Computational pathology in bladder cancer: A scoping review

AI and digital pathology background

There are multiple strategies for using AI to extract meaningful histologic information from WSIs and develop predictive models. These methods generally fall into two broad categories: (1) human-interpretable feature-based models ⁶ and (2) weakly supervised learning (pure deep-learning based pattern recognition). ⁷ While both approaches leverage digitized pathology slides, they differ in terms of workflow, interpretability, performance, and resource requirements. We provide a summary of each method and then highlight key strengths and challenges associated with each approach. A schematic representation of these approaches is shown in Figure 1.

OPEN IN VIEWER

Figure 1.

Schematic overview of representative frameworks for (a) human-interpretable feature-based models and (b) weakly supervised learning-based models.

Eligibility criteria

Studies were eligible for inclusion if they met the following criteria:

Population/Sample: Human bladder cancer tissue specimens (WSIs).

Studies were excluded if they met any of the following criteria:

Studies involving digital cytology as the primary data source.

Search strategy

A systematic search was conducted in PubMed/MEDLINE on June 9, 2025. The full search strategy was developed iteratively by the study team. The final search string was: “((Artificial intelligence) OR (AI) OR (Machine Learning) OR (Deep Learning)) AND (Pathology) AND ((Bladder Cancer) OR (Urothelial Cancer) OR (Bladder Carcinoma) OR (Urothelial Carcinoma))”. No filters were applied, including language, publication year, or study design restrictions. Reference lists of included studies and key review articles were hand-searched to identify additional eligible publications.

Synthesis of results

All retrieved records were imported into Rayyan. ²⁴ Titles and abstracts were independently screened according to eligibility criteria. Full texts of potentially eligible studies were reviewed for inclusion. Included studies were categorized according to one of the primary predictive tasks addressed: diagnosis, grading, staging, molecular classification, prognostication, or treatment response prediction. Data collected from each study included the primary author, year of publication, predictive task-specific aim, data source, number of samples/patients, type of computational approach used (e.g., feature extraction and modeling approach), and model performance. Model performance was described using the primary metrics reported by the authors for the corresponding predictive task. A total of 2055 articles were screened. Fifty-three articles were included after full-text review. Findings were synthesized narratively with summaries organized by predictive task.

Diagnosis

The diagnosis of bladder cancer relies on pathologist review of histologic features from tumor specimens, often initially obtained via transurethral resection of bladder tumor (TURBT). This diagnostic process can be challenging and time consuming, requiring microscopic examination of numerous TURBT tissue chips. Several approaches have demonstrated promise in automating diagnostic evaluation of bladder biopsies using digital pathology (Table 1).

OPEN IN VIEWER

Table 1.

Publications using digital pathology to diagnose bladder cancer.

Authors	Year	Diagnosis-specific aim	Data source	# of samples (patients)	Feature extraction and modeling approach	Model performance
Chen et al. 25	2021	Bladder cancer vs normal bladder; Bladder cancer vs glandular cystitis	TCGA & Shanghai General Hospital	TCGA: 443 Shanghai: 200	CellProfiler-derived human interpretable features; LASSO analysis for feature selection and model construction.	Bladder cancer vs normal bladder: AUC: 0.941 Bladder cancer vs glandular cystitis: AUC: 0.938
Nishikawa et al. 26	2023	Normal urothelium vs atypical urothelium vs dysplasia vs carcinoma in situ	Wakayama Medical University	110 (98)	Custom segmentation (UroSegNet) to extract morphologic features; XGBoost for classification with SHAP-based interpretability.	Normal urothelium vs atypical urothelium: AUC: 0.72 Atypical urothelium vs dysplasia: AUC: 0.758 Dysplasia vs carcinoma in situ: AUC: 0.796
Zheng et al. 27	2022	Bladder cancer vs normal bladder	TCGA & Renmin Hospital of Wuhan University	TCGA: 926 (412) Renmin: 250 (150)	ResNet-50 feature extraction; weakly supervised multiple instance learning with attention for classification.	Bladder cancer vs normal bladder: AUC: 0.987
Jeong et al. 28	2024	Bladder tumor cell vs stromal cell vs immune cells	Asan Medical Center	199	Cell-level patches use to train and validate 9 independent CNNs for cell type classification.	Cell classification by EfficientNet (best performing CNN): Accuracy: 0.87 Sensitivity: 0.94 Specificity: 0.88 Precision: 0.90 F1 Score: 0.92 Neoplastic cell percentage estimation by EfficientNet: ICC with expert pathologists: 0.88

CNN: Convolutional Neural Network; LASSO: Least Absolute Shrinkage and Selection Operator; ICC: Interclass correlation coefficient; ResNet: Residual Neural Network; SHAP: Shapley Additive Explanations; TCGA: The Cancer Genome Atlas; XGBoost: Extreme Gradient Boosting.

Two groups have developed feature-based predictive models to discriminate WSIs containing bladder cancer from those that do not. Chen et al. used CellProfiler to extract quantitative features from H&E-stained WSIs and applied LASSO regression for automated feature selection and model development. ²⁵ Their model distinguished bladder cancer from normal urothelium or glandular cystitis with an AUC of 0.941 and AUC of 0.934, respectively. Similarly, Nishikawa et al. developed a custom segmentation network (UroSegNet) to extract seven biologically interpretable morphologic features, such as cellularity, nuclear polarity, and nuclear size variation, and trained interpretable models to classify flat urothelial lesions into normal urothelium, atypical urothelium, dysplasia, and carcinoma in situ (CIS), achieving near-perfect agreement with expert pathologists. ²⁶

Pure deep learning approaches to diagnose bladder cancer have also been developed at both the whole slide and patch levels. At the whole slide level, Zhang et al. used an attention-based, multiple instance learning (MIL) framework trained on patch features extracted from a pre-trained ResNet-50 model. ²⁷ Their model achieved excellent performance in identifying the presence or absence of bladder cancer on WSIs, with an accuracy of 0.987 and AUC of 0.993 in an external validation cohort. At the patch level, Jeong et al. developed a deep learning pipeline to classify cell patches as tumor, stromal, or immune, with the goal of estimating the neoplastic cell percentage in histologic samples. ²⁸ They trained nine different CNNs and demonstrated that their best performing models were comparable to expert pathologists and can inform downstream molecular evaluation, such as copy number variation (CNV) assessment, which is known to be influenced by non-neoplastic cell content.

Overall, these models demonstrate the potential of AI to distinguish between bladder cancer and non-cancerous tissue in pathology specimens. While not within the scope of this review, there have also been a multitude of AI-based approaches to evaluate urine cytology to aid in bladder cancer diagnosis.^29–32 These computational pathology tools have the potential to assist expert pathologists and expedite traditional diagnostic workflows.

Grade

Tumor grade plays a critical role in determining optimal treatment for bladder cancer, particularly in patients with non-muscle invasive bladder cancer (NMIBC), where grade is a key factor in risk stratification. Historically, bladder cancer grade was classified using a three-tier system (grades 1, 2, and 3) until the World Health Organization introduced a two-tier system (low-grade and high-grade) in 2004 to improve reproducibility and clinical relevance. ⁴⁵ Despite this change, grading remains largely subjective and is associated with substantial inter-observer and intra-observer variability.^46–49 As a result, there has been longstanding interest in using computational methods to automate objective measurement of histologic features to improve the consistency of bladder cancer grading (Table 2).

OPEN IN VIEWER

Table 2.

Publications using digital pathology to predict bladder cancer grade.

Authors	Year	Grade-specific aim	Data source	# of samples (patients)	Feature extraction and modeling approach	Model performance
Spyridonos et al. 33	2002	Bladder cancer grade (1 vs 2 vs 3)	University of Patras, Greece	92	Calculated nuclear features after nuclear segmentation by an ANN pixel-based classifier; grade prediction using ANN classifier in a 2-stage hierarchical tree.	Bladder cancer grades 1 & 2 vs grade 3:  Accuracy: 0.89 Bladder cancer grades 1 vs 2:  Accuracy: 0.79
Slotman et al. 34	2023	Non-invasive papillary urothelial carcinoma grade (high grade vs low grade)	Kingston Health Sciences Center	641 (371)	AI-based tissue and nuclear segmentation with morphometric feature extraction; classification using linear regression, logistic regression, decision tree, and random forest models.	High grade vs low grade (linear regression):  AUC: 0.94  Accuracy: 0.88
Zhang et al. 35	2019	Non-invasive low-grade papillary urothelial carcinoma vs non-invasive or invasive high-grade papillary urothelial carcinoma	TCGA & University of Florida Health Shands Hospital	913	Three neural networks: CNN for tumor detection, composite neural network combining CNN & LSTM for characterizing regions of interest (with natural language and attention maps), and fully-connected neural network for slide-level prediction (low grade vs high grade).	Non-invasive low-grade vs non-invasive or invasive high grade:  AUC: 0.97
Fuster et al. 36	2024	NMIBC low grade vs high grade	Stavanger University Hospital	300	Tissue segmentation algorithm to identify urothelium; extraction of patches at multiple magnification levels; nested attention based multiple instance learning for slide-level prediction.	Low grade vs high grade:  AUC: 0.94  Accuracy: 0.86  Precision: 0.87  Recall: 0.85  F1 Score: 0.85  κ: 0.71
Wetteland et al. 37	2021	NMIBC low grade vs high grade	Stavanger University Hospital	300	CNN-based tissue segmentation; 3 parallel VGG16 networks at varying magnification with global average pooling and fully connected classifier for patch-level prediction, aggregated into slide-level prediction.	Low grade vs high grade:  Precision: 0.92  Recall: 0.90  F1 Score: 0.90
Mundhada et al. 38	2022	Non-invasive papillary urothelial carcinoma low grade vs high grade	Sri Ramachandra Hospital	20	Patch extraction from WSIs at 2 sizes; CNN for patch-level classification; Grad-CAM for interpretability.	Low grade vs high grade (1024 × 1024 pixel patches):  Sensitivity: 0.88  Specificity: 0.87  Precision: 0.84  F1 Score: 0.86 Low grade vs high grade (256 × 256 pixel patches):  Sensitivity: 0.70  Specificity: 0.76  Precision: 0.70  F1 Score: 0.70
Pan et al. 39	2023	High grade muscle invasive vs high grade non-muscle invasive vs low grade non-muscle invasive bladder cancer	Sun Yat-sen Memorial Hospital	854 (692)	CNN for patch-level training and classification; probability weighted value of each subtype for WSI-level classification.	High grade muscle invasive vs high grade non-muscle invasive vs low grade non-muscle invasive (at patch level):  AUC: 0.878 High grade muscle invasive vs high grade non-muscle invasive vs low grade non-muscle invasive (at WSI level):  AUC: 0.847  Accuracy: 0.793  Sensitivity: 0.797  Specificity: 0.899  PPV: 0.816  NPV: 0.902
Shalata et al. 40	2024	NMIBC low grade vs high grade	Urology & Nephrology Center at Mansoura University	21	Multi-scale pyramidal pretrained CNN at 3 patch sizes integrated with random forest algorithm for WSI-level classification.	Low grade vs high grade:  Accuracy: 0.943  Sensitivity: 0.945  Specificity: 0.940  F1 Score: 0.943
Wenger et al. 41	2022	Bladder cancer grade (low grade 1 vs low grade 2 vs high grade 3 vs high grade 4 vs stroma)	Mount Sinai Hospital, Toronto, Ontario, Canada	13	Modified VGG16 model trained on grayscale image patches using consistency regularization and self-ensembling for semi-supervised patch level classification.	Bladder cancer grade (grade 1 vs 2 vs 3 vs 4 vs stroma), fully supervised (1000 labeled patches):  Accuracy: 0.696  Precision: 0.708  Recall: 0.702  F1 Score: 0.696 Bladder cancer grade (grade 1 vs 2 vs 3 vs 4 vs stroma), semi-supervised (1000 labeled patches and 29,000 unlabeled patches):  Accuracy: 0.881  Precision: 0.886  Recall: 0.882  F1 Score: 0.882 Bladder cancer grade (grade 1 vs 2 vs 3 vs 4 vs stroma), semi-supervised (16,000 labeled patches and 14,000 unlabeled patches):  Accuracy: 0.960  Precision: 0.962  Recall: 0.958  F1 Score: 0.960
Garcia et al. 42	2021	Non-tumor vs mild vs infiltrative patterns in MIBC	Hospital Universitario y Politécnico La Fe	136	Deep Convolutional Embedded Attention Clustering (DCEAC): a two-step unsupervised self-learning framework using a convolutional autoencoder with attention module and clustering branch for classification.	Non-tumor vs mild vs infiltrative MIBC (macro-averaged):  AUC: 0.878  Accuracy: 0.903  Sensitivity: 0.840  Specificity: 0.915  F1 Score: 0.847
Hu et al. 43	2025	Ta/T1 NMIBC low grade vs high grade using hyperspectral imaging	Shandong First Medical University	120	ResNet-34 and Vision Transformer feature extraction with cross-attention multiscale feature fusion and slide-level classification using a Kernel Adaptive Network.	Ta/T1 low grade vs high grade:  Accuracy: 0.941  Precision: 0.917  Recall: 0.906  F1 Score: 0.937
Kujdowicz et al. 44	2023	Low grade vs high grade bladder cancer using infrared microscopy	Jagiellonian University Medical College	100 (50)	Infrared spectral feature extraction; classification using partial least squares–discriminant analysis with repeated random sampling double cross-validation.	Low grade vs high grade:  AUC: 0.69  Accuracy: 0.64  Sensitivity: 0.65  Specificity: 0.63

ANN: Artificial Neural Network; CNN: Convolutional Neural Network; Grad-CAM: Gradient-weighted Class Activation Mapping; LSTM: Long Short-Term Memory (a type of Residual Neural Network); NMIBC: Non-muscle invasive bladder cancer; ResNet: Residual Neural Network; TCGA: The Cancer Genome Atlas; VGG: Visual Geometry Group Network.

Early efforts by Spyridonos and colleagues laid the foundation for AI in bladder cancer grading using human-interpretable features. In 2002, they developed an artificial neural-network for nuclear segmentation and grade prediction using morphometric features^33,50 and later expanded upon this work, comparing support vector machines and probabilistic neural networks for grade classification. ⁵¹ While similar human-interpretable feature-based approaches are still being used for automated bladder cancer grading, such as by Slotman et al. in 2023, ³⁴ recent work has increasingly shifted towards pure deep learning-based methods.

Pure deep-learning learning approaches to bladder cancer grading have increasingly relied on sophisticated machine learning architectures. Zhang et al. developed a pipeline involving three sequential neural networks to detect tumor tissue, characterize regions of interest, and generate slide-level predictions of grade. Their model outperformed the average performance of 17 pathologists, differentiating between high and low grade tumors with an AUC of 0.97, and incorporated attention heatmaps and natural language generation to enhance interpretability for human users. ³⁵ Engan and colleagues developed two other deep learning models for bladder cancer grading with innovative architectures: NMGrad and TRIGRADE. NMGrad segments urothelium from WSIs, extracts image patches at 3 magnification levels, and then applies a nested, attention-based multiple-instance learning architecture for slide-level grade prediction. ³⁶ In contrast, TRIGRADE also begins with urothelium segmentation, but processes tile triplets at three magnifications through three parallel CNNs, generating patch-level grade predictions, which are then aggregated using a decision threshold to yield slide-level grade predictions. ³⁷ While these approaches were all weakly supervised, meaning that they were trained using slide-level labels, strongly supervised CNN-based models that leverage patch and pixel-level labels have also proven useful for bladder cancer grading.^38–40

Given the time and cost of image annotation required for supervised machine learning methods, semi-supervised and unsupervised models, which can leverage larger datasets of unlabeled images in addition to a smaller number of labeled images, have also been explored for bladder cancer grading. Wegner et al. demonstrated a semi-supervised approach in which they labeled only 3% of data, but were able to improve accuracy from 69.6% to 88.2% after leveraging the remaining 97% of unlabeled cases using a consistency-regularized, self-ensembling VGG16-based model. ⁴¹ Garcia et al. also developed a fully unsupervised self-learning framework to grade muscle invasive bladder cancer cases from immunohistochemical (IHC) histological images stained for cytokeratin (AE3/1). ⁴²

Beyond H&E and IHC, additional imaging modalities are being evaluated to improve automation of bladder cancer grading, such as hyperspectral imaging ⁴³ and infrared microscopy. ⁴⁴ Together, these research efforts highlight the growing potential of computational pathology methods to support expert pathologists in bladder cancer grading.

Although a wide range of computational approaches have been explored for automated bladder cancer grading, several clear themes emerge. The best performing models generally use a multi-stage workflow, beginning with tumor segmentation and region-of-interest selection, followed by a deep learning approach for predictive modeling. Models that incorporate image patches at multiple resolutions generally outperform single resolution methods, suggesting that histologic findings relevant to grading exist at different scales. Finally, there has been a clear shift from human interpretable feature-based models toward deep learning models, which result in superior performance. Given substantial inter-observer and intra-observer variability in grading, automated computational pathology approaches have potential to support more objective, reliable measurements of bladder cancer grade.

Stage

Accurate tumor staging is essential for guiding treatment decisions and informing prognosis in bladder cancer. However, precise staging by conventional histologic evaluation can be challenging. Recent advances in computational pathology have the potential to improve staging accuracy and assist expert pathologists (Table 3).

OPEN IN VIEWER

Table 3.

Publications using digital pathology to predict bladder cancer stage.

Authors	Year	Stage-specific aim	Data source	# of samples (patients)	Feature extraction and modeling approach	Model performance
Yin et al. 52	2020	Ta vs T1	University of Rochester Medical Center	1177	Human-interpretable feature extraction; classification with probabilistic neural network (PNN), support vector machine, logistic regression, bagging, random forest, and multilayer perceptron.	Ta vs T1 (PNN): AUC: 0.99 Accuracy: 0.97
Niazi et al. 53	2020	Segmentation of urothelium, lamina propria, muscularis propria, and muscularis mucosa	Ohio State University	54	Multi-class, pixel-wise classification using deep learning U-Net CNN.	Classification of tissue layers: Accuracy (by tissue type): Background: 0.999 Lamina Propria: 0.977 Muscularis Propria: 0.975 Mucosa: 0.936 RBC: 0.846 Cautery: 0.551 Inflammation: 0.757
Subramanya et al. 54	2023	Segmentation of muscularis propria vs non-muscularis propria	University of Rochester Medical Center	277 (16)	Classification with CNNs at the patch-level (VGG16, ResNet-18, SqueezeNet, and MobileNetV2) and CNN-based semantic segmentation models at the pixel-level (U-Net, MA-Net, DeepLabv3+, and FPN).	Muscularis propria vs non-muscularis propria (patch-based approach evaluated at pixel-level using MobileNetV2): AUC: 0.87 Accuracy: 0.87 Specificity: 0.85 Precision: 0.73 Recall: 0.89 F1 Score: 0.80 Muscularis propria vs non-muscularis propria (pixel-based approach, evaluated at pixel-level using U-Net): AUC: 0.92 Accuracy: 0.89 Specificity: 0.88 Precision: 0.77 Recall: 0.90 F1 Score: 0.83
Fahoum et al. 55	2024	Segmentation of muscularis propria and tumor; Calculation of proximity	Tel-Aviv Sourasky Medical Center	1639 (187)	Hierarchical Contextual Analysis 56 framework with U-Net CNN; unsupervised initialization and supervised refinement using manual annotations.	Muscularis propria detection: Sensitivity: 0.72 Specificity: 0.46 Tumor detection: Sensitivity: 0.65 Specificity: 0.77
Zheng et al. 57	2023	Predicting lymph node metastasis from primary tumor slides	TCGA, Renmin Hospital of Wuhan University (RHWU), & People's Hospital of Hanchuan City (PHHC)	TCGA: 358 (323) RHWU: 417 (139) PHHC: 230 (78)	Weakly supervised learning using patch-level features extracted by ResNet-50 and an attention-based multiple instance learning architecture.	Lymph node metastasis prediction from primary tumor slides: AUC (TCGA): 0.864 AUC (RHWU): 0.810 AUC (PHHC): 0.824
Harmon et al. 58	2020	Predicting lymph node metastasis from primary tumor slides	TCGA & In-house	TCGA: 294 In-house: 13	Multiresolution deep learning using patch-level ResNet-101 and pre-trained lymphocyte infiltration classifier 59 , integrated via an AdaBoost ensemble classifier; clinicopathologic features incorporated using logistic regression.	Lymph node metastasis prediction from primary tumor slides: AUC: 0.807
Wu et al. 60	2023	Predicting lymph node metastasis from lymph node slides	Nanfang Hospital of Southern Medical University (NFH), The Third Affiliated Hospital of Southern Medical University (SMUTH), Sun Yat-sen Memorial Hospital of Sun Yat-sen University (SYSMH), The Third Affiliated Hospital of Sun Yat-sen University (SYUTH), Zhujiang Hospital of Southern Medical University (ZJH)	7991 (998)	High-Resolution Net CNN for strongly supervised patch-level prediction.	Lymph node metastasis prediction from lymph node slides (reported as a range across all 5 cohorts): AUC: 0.978–0.998 Accuracy: 0.904–0.987 Sensitivity: 0.958–1.000 Specificity: 0.901–0.989 PPV: 0.283–0.878 NPV: 0.997–1.000

AdaBoost: Adaptive Boosting; FPN: Feature Pyramid Network; MA-Net: Multi-scale attention network; RBC: Red Blood Cell; ResNet: Residual Neural Network; TCGA: The Cancer Genome Atlas; VGG: Visual Geometry Group Network.

Molecular subtype

Intrinsic molecular subtypes of bladder cancer, defined by bulk RNA sequencing, have prognostic significance and potential implications for response to both chemotherapy and immunotherapy.^67,68 Predicting molecular subtypes from H&E-stained WSIs using AI offers a compelling alternative to sequencing, given that RNA sequencing is costly, not routinely performed, and molecular subtype distinctions are not visually apparent to expert pathologists.^62,68 Computational pathology approaches to predict molecular subtypes in bladder cancer are summarized in Table 4.

OPEN IN VIEWER

Table 4.

Publications using digital pathology to predict bladder cancer molecular subtype.

Authors	Year	Molecular subtype-specific aim	Data source	# of samples (patients)	Feature extraction and modeling approach	Model performance
Zheng et al. 62	2025	Basal vs luminal subtype	TCGA, Shanghai Tenth People's Hospital (STPH), The Affiliated Guangdong Second Provincial General Hospital of Jinan University (GD2H)	TCGA: 449 (378) STPH: 179 (163) GD2H: 53 (38)	Classification of tumor vs normal patches with ResNet-50; patch-level deep-learning feature extraction with RetCCL model using (1) all patches and (2) tumor only patches from ResNet-50 model; patch aggregation to slide-level features and prediction with logistic regression, random forest and support vector machine.	Basal vs luminal subtype: AUC: 0.85 Accuracy: 0.74 Sensitivity: 0.66 Specificity: 0.82 PPV: 0.79 NPV: 0.71
Woerl et al. 63	2020	MD Anderson molecular subtype (basal vs luminal vs luminal p53-like vs double negative)	TCGA, Comprehensive Cancer Center Erlangen Metropol Region Nuremberg MIBC cohort (CCC-EMN)	TCGA: 407 CCC-EMN: 16	Patch-level predictions using a CNN with a ResNet50 base; slide-level predictions using weighted majority of patch predictions.	MD Anderson molecular subtype (basal vs luminal vs luminal p53-like vs double negative): AUC (micro-average): 0.89 AUC (macro-average): 0.87 Accuracy: 0.70
Stoll et al. 64	2024	Simplified consensus molecular subtype (basal vs stroma-rich vs merged luminal)	TCGA, Dr. Senckenberg Institute of Pathology (SIP)	TCGA: 292 SIP: 103	Patch-level classifications generated using a ResNet-18 CNN; slide-level prediction obtained using an average of patch predictions.	Simplified consensus molecular subtype (basal vs stroma-rich vs merged luminal): AUC (TCGA 5-fold cross validation): 0.73 AUC (TCGA test set): 0.80 AUC (SIP test set): 0.75
Bai et al. 65	2022	IHC-defined basal vs luminal subtype; p53 status, PD-L1 expression	The First Affiliated Hospital, Zhejiang University School of Medicine	119	Weakly supervised segmentation of tumor regions; patch-level prediction of molecular subtype, p53, and PD-L1 using 3 distinct ResNet-50 CNNs; slide-level prediction by average score of tumor patches.	Basal vs luminal: Accuracy: 0.946 Sensitivity: 1.000 Specificity: 0.909 PD-L1 status: Accuracy: 0.897 Sensitivity: 0.875 Specificity: 0.913 p53 status: Accuracy: 0.846 Sensitivity: 0.857 Specificity is 0.750
Hamidi et al. 66	2024	Novel custom molecular classification (luminal-desert vs basal vs stromal vs immune)	Four clinical trials (IMvigor210/211/130/010)	IMvigor210: 354 IMvigor211: 793 IMvigor130: 928 IMvigor010: 728	Human interpretable features (HIF): Deep learning models from PathAI for HIF extraction; random forest for slide-level prediction. Self-supervised deep learning: Contrastive self-supervised learning model (RetCCL with ResNet-50 backbone) for patch-level feature extraction; CLAM-based attention model for slide-level prediction.	Novel custom molecular classification (luminal-desert vs basal vs stromal vs immune): AUC (HIF-based model): 0.72 AUC (Deep learning model): 0.86

CLAM: Clustering-constrained Attention Multiple Instance Learning; IHC: Immunohistochemistry; PD-L1: Programmed Death-Ligand 1; ResNet: Residual Neural Network; RetCCL: Retrieval with Clustering-guided Contrastive Learning; TCGA: The Cancer Genome Atlas.

While there are many molecular classification systems for bladder cancer, early classification systems focused on basal and luminal subtypes.^68–71 Zheng et al. used two deep learning pipelines—one incorporating both tumor and normal tissue, and another using tumor-only regions—to predict basal and luminal subtype with an optimal AUC of 0.85 (95% CI, 0.77–0.92) surpassing the performance of expert pathologists. ⁶² Woerl et al. developed a similar approach to predict MD Anderson molecular subtype classifications (basal, luminal, luminal p53-like, and double negative), ⁶³ while Stoll et al. developed a model for a simplified consensus subtype classification (basal, stroma-rich, and merged luminal subtypes). ⁶⁴ Bai et al. also trained a deep learning model on digital H&E images to predict IHC-defined basal/luminal subtypes, p53 status, and PD-L1 expression. ⁶⁵

While these approaches rely on existing molecular classification systems, Hamidi et al. proposed a novel 4-class system, including luminal-desert, basal, stromal, and immune molecular subtypes, based on a clustering approach using RNA-sequencing data from four late-phase clinical trials. They demonstrated superior overall survival in patients with basal and immune subtypes, and successfully predicted these subtypes from digital H&E slides using a self-supervised learning approach. ⁶⁶

These studies highlight the potential of digital pathology as a surrogate for transcriptomic profiling in bladder cancer, enabling molecular subtype prediction and its associated prognostic insights without direct RNA sequencing. This is of particular relevance given that pathology slides are obtained as standard of care, whereas RNA sequencing is not routinely performed in the diagnostic evaluation of bladder cancer. Similar to computational pathology approaches used for bladder cancer grading, the best performing models for molecular subtype prediction use multi-stage pipelines that start with tumor segmentation. Performance of these models decline when classifying more granular, multi-class molecular subtypes, likely due both to the probabilistic and imperfect nature of molecular subtype classification and the natural limits of image-based models in estimating transcriptomic profiles, particularly from bulk RNA sequencing. Integrating spatial transcriptomics or multiplex immunofluorescence data with digital pathology holds promise for identifying more precise H&E-based correlates of molecular subtype and represents an important direction for future work.

Prognosis

Computational pathology has enabled the extraction of prognostic information directly from H&E-stained WSIs, offering an additional objective source of data to inform risk stratification in bladder cancer. Computational pathology approaches for prognostication range from direct prediction of cancer recurrence and overall survival to the identification of novel histologic biomarkers with prognostic significance. Representative studies that have used digital pathology to inform prognosis are summarized in Table 5.

OPEN IN VIEWER

Table 5.

Publications using digital pathology to predict bladder cancer prognosis.

Authors	Year	Progonsis-specific aim	Data source	# of samples (patients)	Feature extraction and modeling approach	Model performance
Lucas et al. 72	2022	Predict 1- and 5-year recurrence-free survival in NMIBC	Amsterdam UMC	359	U-Net CNN for pixel-wise segmentation of urothelium to facilitate patch selection, with VGG16 CNN for patch-level feature extraction, combined with clinical data using a bidirectional GRU for classification.	1-year recurrence prediction in NMIBC: AUC: 0.62 Accuracy: 0.65 Sensitivity: 0.30 Specificity: 0.80 PPV: 0.40 NPV: 0.73 5-year recurrence prediction in NMIBC: AUC: 0.76 Sensitivity: 0.89 Specificity: 0.57 Accuracy: 0.74 PPV: 0.71 NPV: 0.81
Wang et al. 73	2024	Predict 1-year recurrence in NMIBC	Xuzhou Central Hospital and Suqian Affiliated Hospital of Xuzhou Medical University	210	Patch-based prediction using a ResNet50 model; Patch Likelihood Histogram (PALHI) and Bag of Words (BoW) to generate feature vectors; ensemble random forest classifier for WSI-level prediction.	1-year recurrence prediction in NMIBC: AUC: 0.860 Accuracy: 0.811 Sensitivity: 0.90 Specificity: 0.651 PPV: 0.500 NPV: 0.884 F1 Score: 0.643
Drachneris et al. 74	2024	Predict 2- and 5-year recurrence-free survival in NMIPUC patients treated with BCG	Vilnius University Hospital Santaros Klinikos	157. 981 independent slides for feature extraction only	Feature extraction via Inception-ResNet-v2 with multiresolution patch sampling; deep learning using both multiple instance learning and multiple instance survival learning for WSI-level prediction.	2-year recurrence prediction in NMIPUC patients treated with BCG (256 × 256 pixel patch): Accuracy: 0.620 F1 Score: 0.626 5-year recurrence prediction in NMIPUC patients treated with BCG (256 × 256 pixel patch): Accuracy: 0.490 F1 Score: 0.480
Tokuyama et al. 75	2022	Predict early recurrence in NMIBC	Tokyo Medical University Hospital	125	Automated nucleus segmentation; CellProfiler for nuclear feature extraction; cell feature level co-occurrence matrix for nuclear heterogeneity and pleomorphism measurements; support vector machine (SVM) and random forest (RF) for classification.	2-year recurrence prediction in NMIBC: Accuracy (SVM): 0.900 Accuracy (RF): 0.867
Qaiser et al. 76	2022	Predict overall survival in stage II–IV bladder cancer	In house	198 bladder cancer patients	Feature extraction using ResNet-18, VGG-11bn (batch normalization), DenseNet, and AlexNet models; weakly supervised survival CNN with visual attention for prediction.	Overall survival in stage II-IV bladder cancer comparing AI-based low and high-risk populations: HR: 1.96 (p = 0.0179) C-index: 0.612
He et al. 77	2024	Predict overall survival across all stages of bladder cancer	First Affiliated Hospital of Chongqing Medical University (CMUFH), the Second Affiliated Hospital of Chongqing Medical University (CMUSH), and the Yongchuan Hospital of Chongqing Medical University (YCH), TCGA	CMUFH: 887 CMUSH: 113 YCH: 108 TCGA: 375	ResNeXt-50 CNN for patch-level tissue segmentation and classification with generation of multi-class tissue probability heatmaps (BlaPaSeg); overall survival prediction using a modified ResNeXt-50 CNN applied to macro-level features from BlaPaSeg heatmaps (MacroVisionNet); overall survival prediction using an integrated model combining macro- and micro-level features via a multiple instance learning (MIL)-based attention architecture (UniVisionNet).	Overall survival comparing AI-based low and high-risk populations (UniVisionNet): CMUFH (validation only):  HR (univariate analysis): 6.57  (p < 0.001)  C-index: 0.797  AUC (at 1 year): 0.782  AUC (at 3 years): 0.819  AUC (at 5 years): 0.845 CMUSH:  HR (univariate analysis): 10.59  (p < 0.001)  C-index: 0.811  AUC (at 1 year): 0.889  AUC (at 3 years): 0.835  AUC (at 5 years): 0.822 YCH:  HR (univariate analysis): 18.21  (p < 0.001)  C-index: 0.820  AUC (at 1 year): 0.830  AUC (at 3 years): 0.835  AUC (at 5 years): 0.916 TCGA:  HR (univariate analysis): 2.20  (p < 0.001)  C-index: 0.661  AUC (at 1 year): 0.674  AUC (at 3 years): 0.662  AUC (at 5 years): 0.678
Jiang et al. 78	2022	Develop WSI-based clusters to predict overall survival	TCGA, Gene Expression Omnibus (GEO), E-MTAB-4321, IMvigor210	TCGA: 435 (363) GEO: 389 IMvigor210: 348 E-MTAB-4321: 476	Feature extraction using pretrained Inception V3 model, flatten features, and histogram features; clustering on features using mini batch K-means, hierarchical clustering, and Gaussian mixture model; clusters used as labels for a weakly supervised deep learning classifier.	Overall survival comparing low and high-risk populations: AUC (at 1 year): 0.838 AUC (at 2 years): 0.817 AUC (at 5 years): 0.847
Barrios et al. 79	2022	Low risk vs high risk	New Hampshire Bladder Cancer Study (NHBCS), TCGA	NHBCS: 607, TCGA: 378	Ensemble of four binary ResNet-18 CNN patch-level classifiers; aggregation of patch-level predictions with a Gaussian process classifier for whole-slide prediction.	Low vs high risk prediction: NHBCS:  Accuracy: 0.91  Precision: 0.91  Recall: 0.91  F1 Score: 0.91 TCGA:  Accuracy: 0.99  Precision: 0.99  Recall: 0.99  F1 Score: 0.99 Overall survival comparing low vs high risk populations: HR: 1.958 (p = 0.005)
Xu et al. 80	2022	Predict TMB status, spatial TMB heterogeneity, and TILs; evaluate utility in predicting overall survival	TCGA	TCGA: 457 (386)	TMB prediction: CNN-based tumor detection with affinity propagation clustering for representative patch identification, feature extraction using a pre-trained Xception model with weighted aggregation for patient-level representations; support vector machine for TMB classification; Shannon entropy of predicted TMB class for spatial heterogeneity TIL detection: ResNet-18 for patch-level TIL classification.	TMB prediction: AUC: 0.752 Accuracy: 0.731 Sensitivity: 0.705 Specificity: 0.758 TIL detection: Accuracy: 0.800 Overall survival comparing TIL-high, TMB-high, and low spatial heterogeneity (HHL) vs TIL-low, TMB-high, and low spatial heterogeneity (LHL) vs other patients: Log-rank p = 0.0027
Zheng et al. 81	2022	Develop quantitative TIL score from WSI; evaluate utility in predicting overall survival in MIBC	TCGA, Renmin Hospital of Wuhan University	TCGA: 247 RHWU: 133	Watershed cell detection for cell segmentation; deep learning neural network for cell classification; quantitative TIL assessment for prognostic modeling.	Overall survival comparing high TILs vs low TILs: HR (TCGA): 0.333 (p < 0.0001) HR (RHWU): 0.264 (p < 0.0001)
Van Rijthoven et al. 82	2024	Develop quantitative TLS and GC from WSI; evaluate utility to predict overall survival in MIBC	TCGA	345	Multi-resolution feature extraction with a dual-branch U-Net deep learning architecture for TLS identification at the pixel-level with thresholding for WSI- level segmentation and object identification; TLS and GC density assessment for prognostic modeling.	Overall survival comparing high TLS density vs low TLS density: Log-rank p = 0.81 Overall survival comparing high GC density vs low GC density: Log-rank p = 0.46
Gavriel et al. 83	2021	Develop quantitative markers related to lymphocytes, macrophages, tumor buds, and PD-L1; evaluate utility in predicting 5 year survival in MIBC	Royal Infirmary and Western General Hospital in Edinburgh	78	Machine learning-based image analysis for cell localization and quantitative feature extraction; decision tree, random forest, support vector machine, logistic regression, K-nearest neighbor classifier, and an ensemble classifier.	Overall survival at 5 years: AUC: 0.893 Accuracy: 0.800 Sensitivity: 1.000 Specificity: 0.714 F1 Score: 0.833 Overall survival comparing low vs high risk populations: HR: 0.325
Debatin et al. 84	2024	Develop quantitative and spatial markers related to immune cell populations; evaluate utility in predicting overall survival	Institute of Pathology, University Hospital Hamburg	2463	U-Net for cell segmentation, DeepLabv3 + deep learning model for multiplex immunofluorescent marker detection and tumor area quantification, and unsupervised X-shift clustering for immune and tumor cell subpopulation classification; unsupervised hierarchical clustering for phenotyping and subsequent prognostic evaluation.	Overall survival at 4 years comparing high vs low intra-intraepithelial CD8+ T-cell density: AUC: 0.70
Chen et al. 85	2021	Develop a pathomics-based molecular signature to predict overall survival	Shanghai General Hospital (SGH), TCGA	SGH: 102 TCGA: 406	Morphologic features were extracted from WSIs; LASSO regression to select prognostic features for prognostic modeling.	Overall survival comparing high vs low risk populations: HR (TCGA): 2.19 (p < 0.0001) HR (SGH): 3.20 (p = 0.0014)
Zheng et al. 86	2023	Quantify tumor-stroma ratio; evaluate utility in predicting overall survival in MIBC	Renmin Hospital of Wuhan University, TCGA	TCGA: 261 RHWU: 133	Watershed cell detection method for cell segmentation; deep neural network for cell classification with automated tumor-stroma ratio quantification used for prognostic modeling.	Quantification of tumor-stroma ratio: Spearmen correlation (AI vs human): 0.911 Overall survival comparing high vs low tumor-stroma ratio: HR (TCGA): 2.346 (p < 0.0001) HR (RHWU): 2.851 (p < 0.0001)

BCG: Bacillus Calmette-Guérin; DenseNet: Densely Connected Convolutional Network; GC: Germinal Center; GRU: Gated Recurrent Unit (a type of recurrent neural network); HR: Hazard Ratio; LASSO: Least Absolute Shrinkage and Selection Operator; MIBC: Muscle Invasive Bladder Cancer; NMIBC: Non-Muscle Invasive Bladder Cancer; NMIPUC: Non-Muscle Invasive Papillary Urothelial Carcinoma; ResNet: Residual Neural Network; TIL: Tumor-Infiltrating Lymphocyte; TMB: Tumor Mutational Burden; TLS: Tertiary Lymphoid Structure; VGG: Visual Geometry Group; Xception: Extreme Inception.

Digital pathology-based prognosis prediction

Recent studies have used machine learning on digital WSIs to predict cancer recurrence and overall survival in bladder cancer. In patients with NMIBC, both feature-based and pure deep learning models have been used to predict recurrence after TURBT.^72–75 For more advanced disease, weakly supervised deep learning approaches have been employed to predict overall survival in stage II–IV bladder cancer ⁷⁶ and across all stages. ⁷⁷

Several groups have also used machine learning to define novel prognostic categories directly from H&E-stained images. Jiang et al. used an unsupervised deep learning-based clustering approach to group H&E-stained WSIs into three distinct categories, revealing significant differences in overall survival between WSI-based clusters. ⁷⁸ Similarly, Barrios et al. developed a strongly supervised model with patch-level annotations to classify patients into low and high risk groups, and showed that these classifications were associated with differences in overall survival. ⁷⁹

Histologic biomarkers of prognosis

Numerous supervised computational approaches have also been used to identify histology-based biomarkers with prognostic significance. These can be broadly categorized into models focused on immune-related and non-immune histologic patterns.

Several studies have sought to identify prognostic biomarkers related to the tumor immune microenvironment. Xu et al. developed a deep learning pipeline to predict tumor mutation burden (TMB) and the spatial distribution of tumor-infiltrating lymphocytes (TILs), demonstrating that tumors with high TMB, low spatial heterogeneity, and high TIL density were associated with improved survival. ⁸⁰ Similarly, Zheng et al. showed that in MIBC, higher TIL counts, quantified using an automated machine learning algorithm, were associated with better prognosis. ⁸¹ A deep learning model developed by van Rijthoven et al. for the automated quantification of tertiary lymphoid structure (TLS) achieved human-level performance across multiple cancer types, including bladder cancer. ⁸² However, unlike TIL-based models, their model did not identify prognostic differences in bladder cancer based on calculated TLS density.

Beyond H&E-based analyses, there is growing interest in using multiplex immunofluorescence to identify immune-specific prognostic biomarkers. Gavriel et al. applied machine learning to WSIs stained with multiplex immunofluorescence to quantify spatial and immune features, including lymphocytes, macrophages, tumor buds and PD-L1, and demonstrated that a model incorporating these features outperformed TNM stage in predicting 5-year prognosis in MIBC. ⁸³ Similarly, Debatin et al. used a deep neural network-based approach with multiplex fluorescence immunohistochemistry to identify 16 immune parameters with prognostic significance in MIBC, independent of T stage, N stage, and grade. ⁸⁴ Among these, intraepithelial CD8+ T-cell density showed the strongest association with overall survival.

Other groups have investigated non-immune histologic biomarkers to predict overall survival. Chen et al. developed a pathomics-based machine learning signature incorporating features such as nuclear area, circularity, and cell size to predict overall survival. ⁸⁵ Zheng et al. created an algorithm to quantify tumor-stroma ratio in MIBC, showing that a higher ratio correlated with improved survival. ⁸⁶

Altogether, these advances demonstrate that computational pathology has the potential to identify novel prognostic features and improve risk stratification in bladder cancer. Although deep learning-based approaches generally outperform models built on predefined, human-interpretable features for prognostication, the latter remain valuable for isolating individual prognostic biomarkers, such as TILs or intraepithelial CD8+ T-cell density. Of note, many of the prognostic models in this review do not account for key variables such as age, stage, treatment history, or comorbidities, limiting conclusions about their clinical utility. When integrated with established prognostic factors, digital pathology can help improve prognostic precision and inform clinical decision-making.

Treatment response prediction

Predicting treatment response using digital pathology holds promise for optimizing patient care by guiding therapy selection and minimizing unnecessary treatment-related toxicity. Computational pathology may be used to directly predict response to specific therapies, such as intravesical BCG for NMIBC or neoadjuvant chemotherapy for MIBC, or to identify predictive biomarkers of treatment response. Studies evaluating these strategies are summarized in Table 6.

OPEN IN VIEWER

Table 6.

Publications using digital pathology to predict treatment response in bladder cancer.

Authors	Year	Treatment response-specific aim	Data source	# of samples (patients)	Feature extraction and modeling approach	Model performance
Jiang et al. 87	2025	Predicting response to intravesical BCG in NMIBC	SunYat-sen Memorial Hospital of Sun Yat-sen University, The Fifth Affiliated Hospital of Sun Yat-sen University, Shenshan Medical Centre, Zhujiang Hospital of Southern Medical University, the Third Affiliated Hospital of Sun Yat-sen University	4395 (1275)	Patch-level feature extraction using a pre-trained model, aggregated into case-level representations; ensemble model averaging the output of a multiple instance learning framework and top-N patch prediction.	Response to intravesical BCG in NIMBC: Accuracy: 0.841 Sensitivity: 0.690 Specificity: 0.883
Lotan et al. 88	2025	Predicting recurrence, progression, BCG-unresponsive disease, and cystectomy in patients with high-risk NMIBC treated with intravesical BCG	12 academic centers	944	Computational histology AI (CHAI) platform: computed features extracted after strongly supervised deep learning–based segmentation to identify nuclei and tissue regions; cox proportional hazards regression for predictive modeling.	High-grade recurrence-free survival in high vs low risk populations: HR: 2.08 (p < 0.0001) Progression-free survival in high vs low risk populations: HR: 3.87 (p < 0.001) Cystectomy risk in high vs low risk populations: HR: 3.35 (p < 0.0001) Time to BCG-unresponsive disease in high vs low risk populations: HR: 2.31 (p < 0.0001)
Packiam et al. 89	2025	Evaluation of an AI-based histologic biomarker in predicting response to BCG and Gemcitabine/Docetaxel in high-grade NMIBC	University of Iowa and University of Kentucky	253	CHAI platform.	High-grade recurrence free survival in CHAI biomarker positive patients treated with BCG vs Gemcitabine/Docetaxel: HR: 5.4 (p = 0.007) High-grade recurrence free survival in CHAI biomarker negative patients treated with BCG vs Gemcitabine/Docetaxel: HR: 1.3 (p = 0.5)
Mi et al. 90	2021	Prediction of cisplatin-based neoadjuvant chemotherapy response in MIBC	Johns Hopkins University & University Hospital Bern	Hopkins: 66 Bern: 56	Deep learning and watershed algorithms for cellular segmentation; computed human interpretable features at region, cluster, and cell-levels; support vector machine and random forest for classification.	Prediction of cisplatin-based neoadjuvant chemotherapy response in MIBC: AUC: 0.67 Accuracy: 0.73
Bai et al. 91	2025	Prediction of cisplatin-based neoadjuvant chemotherapy response in MIBC	SWOG1314 Clinical Trial	180	Classification using a graph-based multimodal deep learning model with late-fusion that integrates: (1) GNN-based branch using ResNet-50 image feature embeddings, a (2) GNN-based branch using human-interpretable HoVer-Net-derived morphologic features, and (3) a multilayer perceptron to process gene expression data.	Prediction of cisplatin-based neoadjuvant chemotherapy response in MIBC (multimodal model): AUC 0.72
Velmahos et al. 92	2021	Prediction of FGFR mutation status	TCGA	290	Pre-trained CNN for patch-level TIL detection with automated calculation of slide-level TIL percentage; logistic regression to predict mutation status.	Prediction of FGFR gene aberrations: AUC = 0.76 Sensitivity = 0.89 Specificity = 0.42 Prediction of FGFR2/FGFR3 mutations: AUC = 0.86 Sensitivity = 0.82 Specificity = 0.85
Bannier et al. 93	2024	Prediction of FGFR3 mutations status	Erlangen University Hospital, TCGA	Erlangen: 767 TCGA: 307	U-Net CNN for tissue segmentation; Vision Transformer for patch-level feature extraction; ensemble of 125 multiple-instance learning models for classification.	Prediction of FGFR3 mutation status: TCGA MIBC: AUC: 0.82 Sensitivity: 0.96 Specificity: 0.47 PPV: 0.14 NPV: 0.99 Erlangen MIBC: AUC: 0.89 Sensitivity: 0.96 Specificity: 0.47 PPV: 0.21 NPV: 0.99 Erlangen mUC: AUC: 0.82 Sensitivity: 1.00 Specificity: 0.32 PPV: 0.17 NPV: 1.00
Ramon et al. 94	2024	Prediction of FGFR mutation status	TCGA & 3 clinical trials (BLC2002, BLC3001, ANNAR)	TCGA: 407 BLC2002: 184 BLC3001: 3161 ANNAR: 188	Patch-level feature extraction using ResNet-34 CNN; attention-based multiple instance learning network for slide-level classification.	Prediction of FGFR mutation status: AUC: 0.75 Sensitivity: 0.887 Specificity: 0.318
Rüschoff et al. 95	2024	Evaluate an AI-powered calculation of percent tumor cell expression of PD-L1	Institute of Pathology Nordhessen	50	Strongly supervised deep learning model with automated quantification of slide-level PD-L1 expression.	Prediction of PD-L1 expression (IHC 28–8 assay): CCCbetween: 0.84 Prediction of PD-L1 expression (IHC 22C3 assay): CCCbetween: 0.89
Lee et al. 96	2024	Evaluate an AI-powered calculation of PD-L1 combined positive score (CPS)	Five US institutions	943	Strongly supervised cell annotation and tissue segmentation with automated quantification of PD-L1 combined positive score (CPS).	Prediction of PD-L1 CPS (≥ 10 or < 10): Accuracy: 0.895

BCG: Bacillus Calmette-Guérin; CCC_between: Concordance Correlation Coefficent between AI model and pathologists; CNN: Convolutional Neural Network; CPS: Combined Positive Score; FGFR: Fibroblast Growth Factor Receptor; HoVer-Net: Horizontal and Vertical Network; HR: Hazard Ratio; MIBC: Muscle-Invasive Bladder Cancer; mUC: Metastatic Urothelial Cancer; NMIBC: Non-Muscle Invasive Bladder Cancer; PD-L1: Programmed Death-Ligand 1; ResNet: Residual Neural Network; TIL: Tumor-Infiltrating Lymphocyte.

Treatment response prediction – intravesical BCG

For patients with high-risk NMIBC, intravesical Bacillus Calmette-Guérin (BCG) is considered standard first-line adjuvant therapy following TURBT. However, there are currently no prospectively validated biomarkers to identify which patients are most likely to benefit from this treatment, or when intravesical chemotherapy may be a preferred alternative.

Recent efforts using digital WSIs have shown great promise in predicting BCG response, highlighting the potential of computational pathology to guide treatment decisions. Jiang et al. introduced two ensemble models to predict both early recurrence and BCG response in patients with NMIBC. ⁸⁷ Similarly, Lotan et al. developed a strongly supervised, interpretable AI-based histologic model capable of predicting BCG-unresponsive disease, recurrence, progression, and need for cystectomy in high-risk NMIBC. ⁸⁸ This model, termed the computational histology AI (CHAI) platform, was subsequently evaluated by Packiam et al. in a multi-center retrospective international cohort study. ⁸⁹ Notably, Packiam et al. demonstrated that CHAI biomarker-positive patients had worse recurrence free survival with adjuvant BCG compared to sequential intravesical gemcitabine and docetaxel (Hazard Ratio [HR] 5.4, 95% CI 1.6–18.3, p = 0.007), whereas biomarker-negative patients had similar recurrence-free survival with either regimen (HR 1.3, 95% CI 0.6–2.6, p = 0.5). This result is particularly relevant given recent BCG shortages and increasing evidence to support intravesical gemcitabine and docetaxel as a possible BCG alternative for treatment-naïve high-risk NMIBC.^97–99 With prospective validation, similar AI-based digital pathology tools could support individualized treatment decisions in high-risk NMIBC.

Treatment response prediction – neoadjuvant chemotherapy

Standard treatment for MIBC includes neoadjuvant cisplatin-based chemotherapy. However, only about half of patients experience pathologic downstaging following treatment, ¹⁰⁰ and it remains unclear which individuals derive the greatest benefit.

Early predictive models based on histologic features, such as image texture, nuclear morphology, cellular clustering, and spatial correlations demonstrated only modest accuracy in predicting chemotherapy response. ⁹⁰ More recently, Bai et al. introduced a human-interpretable, graph-based multimodal deep learning framework that integrates digital pathology with RNA sequencing data. ⁹¹ This approach demonstrates that a multimodal deep learning model that integrates digital pathology and molecular data can predict chemotherapy response in MIBC with an AUC of 0.72 and outperforms unimodal models. Together, these findings suggest that multimodal, AI-driven strategies incorporating digital pathology may help personalize neoadjuvant chemotherapy decisions and optimize treatment selection in MIBC.

Predictive biomarkers – FGFR3

Patients with locally advanced or metastatic urothelial carcinoma who have progressed on prior first-line therapy and harbor FGFR3 mutations are eligible for targeted treatment with erdafitinib, an FGFR1-4 inhibitor. However, molecular testing can be costly and may delay treatment initiation. Machine learning applied to digital pathology offers a promising strategy to pre-screen patients for molecular testing.

Several studies have shown that AI algorithms can detect actionable genetic mutations directly from pathology images. Velmahos et al. used a CNN to quantify TILs and demonstrated that this metric could predict FGFR-activating mutations in bladder cancer. ⁹² Bannier et al. further validated the association between TIL depletion and FGFR mutation status, and developed a deep learning model that predicted FGFR3 mutations from digital pathology with a sensitivity of 93–100%. ⁹³ Notably, Ramon et al. developed a predictive model for FGFR3 mutations that was prospectively validated under an Investigational Device Exemption (IDE) regulatory designation and implemented in clinical workflows to guide decisions regarding molecular testing. ⁹⁴ This prospective validation represents a critical step toward the clinical adoption of AI-based tools in pathology for bladder cancer.

Predictive biomarkers – PD-L1

Although PD-L1 status is not currently included in clinical guidelines to guide immunotherapy decisions in bladder cancer, it remains a valuable biomarker for predicting immunotherapy response in other tumor types. Two AI-based tools have been developed to automate PD-L1 assessment in urothelial carcinoma, both demonstrating high accuracy and strong concordance with expert pathologists.^95,96 Notably, the model developed by Lee et al. was shown to improve inter-pathologist agreement in previously discrepant cases, suggesting that AI may help standardize PD-L1 combined positive score quantification and facilitate diagnostic consensus. ⁹⁶

Overall, this evidence suggests that computational pathology has the potential to inform personalized treatment decisions. AI-driven digital pathology models can function as tool to pre-screen patients for actionable molecular traits, such as FGFR3 mutations, and as a standalone predictive biomarker of treatment response. However, the existing models for bladder cancer have been developed and validated using only retrospective data, and rigorous validation in prospective clinical trials will be essential for these tools to move towards widespread clinical adoption.

Recent advances

Multimodal machine learning

There is growing interest in integrating information from multiple data types to develop more accurate and informative predictive models. With increasing access to clinical data, genomic data, radiologic imaging, and now digital pathology, researchers are well-positioned to leverage this multimodal data to enhance model performance across a range of clinical applications. In cancer research, combining histopathology with multi-omic, radiomic, and clinical data holds promise for identifying novel biomarkers, predicting patient outcomes, and informing personalized treatment strategies. In bladder cancer, such approaches are beginning to emerge. In March 2025, Bai et al. published a model that integrated histologic features from WSIs and microarray gene expression data to predict response to neoadjuvant chemotherapy in muscle-invasive bladder cancer, demonstrating improved accuracy over unimodal models. ⁹¹ Such work highlights the promise of multimodal learning in precision oncology, demonstrating that models incorporating multiple data sources are likely to outperform those relying on digital pathology alone.

3D digital pathology

There have been recent advancements in non-destructive 3D pathology, which is a technology that enables imaging of large tissue specimens to generate a three-dimensional complement to traditional 2D slide-based pathology. ¹⁰³ This approach offers several potential advantages: (1) visualization of 3D tissue architecture and identification of novel 3D biomarkers, (2) more comprehensive sampling compared to individual 2D sections, and (3) preservation of the original specimen for downstream molecular analysis. ¹⁰⁴

Despite its promise, this technology also presents several challenges, including a new pathology image format without clear standards for human interpretation and large volumes of data, which currently cannot be managed at scale. As such, there are emerging approaches to “2.5D pathology”, which use non-destructive 3D pathology coupled with a machine learning triage approach to identify the highest-risk 2D slices, ultimately allowing pathologists to efficiently review the most important sections of tissue in a familiar format.^105–107

Three-dimensional pathology has been used in studies related to prostate cancer^108–110 and other malignancies. ¹⁰⁷ However, its use in bladder cancer remains largely unexplored and represents a valuable area for future investigation.

Large language models

One of the major challenges in developing machine learning models for digital pathology is developing patient cohorts and extracting relevant data from clinical documentation, which can serve as case or slide-level labels. Generating usable datasets from electronic health records often requires substantial manual effort to review pathology reports, clinical notes, and molecular testing results, depending on the outcome of interest. To address this bottleneck, there is growing interest in automating the extraction of structured data from unstructured clinical documentation using large language model (LLMs). These models have the potential to generate high-quality labels without the need for extensive manual annotation.^111–113 As natural language processing-based approaches to image annotation are increasingly utilized, we can expect to see a significant increase in the scale of annotated datasets for computational pathology in bladder cancer and other tumor types.

Challenges and future directions

There are numerous challenges facing the field of computational pathology and potential areas for growth. We focus on those issues that are especially pronounced in the context of bladder cancer.

The development of predictive models using digital pathology data is subject to several potential sources of bias from pathology images. These include, but are not limited to, artifacts introduced by variation in staining procedures, imaging techniques, image quality, and institution-specific practices. Although methods such as stain normalization, data augmentation, and noise reduction exist to mitigate these biases, they are not consistently employed in the studies presented in this review. These limitations can substantially affect generalizability of predictive models and highlight the importance of cautious interpretation of results, particularly in the absence of external validation.

Due to the sensitive and often limited availability of well-annotated digital pathology datasets, many predictive models are developed and evaluated using small institutional cohorts or limited publicly available datasets without external validation. Compounding this issue, approaches to performance evaluation vary widely across studies, with little consistency in how results are reported. This inconsistency makes it difficult, if not impossible, to compare computational pathology approaches across studies, ultimately hindering progress. There is growing recognition of the need for standardized benchmarks, performance metrics, and validation protocols in computational pathology to facilitate meaningful comparisons and promote robust model development. ¹¹⁴ Additionally, the field would benefit from standard datasets for external validation, as is common practice for other computer vision problems.

Machine learning applications in digital pathology for bladder cancer will significantly benefit from efforts to create and maintain large, multi-institutional, publicly accessible datasets particularly suited for external validation. Such datasets are instrumental to progress in machine learning research.^115,116 Although many public digital pathology datasets exist at the pan-cancer level, such as The Cancer Genome Atlas, they often lack the detailed data necessary to evaluate task-specific models that are most clinically relevant to bladder cancer patients. Bladder cancer-specific datasets that combine digital pathology with robust clinical annotations, and potentially include other data modalities, would be especially valuable for moving the field forward.

While numerous digital pathology approaches have been developed for bladder cancer, as highlighted in this review, none has yet seen widespread adoption in clinical practice. Ultimately, the integration of these tools into practice will depend on their impact on patient outcomes and utility to practicing pathologists. In addition to demonstrating model performance through prospective validation, successful clinical implementation will require overcoming a range of technical, operational, regulatory, financial, and cultural barriers. As computational pathology in bladder cancer continues to show promise, active engagement with key stakeholders, including institutional review boards, regulatory bodies, and patient advocacy groups, will be essential for navigating clinical integration. This will be critical to ensure computational pathology fulfills its potential not just as a research innovation, but as a transformative clinical tool in the management of bladder cancer.

Comprehensive reviews of the literature, such as this one, are essential for understanding the current state of the field and guiding its future development. Limitations of this review include the use of a single database for article identification and the absence of a formal critical appraisal process. Future scoping reviews should build upon this work to further highlight important developments and identify opportunities for advances in computational pathology for bladder cancer.