Al-Powered classification of Ovarian cancers Based on Histopathological lmages

Haitham Kussaibi,Elaf Alibrahim,Eman Alamer,Ghada Alhaji,Shrooq Alshehab,Zahraa Shabib,Noor Alsafwani,Ritesh G. Meneses

DOI: https://doi.org/10.1101/2024.06.05.24308520

2024-06-06

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.

What problem does this paper attempt to address?

This paper discusses the use of artificial intelligence (AI) to assist in identifying subtypes of ovarian cancer. The research is based on a large dataset of whole slide images (WSIs) of tissue pathology images, consisting of over 2 million image slides, covering the major epithelial ovarian cancer (EOC) subtypes: clear cell carcinoma, endometrioid carcinoma, mucinous carcinoma, and serous carcinoma. By using a convolutional neural network (ResNet50) to extract features and input them into two classifiers (CNN and LightGBM), the paper aims to predict cancer subtypes.

The study found that the AI classifier achieved a block-level accuracy of 97-98% on the test set, and the discriminative performance between subtypes was improved by adding a weighted cross-entropy loss function. The conclusion of the paper states that AI models trained on pathological image data can accurately classify EOC subtypes, assisting pathologists' work and reducing subjectivity in diagnosis.

The paper mentions that despite advancements in early diagnosis techniques for ovarian cancer, issues such as high recurrence rate and late-stage diagnosis still exist, leading to high mortality rates. Therefore, developing an AI system that can automatically identify EOC subtypes is an important step in improving diagnostic efficiency and accuracy. The research used publicly available and institutional data, but the sample size and subtype balance need to be improved, which may affect the generalization ability of the model.

In summary, this paper aims to address how AI technology can improve the analysis of ovarian cancer pathology images to assist pathologists in identifying different types of ovarian cancer more accurately and quickly, thereby improving the quality and efficiency of diagnosis.