Integrative Deep Learning from H&E Images Reveals Prognostically Distinct Pathology-Based Subtypes in Bladder Cancer.

Area of Science:

• Oncology
• Computational Pathology
• Digital Pathology

Background:

• Molecular subtyping of bladder cancer (BCa) is crucial for treatment but often relies on complex RNA profiling.
• Developing pathology-based subtypes using routinely available H&E-stained whole-slide images (WSIs) offers a more accessible approach.

Purpose of the Study:

• To create novel bladder cancer (BCa) subtypes based on deep learning features extracted from H&E-stained WSIs.
• To validate these pathology-based subtypes against clinical outcomes and molecular data.
• To provide an interpretable and cost-effective alternative to traditional molecular profiling.

Main Methods:

• A Resnet50 deep learning model was employed to extract features from H&E-stained WSIs.
• Weighted gene co-expression network analysis (WGCNA), Cox regression, and K-means clustering were used to define subtypes.
• External validation was conducted using WSIs from multiple centers and transcriptomic data; Grad-CAM was used for interpretability.

Main Results:

• Four distinct bladder cancer (BCa) subtypes were identified with significant differences in clinical outcomes.
• Subtypes showed distinct molecular profiles, with some enriched for specific pathways (e.g., FGFR3, EGFR) and others having higher immune/stromal scores.
• Deep learning features and Grad-CAM analysis provided interpretable insights into subtype-specific nuclear morphologies.

Conclusions:

• A novel, pathology-based bladder cancer (BCa) subtyping method was developed using deep learning on H&E WSIs.
• This approach is portable, prognostically informative, and aligns with RNA-defined biology.
• It presents a cost-effective and accessible alternative for personalized treatment and improved patient stratification in bladder cancer care.