| Autor: |
Bourgade, Raphaël, Rabilloud, Noémie, Perennec, Tanguy, Pécot, Thierry, Garrec, Céline, Guédon, Alexis F., Delnatte, Capucine, Bézieau, Stéphane, Lespagnol, Alexandra, de Tayrac, Marie, Henno, Sébastien, Sagan, Christine, Toquet, Claire, Mosnier, Jean-François, Kammerer-Jacquet, Solène-Florence, Loussouarn, Delphine |
| Zdroj: |
Modern Pathology; November 2023, Vol. 36 Issue: 11 |
| Abstrakt: |
BRCA1and BRCA2 genes play a crucial role in repairing DNA double-strand breaks through homologous recombination. Their mutations represent a significant proportion of homologous recombination deficiency and are a reliable effective predictor of sensitivity of high-grade ovarian cancer (HGOC) to poly(ADP-ribose) polymerase inhibitors. However, their testing by next-generation sequencing is costly and time-consuming and can be affected by various preanalytical factors. In this study, we present a deep learning classifier for BRCAmutational status prediction from hematoxylin-eosin-safran–stained whole slide images (WSI) of HGOC. We constituted the OvarIA cohort composed of 867 patients with HGOC with known BRCAsomatic mutational status from 2 different pathology departments. We first developed a tumor segmentation model according to dynamic sampling and then trained a visual representation encoder with momentum contrastive learning on the predicted tumor tiles. We finally trained a BRCAclassifier on more than a million tumor tiles in multiple instance learning with an attention-based mechanism. The tumor segmentation model trained on 8 WSI obtained a dice score of 0.915 and an intersection-over-union score of 0.847 on a test set of 50 WSI, while the BRCAclassifier achieved the state-of-the-art area under the receiver operating characteristic curve of 0.739 in 5-fold cross-validation and 0.681 on the testing set. An additional multiscale approach indicates that the relevant information for predicting BRCAmutations is located more in the tumor context than in the cell morphology. Our results suggest that BRCAsomatic mutations have a discernible phenotypic effect that could be detected by deep learning and could be used as a prescreening tool in the future. |
| Databáze: |
Supplemental Index |
| Externí odkaz: |
|
