Abstract 5430: Improved colorectal cancer survival prediction with deep learning-based WSI analysis on PETACC8 and PRODIGE13 cohorts

Le résumé fourni par la source

Abstract Colorectal cancer (CRC) is the third most common cancer worldwide and represents the third leading cause of cancer deaths. Robust postoperative prediction of CRC patient prognosis may prove useful to better stratify patients, guide therapeutic choices and improve clinical trial designs. Deep learning-based analysis of whole slide images (WSI) has recently proved successful at various prediction tasks, including for example survival prediction for malignant pleural mesothelioma. We developed a deep learning model that predicts overall survival (OS) of CRC patients, using WSI stained with haematoxylin/eosin as input, and also evaluated the difference in prognostic power obtained when combining our model prediction with other clinical factors, namely tumor grade, sex and age. The model was trained on PETACC8 cohort, constituted of 1939 WSI from patients with stage III colon cancer. The PRODIGE13 cohort, constituted of 1155 WSI from patients with stage II and III colon and rectal cancers, was used for validation; only stage III colon cancers (N=428) were selected. Patients from both cohorts received standard chemotherapy treatment, with half of the PETACC8 population also receiving CETUXIMAB antibody treatment. Both cohorts were provided by the FFCD. Our model first extracts information on small tiles of size 112µm from the WSI, using a deep learning network trained in a self-supervised fashion, then aggregates the information of these tiles using a Multiple Instance Learning model (MIL) at the slide level to establish the final prediction. Our model was able to predict OS from WSI, reaching a c-index of 0.63 [0.61 - 0.66] in cross-validation over the PETACC8 cohort and a c-index of 0.59 [0.53 - 0.65] when transferring the PETACC8-trained model onto the PRODIGE13 cohort. The model was also able to significantly stratify patients into high and low risk groups [HR: 2.67; p<0.0001]. We observed a significant c-index gain when combining our prediction and the pT classification in a linear Cox model as compared to pT alone, with c-index increasing from 0.61 to 0.66 (p=0.03). Overall, we demonstrated that our model is able to robustly predict OS from WSI in stage III colon cancers and provides increased prognostic power, on top of more traditional clinical markers such as tumor grading. Further investigation of the WSI regions targeted by our model could provide valuable insights into postoperative histopathological features of prognostic significance. Citation Format: Jean-Eudes Le Douget, Julien Taïeb, Paul Jacob, Frédéric Bibeau, Karine Le Malicot, Jean-François Emile, Aurélien de Reyniès, Mehdi Morel, Simon Jégou, Côme Lepage, Pierre Laurent-Puig. Improved colorectal cancer survival prediction with deep learning-based WSI analysis on PETACC8 and PRODIGE13 cohorts. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5430.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.