Muscle-invasive bladder cancer (MIBC) has a recurrence rate of up to 50% after radical cystectomy, suggesting that microscopic metastatic deposits are often present at the time of surgery even when the surgical margins are clear. Better risk stratification before surgery is urgently needed to identify which patients should receive neoadjuvant chemotherapy (NAC) to reduce recurrence.
Current prognostic factors for MIBC -- T stage, lymphovascular invasion (LVI), and variant histologic features -- are inconsistently reported and have highly variable concordance between pre-surgical biopsy and final cystectomy pathology. Clinical T stage frequently disagrees with pathologic T stage, limiting its ability to select patients for intensified treatment.
Lymph node (LN) positivity at cystectomy is one of the strongest adverse prognostic indicators, yet no reliable non-invasive method exists to identify LN-positive patients before surgery. Identifying them prospectively would enable better selection of patients for NAC, which provides only a 5% overall survival benefit in unselected populations but presumably more in high-risk subgroups.
AI approaches applied to whole-slide H&E images offer a fundamentally different approach: rather than relying on pathologist-reported categorical features, deep learning models can learn quantitative imaging features across the entire slide that correlate with biological behaviors such as metastatic potential -- information that qualitative pathology assessment does not systematically capture.
All 457 digital H&E pathology slides from 386 bladder cancer cystectomy specimens in The Cancer Genome Atlas (TCGA) were downloaded from the NCI Genomic Data Commons. After excluding 92 patients for missing lymph node staging information, poor slide quality, pure small cell carcinoma, or TURBT specimens, 294 patients were included from TCGA.
A board-certified pathologist manually annotated tumor regions within each slide, providing the spatial boundaries needed for patch extraction. Only the first diagnostic slide (DX1) per patient was used for statistical analysis, consistent with prior TCGA-based studies. Median one slide was available per patient (range 1 to 8).
An additional 13 patients from the authors' institution who underwent cystectomy between 2013 and 2019 provided an external validation cohort, with 30 of 43 institutional patients excluded due to missing slides, exclusion criteria, or more than two types of neoadjuvant therapy.
Patients were split: 146 for training, 73 for validation, and 88 for testing (75 TCGA plus 13 in-house). LN positivity rates were 35.7% in TCGA and 23.1% in the in-house cohort, with 94% to 100% high histologic grade across cohorts. There were no statistically significant demographic differences between training and test cohorts.
Four separate ResNet-101 convolutional neural network models were trained on image patches at 2.5x, 5x, 10x, and 20x effective magnifications to capture features at different spatial scales -- from broad tumor architecture at low magnification to cellular morphology at high magnification.
Within each 1,200 by 1,200 pixel tumor region at 10x reference resolution, patches were extracted hierarchically: 1 patch at 2.5x (300 by 300 pixels downsampled to 1/4 size), 4 patches at 5x (1/2 size), 16 patches at 10x (full size), and 64 patches at 20x (4,800 by 4,800 mapped to 300 by 300 patches). Each patch was assigned the patient-level LN status as its label.
Stain normalization using the Macenko method and background artifact removal were applied before training. Data augmentation including rotation, flipping, intensity warping, brightness, and contrast variation was applied randomly during training to reduce overfitting to patient- and slide-specific characteristics.
At the highest magnification (20x), 664,754 training patches and 379,404 validation patches were used, with weighted cross-entropy loss applied to address the class imbalance between LN-positive and LN-negative patients. All four resolution-specific models generated LN-positivity probability scores for every patch in every cohort.
Patch-level predictions from all four resolution models were spatially referenced back to original slide coordinates to create multiresolution probability maps, in which every pixel represents the regional probability of LN positivity. This spatial representation captures the heterogeneous distribution of high-risk features within the tumor.
Density-Based Spatial Clustering of Applications with Noise (DBSCAN) was applied to the probability maps across 12 different probability threshold conditions to identify spatially coherent clusters of high-risk tissue regions. From these clusters, metrics including total cluster area, size, density, and distance between clusters were extracted.
Tumor-infiltrating lymphocyte (TIL) maps were generated in parallel from the same spatial coordinates using 576 additional 100 by 100 pixel patches per region, applying a publicly available lymphocyte detection algorithm from Saltz et al. TIL features including burden, cluster size, and distance from high-risk AI clusters were extracted separately.
In total, 716 features were extracted per patient from multiresolution AI probability maps and TIL maps. An AdaBoost ensemble of decision trees was then fit using 10-fold cross-validation on the combined training and validation cohort, selecting the top 26 consistently important features including 8 TIL-related variables, to generate the final patient-level AI score.
The clinicopathologic model using age, T stage, and LVI achieved an AUC of 0.755 in training and validation cohorts and 0.678 in the test cohort. The AI score significantly outperformed clinicopathologic features in training and validation (AUC 0.866, p = 0.021) and showed superior performance in the test cohort (AUC 0.784) despite not reaching statistical significance (p = 0.208) in that smaller cohort.
The combined model integrating both AI score and clinicopathologic features significantly outperformed clinicopathologic features alone in the test cohort (AUC 0.807 versus 0.678, p = 0.047), confirming the AI score provides independent prognostic information beyond what standard pathologic variables capture.
The AI score remained highly significant after multivariable adjustment for all clinicopathologic variables (odds ratio 6.36, 95% CI 3.66 to 12.1, p = 1.08 times 10 to the negative 9), demonstrating that the imaging features learned by the CNN are not simply proxies for T stage or LVI but represent genuinely independent prognostic information.
On the test cohort, the AI model achieved 84.5% specificity with 66.7% positive predictive value. Patient-level accuracy in the in-house cohort was 84.6%, with 11 of 13 patients correctly classified. The simplified majority vote method (AUC 0.762) performed slightly worse than the full DBSCAN-AdaBoost pipeline (AUC 0.784), validating the post-processing approach.
Binary AI classification was significantly associated with high-risk variant histologies including micropapillary, neuroendocrine, and plasmacytoid features in both training (p = 0.019) and test cohorts (p = 0.044), confirming that the CNN learned to recognize histologic patterns already known to correlate with adverse outcomes.
Of the 26 features selected for the AdaBoost ensemble, 8 captured the presence and spatial distribution of tumor-infiltrating lymphocytes, highlighting that the immune microenvironment provides substantial prognostic information beyond tumor cell morphology alone. High-risk regions identified by AI were significantly less lymphocyte-rich than low-risk regions (Wilcoxon p = 3.66 times 10 to the negative 8).
The multiresolution probability maps revealed intratumoral heterogeneity in the distribution of high-risk imaging features, with distinct spatial clusters of high-probability regions rather than uniform high or low risk across the tumor. This heterogeneity itself appears to contain prognostic signal captured by the cluster metrics.
The CNNs appear to have learned features not captured in traditional clinical reporting -- features associated with metastatic potential that pathologists do not routinely quantify -- explaining why the AI score provides independent predictive value beyond what standard staging and histologic characterization report.
The most clinically actionable application of this approach would be to apply it to TURBT biopsy specimens before cystectomy, enabling identification of high-risk patients likely to have LN-positive disease who should receive NAC. This would allow treatment intensification in patients most likely to benefit rather than in all MIBC patients uniformly.
The current study was conducted on cystectomy specimens, where the full tumor is available. Validation in TURBT samples -- which provide limited and potentially unrepresentative tissue -- is needed before clinical deployment. Clinical T stage on TURBT is already known to frequently disagree with pathologic T stage at cystectomy, and similar challenges may apply to AI scores derived from pre-operative biopsy specimens.
Limitations include training on only 146 patients (though thousands of patches per resolution), evaluation limited to a single TCGA slide per patient, variable numbers of lymph nodes sampled across patients (affecting LN-positive ground truth reliability), and 16% missing LVI data in the training set. The model requires pathologist tumor annotation before inference, which limits scalability.
All code and methodology for reproducing the study is publicly available at github.com/NIH-MIP/BLCA_LNprediction, enabling external groups to validate and extend this approach on their own datasets.
This study provides proof-of-concept that deep learning algorithms applied to primary tumor H&E slides can identify features associated with lymph node metastasis that are independent of established clinicopathologic prognostic variables. The AI score remained strongly significant after adjusting for T stage, LVI, age, and variant histology, confirming genuine additive prognostic value.
The multiresolution approach captures complementary information at different spatial scales: low magnification encodes broad tumor growth patterns and architectural features, while high magnification captures nuclear and cellular morphology. Combining all four resolutions with microenvironmental TIL information produces a richer prognostic model than any single scale alone.
Unlike pathologist-based scoring, the AI approach is not subject to interobserver variability and provides a continuous probability score rather than a categorical assessment, enabling more nuanced patient risk stratification. The approach can be generalized across multiple institutions, as demonstrated by performance in the 22-center TCGA test set.
Future directions include prospective validation in TURBT specimens, integration with molecular biomarkers (genomics, transcriptomics), and extension to predict response to NAC rather than only LN status -- ultimately enabling a pre-surgical AI-based biomarker panel to guide individualized treatment selection for MIBC patients.