Machine Learning Quantification of Tumor-Stroma Ratio in Early Muscle Invasive Urothelial Carcinomas

Tumor-Stroma Ratio, a marker of tumor microenvironment, proved to be a reliable independent prognostic predictor in many solid tumors but it’s value in transitional carcinoma is still under research. Visual quantification of tumoral and stromal areas is possible but is time consuming and subjective. Machine learning image segmentation can improve diagnostic precision. Our research interest is to evaluate how precision pathology tools (machine learning segmentation of whole slide images) may improve quantification of the tumor-stroma ratio in early muscle invasive bladder tumors and increase histologic diagnostic prognostic value. 10 cases of pathology stage T2A bladder cancers whole slide images were carefully matched (sex, age and smoking status) with 10 cases of pT2B form the same open database (Cancer Genome Atlas Urothelial Bladder Carcinoma dataset, TCGA-BLCA). The machine learning segmentation used a trained approach and was performed under 3 labels (tumor, stroma, other). The mean tumor to stroma ratio was significant higher (tumor>stroma) in pT2A cohort (p<0.0001). Vital prognostic was different between groups: 90% of subjects were alive at 3 years after diagnostic in pT2A cohort and only 40% in pT2B cohort. Our proof-of-concept study suggest the utility of the tumor-stroma ratio in differentiating challenging diagnostics of early muscle invasive urothelial carcinoma. A larger, real world data study will have to confirm the benefits of this marker in everyday clinical settings.