
To maximize CAD and AI contributions clinically, advances are needed in two key first, the availability of accessible and accurately annotated representative frames and sequences for trainees and AI models; and second, methods to incorporate new types of data into the medical record. Suggested methods of improved cystoscopy data collection and storage include frameworks designed to automatically capture and barcode notable clips based on a surgeon’s screenshots of images and development of sharable data annotated on the case, lesion and frame level^68,69^. How mosaics or reconstructions fit into the medical record and can be communicated among physicians has also been investigated. Digital reporting software based on stitching to generate panoramic images from available cystoscopy videos, specialized cystoscopes paired with frames recreated in a 3D surface model, and advanced computer vision techniques for 3D reconstruction are some examples of methods aimed at creating photo-documentation; however, these methods involve either specialized software or hardware and lack easy integration into standard electronic health records^55,57,58^. With the promising potential of AI for bladder cancer detection, the next steps towards integration are incorporation into the clinical workflow, use at multiple clinical sites and evaluation of its handling of prospective data.
AI has several advantages over existing approaches to enhancing endoscopic bladder cancer diagnosis. The advantages of AI are that it learns rapidly, adapts to new data, and does not fatigue. CAD with AI can, therefore, help a practitioner to obtain the same diagnostic accuracy whether it is the first or last patient of the week, or whether they are an experienced urologist or a trainee.
Prospective evaluation is the next step towards clinical deployment. Formally characterizing clinical benefit is not straight-forward; unlike cancer treatment, a new diagnostic tool might not demonstrate improvement of cancer-specific outcomes such as recurrence and progression readily. With the above framework in mind, which short-term outcomes are appropriate to measure remain to be elucidated. Early-stage feasibility studies could focus on feedback from urologists to refine the user interface and evaluate subjective benefit. Studies in low-resource settings could evaluate if AI-enhanced endoscopy in the hands of local non-specialized practitioners or advanced practice providers can improve access to bladder cancer diagnosis and surveillance. Subsequent studies could compare the cost of deployment between AI-enhanced endoscopy and existing practices including WLC and BLC. This evaluation could include both financial costs and time spent per patient. Cancer detection, recurrence and ultimately progression rate can be compared for non-inferiority. In the context of TURBT, AI-enhanced cystoscopy might reduce the rate of residual disease on repeat resection by improving the quality of the index resection.
Considerable strides have been made to enhance cystoscopy via digitization, bladder mapping and AI, but several key bottlenecks, both technical and clinical, need to be overcome in order to successfully implement and deploy these systems at scale. Fortunately, the relatively advanced state of CAD and AI systems in other endoscopic applications, such as gastrointestinal endoscopy and colonoscopy, can serve as models for the development of AI systems for cystoscopy^93,94^.
From its inception, endoscopy has continuously evolved to meet the needs of affected patients. Cystoscopy has been instrumental in urology, particularly in the diagnosis and management of bladder cancer. Advances in technology and the addition of enhanced cystoscopy has further improved bladder cancer diagnostic capabilities, and numerous novel tools are under development to continue this push forward. However, the current state of cystoscopy and TURBT clinical workflows remains relatively similar to where they were shortly after the advent of the resectoscope, and generating further substantial improvement might require a paradigm shift; AI could be the impetus for this paradigm shift. CAD and AI have several unique advantages over existing approaches to enhancing endoscopic bladder cancer diagnosis and have already demonstrated promise in the identification of bladder tumours. However, these tools are at an early stage. Delineating the strengths — and more importantly the key limitations — ensures that the technology is deployed in a role that provides tangible patient benefit.