The stage of bladder cancer determines everything about treatment. Bladder cancer is classified as either non-muscle-invasive (NMIBC) or muscle-invasive (MIBC) based on how deeply the tumor has grown into the bladder wall. These two categories require fundamentally different treatments, making accurate staging a matter of life and death.
NMIBC is treated with transurethral resection followed by bladder-preserving therapies and has a five-year survival rate of up to 90%. MIBC requires radical cystectomy, pelvic lymph node dissection, and often chemotherapy or immunotherapy, with a much lower five-year survival rate of around 66%.
Histologic grade adds another layer of complexity. High-grade tumors are more aggressive and prone to recurrence and progression, requiring closer long-term follow-up. Misclassifying grade or invasion depth can lead to undertreatment of dangerous cancers or overtreatment of indolent ones.
Current imaging methods including CT and MRI achieve only 64.7 to 83% accuracy for preoperative staging. Tissue obtained during transurethral resection is often fragmented, scarce, or lacking complete muscle layers, making pathological diagnosis challenging even for experienced pathologists and creating significant inter-observer variability.
854 whole slide images from 692 patients formed the dataset. All patients underwent transurethral resection of bladder tumor at Sun Yat-sen Memorial Hospital between 2013 and 2019. H&E stained slides were digitally scanned at 40-fold magnification and classified into three categories: high-grade muscle invasion (HGMI), high-grade non-muscle invasion (HGNMI), and low-grade non-muscle invasion (LGNMI).
For fully annotated images, pathologists labeled six tissue types at the patch level: HGMI, HGNMI, LGNMI, illegible area, normal interstitial area, and noise area. Disagreements were resolved by a pathologist with over 30 years of experience. Images were split into training, validation, and independent comparison sets.
The model, called the Pathological AI Diagnostic Model (PAIDM), was built using ScanNet, a convolutional neural network architecture previously validated for breast cancer metastasis detection. The model was initialized with ImageNet pre-trained weights and trained for 30 epochs with cross-entropy loss and stochastic gradient descent optimization.
During inference, each whole slide image is divided into patches. The CNN classifies each patch and stitches predictions into a six-channel heatmap. A confidence score for whole slide classification is computed by weighting contour areas and their mean probabilities, producing both patch-level and slide-level diagnostic outputs.
Strong performance across both patch-level and whole-slide classification. At the patch level using 112,472 patches from 100 whole slide images, the PAIDM achieved an overall AUC of 0.878 (95% CI 0.875 to 0.881). Individual subtypes performed well: HGMI reached AUC 0.904, LGNMI reached 0.889, and HGNMI reached 0.840.
At the whole slide image level using 261 slides in a separate validation set, the overall AUC was 0.870 (95% CI 0.805 to 0.923). For the clinically critical task of distinguishing muscle-invasive from non-muscle-invasive disease, the PAIDM achieved accuracy of 0.850, specificity of 0.941, and negative predictive value of 0.963.
For grade classification (low-grade versus high-grade), the PAIDM achieved accuracy of 0.862, sensitivity of 0.867, and specificity of 0.849 at the whole slide level. The system was slightly less accurate for HGNMI compared to the other two categories, likely because high-grade non-muscle-invasive tumors lack the distinguishing muscle tissue that makes HGMI easier to identify.
Visual heatmaps generated by the PAIDM showed high consistency with manually annotated regions, demonstrating that the model focuses on biologically relevant tissue features rather than spurious image artifacts. This interpretability is important for clinician trust and adoption.
The AI model matched intermediate-level pathologists. In a direct comparison on an independent validation cohort of 83 patients, the PAIDM achieved an AUC of 0.847 (95% CI 0.779 to 0.905), outperforming both junior pathologists (AUC 0.752 and 0.792) and matching one intermediate pathologist (AUC 0.822). Senior pathologists scored higher with AUCs of 0.918 and 0.930.
The PAIDM was faster than all human pathologists. Junior pathologists averaged 252 seconds per slide, intermediate pathologists 195 seconds, and senior pathologists 178 seconds. The AI model completed analysis in just 144 seconds, faster than even the most experienced clinicians.
Unlike human pathologists who may vary in performance based on fatigue, workload, or case complexity, the AI system provides consistent output regardless of caseload. It processes slides without fatigue and produces the same quality analysis for the hundredth case as for the first.
The PAIDM also highlighted prediction masks in images, enabling pathologists to quickly focus on suspicious regions. This combination of speed, consistency, and visual assistance positions the model as a powerful assistant that augments rather than replaces human expertise.
AI as an equalizer for unequal healthcare resources. In China and many other countries, there are significant disparities in pathology expertise between large urban hospitals and rural or primary care facilities. The PAIDM could directly address this gap by providing junior pathologists in underserved areas with AI-assisted diagnoses that approach the quality of expert centers.
In high-volume centers, the PAIDM could handle routine cases efficiently, freeing senior pathologists to focus on complex or ambiguous cases that genuinely require their expertise. This division of labor could improve both throughput and diagnostic quality across healthcare systems.
The system supports a fully automated workflow. Pathologists place stained slides into the scanner, which automatically scans, uploads, and processes the images through the PAIDM without additional manual steps, producing an end-to-end diagnostic output ready for clinical review.
This is the first study to apply AI specifically to the identification of muscle invasion in bladder cancer pathology slides, filling a critical gap in the literature. Previous AI studies for bladder cancer pathology focused only on grade classification, not the invasion depth that most critically determines treatment choice.
Single-center design limits immediate generalizability. All data came from one institution, raising concerns about overfitting to that center's specific staining protocols, scanning hardware, and patient demographics. Multi-center prospective validation is needed before widespread clinical adoption.
The model currently cannot classify carcinoma in situ, a non-muscle-invasive subtype that is poorly differentiated and carries high progression risk. Adding this subtype to future training sets will be necessary to make the PAIDM comprehensively useful in clinical practice.
Annotation was performed with full pixel-level labeling, which maximizes information use but is extremely time-consuming and limits the number of images that can be labeled for training. Future work incorporating partial annotation and weak supervision methods could enable training on much larger datasets.
The model also does not yet classify other clinically relevant findings such as dysplasia and inflammation. Expanding the training data to include these findings would broaden the PAIDM's applicability and move it closer to a comprehensive pathology assistant for bladder cancer specimens.