Al-Powered classification of Ovarian cancers Based on Histopathological lmages

Abstract

Background: Ovarian cancer is the leading cause of gynecological cancer deaths due to late diagnosis and high recurrence rates. While histopathological analysis is the gold standard for diagnosis, artificial intelligence (AI) models have shown promise in accurately classifying ovarian cancer subtypes from histopathology images. Herein, we developed an AI pipeline for automated identification of epithelial ovarian cancer (EOC) subtypes based on histopathology images and evaluated its performance compared to the pathologists' diagnosis. Methods: A dataset of over 2 million image tiles from 82 whole slide images (WSIs) of the major EOC subtypes (clear cell, endometrioid, mucinous, serous) was curated from public and institutional sources. A convolutional neural network (ResNet50) was used to extract features which were then input to classifiers (CNN, and LightGBM) to predict the cancer subtype. Results: Both AI classifiers achieved patch-level accuracy (97-98%) on the test set. Furthermore, adding a class-weighted cross-entropy loss function to the pipeline showed better discriminative performance between the subtypes. Conclusion: AI models trained on histopathology image data can accurately classify EOC subtypes, potentially assisting pathologists and reducing subjectivity in ovarian cancer diagnosis.