Related Experiment Video
Updated: Sep 9, 2025
Introduction of an Integrated Pathology Image Management, Artificial Intelligence, and Reporting System
Published on: July 11, 2025
Deep learning-driven whole-slide image analysis predicts chemo-resistance and motility subtypes in muscle-invasive
Il-San Jeong1,2,3, Jee-Woo Seo2,3, Seung-Jin Park2
1Division of Bio & Medical Bigdata Department (Brain Korea 21 Four), Gyeongsang National University, 52828, Jinju, Republic of Korea.
Background:
Muscle-invasive bladder cancer (MIBC) is a clinically aggressive and heterogeneous disease with variable treatment responses. Transcriptome-based classifications, such as the Chemoresistance-Motility (CrM) signature, are valuable for understanding therapeutic resistance, but their clinical use is often hindered by high cost and tissue requirements. This study explores an alternative, scalable approach using deep learning analysis of whole slide images (WSIs).
Objective:
We aimed to evaluate whether patch-level predictions from deep learning models applied to WSIs can accurately predict transcriptome-derived CrM subtypes and reflect tumor microenvironment (TME) characteristics in MIBC.
Methods:
We analyzed 192 WSIs from 152 TCGA-BLCA patients. A pretrained deep learning model (densenet169-kather100k) was used to classify eight distinct tissue types per patch. A key histological metric, the SAM-to-DNT ratio, which represents the ratio of stromal, adipose, and smooth muscle (SAM) tissue types to debris, normal, and tumor epithelium (DNT) tissue types, was derived from these proportions. A random forest model was then trained on these features to predict CrM subtypes.
Results:
The WSI-derived SAM-to-DNT ratio showed a strong positive correlation with the CrM score (R = 0.453) and transcriptome-based TME scores, such as cancer-associated fibroblasts (R = 0.398). Our random forest model successfully classified CrM subtypes with a balanced accuracy of 0.75, outperforming other algorithms. Feature importance analysis identified adipose tissue (ADI) and tumor epithelium (TUM) as the most predictive features for CrM status.
Conclusions:
Deep learning analysis of routine histological WSIs can serve as a practical, low-cost surrogate for molecular profiling, effectively capturing transcriptomic subtypes associated with chemoresistance in MIBC. This approach provides a viable method for patient stratification and establishes a foundation for future multi-modal precision oncology applications.
