The Clinical Stakes of Chemotherapy Response. Neoadjuvant chemotherapy (NAC) followed by radical cystectomy is the standard treatment for muscle-invasive bladder cancer (MIBC). Patients who achieve a pathologic response to NAC (no residual muscle-invasive disease at surgery) have a 5-year survival rate of 80-90%, while non-responders have only 30-40% 5-year survival.
Despite this dramatic survival difference, NAC provides only a modest 5-10% average survival benefit across all MIBC patients because only a minority respond. The challenge is identifying in advance which patients will respond -- before administering months of toxic chemotherapy that may provide no benefit to non-responders.
The Biomarker Gap. Various biomarker approaches have been explored, including clinical factors, tumor molecular subtypes (luminal vs. basal), somatic mutations, gene expression profiles, and tumor microenvironment composition. No biomarker has yet been established as standard clinical practice for guiding NAC decisions in MIBC. This study pursues a novel approach: computational analysis of routine pathology slides to predict response before treatment begins.
Tissue Microarray Datasets. Pre-NAC biopsy specimens from two cohorts were analyzed: a Johns Hopkins discovery cohort (66 evaluable patients, including 37 responders) and an independent University Hospital Bern validation cohort (56 evaluable patients). Discovery cohort samples were stained with H&E and nine immunohistochemistry (IHC) stains (cyclin D1, P16, P53, P63, Ki67, CK20, CK5/6, GATA3, and Her2Neu). Only H&E was available for the validation cohort.
A comprehensive computational feature extraction pipeline was built using QuPath software. Features were computed at five levels: (1) image texture (265 features per tile from grayscale H&E), (2) nucleus morphology (18 shape descriptors per nucleus), (3) nuclear clustering and cell/nucleus orientation entropy (COrE), (4) spatial correlations between cell types using Ripley's K function, and (5) cell population composition classifying nuclei as lymphocytes, cancer cells, or stromal cells using a pathologist-trained random tree classifier. Together, 1,187 features per tissue core were generated.
Advanced Spatial Analysis. A key innovation was the use of multi-type Ripley's K functions to capture spatial relationships between cell types -- for example, how lymphocytes relate spatially to cancer cells versus stromal cells. Additionally, the degree of colocalization (DoC) algorithm quantified spatial correlations between cell type pairs, and spatial Shannon's entropy captured the diversity of cell species across tissue regions. These spatial metrics aimed to characterize the tumor microenvironment structure that may drive chemotherapy response.
Because IHC sections required alignment, a three-step registration pipeline using t-SNE and k-means clustering identified adjacent sections from the TMAs suitable for image registration, achieving average DSC scores above 0.95. Nuclear segmentation used StarDist for H&E and nuclear IHC stains, with a watershed algorithm for membranous or cytoplasmic IHC stains. All segmentation algorithms showed strong agreement with manual counts (Spearman r = 0.9966 for H&E).
Imaging Features Substantially Improve Prediction. The clinical and demographic baseline model (using clinical T stage and age) achieved only 59% accuracy with AUC 0.52 in cross-validation -- essentially no better than chance. Adding H&E-derived imaging features (strategy V) improved accuracy to 69.17% (AUC 0.64) in the discovery cohort and 65.82% (AUC 0.61) in the independent external validation cohort, representing a 16.6% improvement in accuracy and 23.1% improvement in AUC.
The best overall model, which used all five feature categories including IHC-derived spatial correlations, achieved 72.52% accuracy (AUC 0.67) in the discovery cohort cross-validation. Critically, the H&E-only model (strategy V) worked without any immunohistochemistry or special staining -- it relied only on the standard diagnostic stain that is already routinely prepared from every bladder cancer biopsy.
Enriching Responders Among Predicted Responders. The most clinically meaningful result was that the model could enrich the response rate among predicted responders. In the discovery cohort, predicted responders had a response rate of 64.64% compared to the 38.03% baseline -- nearly a doubling. In the independent validation cohort, predicted responders had a 64.71% response rate versus the 40.51% baseline. This enrichment was reproducible across both cohorts and represents a meaningful clinical improvement in patient selection for NAC.
Response-Associated Features. Analysis of the top 10 predictive features revealed distinct patterns between responders and non-responders. Responders showed high CK20 expression (consistent with a luminal MIBC subtype), elevated Ki67 and P16 staining density (indicating rapid cell proliferation and aberrant cell cycle regulation), tight spatial proximity between cells, and more elongated cell shapes. This profile is consistent with the general principle that highly proliferative tumors should be more sensitive to platinum-based chemotherapy.
Non-Response Features. Non-responders were characterized by greater nuclear shape diversity (high coefficient of variation in minimum nucleus diameter and moment of inertia), more chaotic cell orientations, low CD8/FoxP3 ratio (indicating regulatory T cell dominance over cytotoxic T cells), and a stromal barrier impeding lymphocyte infiltration into the tumor. This pattern suggests that immunosuppression mediated by regulatory T cells combined with a dense stromal barrier creates a microenvironment hostile to cytotoxic immune responses and favorable to chemotherapy resistance.
The spatial feature category (derived from IHC stains) was consistently the most predictive across all model evaluations, outperforming image texture and nuclear morphology features individually. This underscores the importance of cellular spatial context -- not just individual cell characteristics -- in determining treatment response. Incorporating pathologist-measured CD8/FoxP3 ratio further improved model accuracy to 72.62% with AUC 0.68, suggesting that immune microenvironment assessment adds independent predictive value.
Predicted Response Correlates with Overall Survival. Beyond prediction accuracy, the machine learning model was evaluated for its ability to stratify patients by clinical outcome. Using a model output threshold of 0.25, patients predicted to respond (classified as low-risk) showed significantly better overall survival than those predicted not to respond (log rank p less than 0.01) in the discovery cohort.
This finding is consistent with the well-established clinical observation that MIBC patients who achieve pathologic response to NAC have dramatically better long-term outcomes. The ability of a pre-treatment computational pathology model to identify this survival difference from routine biopsy images alone suggests the method captures biologically meaningful information about tumor biology and treatment susceptibility.
Survival data were not available for the external validation cohort, limiting full external validation of this survival stratification capability. Future studies with complete survival data in both discovery and validation populations will be needed to confirm this finding.
A Clinically Practical Approach. The most important finding is that meaningful NAC response prediction can be achieved from routine H&E-stained biopsy specimens alone, without requiring molecular profiling, special stains, or clinical information beyond age and T stage. Since H&E slides are already standard in every diagnostic workup, this approach could in principle be deployed without additional biopsy procedures or laboratory tests.
The framework's design at the tissue core level rather than the patient level preserves intra-tumor heterogeneity and effectively increases the statistical power of the model by treating each TMA core as an independent observation. This approach would naturally extend to whole-slide imaging, where multiple sampled regions could be analyzed and aggregated for each patient.
Limitations and Next Steps. Cohort sizes were small (66 discovery, 56 validation), and survival data were only available for the discovery cohort. The feature engineering approach, while interpretable, does not leverage the full power of deep learning that might extract additional predictive information. Future work should validate on larger cohorts, explore deep learning feature extraction, and investigate whether the model generalizes across different NAC regimens and patient populations. Extending the approach to whole-slide imaging rather than tissue microarrays would be a critical step toward clinical implementation.