le 23/03/2023 à 14h
Centrale Nantes, LS2N, salle du conseil.

Raphaël Bourgade viendra nous donner un séminaire et sera accompagné de Delphine LOUSSOUARN, médecin pathologiste au CHU en charge du projet de pathologie numérique.
Ils nous présenteront le service d’anatomopathologie et leurs activités de recherche en pathologie computationnelle.

Title: Deep learning for detecting BRCA 1/2 mutations in high-grade ovarian cancer based on an innovative tumor segmentation method from whole-slide-images

Abstract: BRCA1/2 mutations are widely recognized as effective predictors of the sensitivity of high-grade ovarian cancer to poly(ADP-ribose) polymerase inhibitors (PARPi). Their identification is crucial for selecting appropriate therapeutic options and is currently performed using next-generation sequencing (NGS) on DNA extracted from formalin-fixed paraffin-embedded (FFPE) tissue samples. However, this testing is costly, time-consuming, and can be affected by various preanalytical factors, such as low quality sample or DNA fragmentation. Several studies have thus explored the interest of deep learning to predict gene mutations from pathology images. In this work, we present a novel approach for predicting BRCA1/2 mutations in epithelial ovarian cancer (EOC) using deep learning. We first trained an accurate ovarian cancer segmentation model following an innovative and efficient sampling method, achieving a Dice score of X and IoU of X with only 2 hours of manual annotations. We then trained a deep-learning model to predict BRCA 1/2 mutations with high accuracy (area under the receiver-operator characteristics curve X). Our results suggest that the relevant information for predicting BRCA mutations is more related to the spatial conformation of the tumor rather than to cell and nuclear morphology.