Pathology-based deep learning features for predicting basal and luminal subtypes in bladder cancer.

Area of Science:

• Oncology
• Computational Pathology
• Digital Pathology

Background:

• Bladder cancer (BLCA) exhibits significant molecular heterogeneity, with basal and luminal subtypes displaying distinct prognostic and therapeutic implications.
• Conventional methods for BLCA molecular subtyping are often labor-intensive and require substantial resources.
• Hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) offer a rich source of morphological data for computational analysis.

Purpose of the Study:

• To develop and validate machine learning models for predicting BLCA basal and luminal subtypes.
• To leverage deep learning features extracted from H&E-stained WSIs for subtype classification.
• To assess the performance of these models against human expert pathologists.

Main Methods:

• RNA sequencing data and clinical outcomes were aggregated from multiple public BLCA databases (TCGA, GEO, IMvigor210C).
• Deep learning models were employed to identify tumor patches within WSIs, followed by feature extraction using the RetCCL model.
• Logistic regression (LR), support vector machine (SVM), and random forest (RF) models were trained using extracted features for subtype classification.
• Model performance was evaluated on internal (TCGA) and external validation cohorts (STPH, GD2H).

Main Results:

• Basal BLCA patients demonstrated significantly poorer overall survival compared to luminal BLCA patients (HR=1.47, P<0.001).
• The LR model, utilizing Resnet50-selected tumor patch features, achieved an AUC of 0.88 (internal validation) and 0.81/0.64 (external validation).
• The developed LR model surpassed both junior and senior pathologists in differentiating basal and luminal subtypes (AUC=0.85, accuracy=74%, sensitivity=66%, specificity=82%).

Conclusions:

• Machine learning models effectively predict BLCA basal and luminal subtypes from H&E-stained WSIs.
• The high performance of the LR model highlights the potential of AI in enhancing diagnostic accuracy for molecular subtyping.
• Integrating AI tools could facilitate personalized treatment strategies for BLCA patients by improving subtype classification.