A Deep Learning Model for Guiding Precision Therapy in Bladder Cancer - Expert Commentary

Molecular subtyping of muscle-invasive bladder cancer (MIBC) tumors has the potential to guide treatment decisions and predict outcomes. However, RNA sequencing is costly and not readily available. Artificial intelligence has made its way into various clinical procedures, and recent studies have validated the use of deep learning models in the analysis of patient data for enhanced clinical management. In this study, Jiang et al. developed a deep learning model that determined subtypes based on whole slide images (stained with hematoxylin and eosin) from bladder cancer patient tumors.

Jiang et al. acquired digital whole slide images (WSI) for 363 bladder cancer patients from The Cancer Genome Atlas (TCGA). They then developed a deep learning model that predicts three-year survival based on pathological features in WSI, clinical variables, and tumor microenvironment features. Overall, WSI clustering was found to have predictive value. The analysis resulted in three distinct clusters: C0, C1, and C2. The median survival age was highest in C2, although this cluster had the worst clinical stages. C0 had more luminal subtypes and less basal subtypes than C1 and C2. There was a major difference in survival between C0 and C2. Upon analysis of differentially expressed genes (DEGs) in these clusters, the terms extracellular matrix and organization were significantly enriched, in addition to the PI3K-Akt pathway. In C2 specifically, enriched gene ontology terms included T cell activation, inflammation, and immune-related pathways.

Researchers calculated an immunophenoscore as a measure of immune activation and found that C1 exhibited the highest score. Using an alternate method for analysis (ESTIMATE), C0 had lower immune and stromal scores, while C1 had higher immune and stromal scores. C0 had the highest resting and activated CD4+ T memory cells, memory B cells, and T-reg cells. In contrast, C0 had the lowest abundance of Th1 cells, Th2 cells, and macrophages (M1 and M2). In contrast, C1 exhibited the highest amount of these cells. C2 had the highest abundance of mast cells and the lowest abundance of lymphocytes. Regarding mutational burden, C0 showed enrichment for mutations in KDM6A, ARID1A, and FGFR3. This may predict the sensitivity of patients in C0 to FGFR inhibitors. C1 exhibited enrichment for mutations in PIK3CA and RB1. STAG2 was commonly mutated in C2, correlated with poor prognosis. Across clusters, genes involved in cell cycle regulation and PI3K signaling pathways were frequently mutated.

This study reveals the potential for using artificial intelligence to generate insights into the phenotypes of bladder cancers. Furthermore, the deep learning model offers the potential for cost-effective techniques for predicting prognosis and treatment guidance.

Written by: Bishoy M. Faltas, MD, Director of Bladder Cancer Research, Englander Institute for Precision Medicine, Weill Cornell Medicine


  1. Jiang Y, Huang S, Zhu X, et al. Artificial Intelligence Meets Whole Slide Images: Deep Learning Model Shapes an Immune-Hot Tumor and Guides Precision Therapy in Bladder Cancer. J Oncol. 2022;2022:8213321. Published 2022 Sep 19. doi:10.1155/2022/8213321
Read the Abstract