Use of Artificial Intelligence for the Interpretable Prediction of the Pathologic Diagnosis and Molecular Abnormalities of Flat Urothelial Lesions

Abstract

Flat urothelial lesions are important because of their potential for carcinogenesis and development into invasive urothelial carcinomas; however, it is difficult for pathologists to detect early flat urothelial changes and accurately diagnose flat urothelial lesions. To predict the pathologic diagnosis and molecular abnormalities of flat urothelial lesions from pathologic images, we used artificial intelligence with an interpretable method. We studied 110 hematoxylin-and-eosin-stained slides of normal urothelium and flat urothelial lesions, including atypical urothelium, dysplasia, and carcinoma in situ, and performed next-generation sequencing to detect 17 kinds of molecular abnormalities. To generate an interpretable prediction, we developed a new method for segmenting urothelium and a new pathologic criteria-based artificial intelligence (PCB-AI) model. κ Statistics and accuracy measurements were used to evaluate the ability of the model to predict the pathologic diagnosis. The likelihood ratio test was performed to evaluate the logistic regression models for predicting molecular abnormalities. The diagnostic prediction of the PCB-AI model was almost in perfect agreement with the pathologists' diagnoses (weighted κ=0.98). PCB-AI significantly predicted some molecular abnormalities in an interpretable manner, including abnormalities of TP53 (P=0.02), RB1 (P=0.04), and ERCC2 (P=0.04). We generated a new method of obtaining accurate urothelial segmentation, interpretable prediction of pathologic diagnosis, and interpretable prediction of molecular abnormalities.