Deep Learning and Histopathology images can be used to predict BRCA gene mutations in breast cancer.
Background: Breast cancer, the most common form of cancer and the number one cause of cancer-related death among women in the world, is a very serious disease. A mutation in a particular gene, such as BRCA1/2, may be linked to a genetic predisposition for breast cancer. Patients with germline pathogenic mutations in BRCA1/2 genes are at a higher risk of breast cancer. They may benefit from targeted treatment. Genetic testing can be time-consuming and expensive. This study aims at predicting the risk of gBRCA by using whole-slide features of breast cancer H&E stains, and the patients’ gBRCA status.
Methods: We trained a deep convolutional network (CNN), based on ResNet, to predict the gBRCA mutatation in breast cancer using whole-slide image (WSI) data. We divided the WSIs into smaller tiles to train the CNN, as the dimensions were too large for a slide-based approach. The positive classification results were added to the tile-based classification in order to calculate the combined accuracy of the slide-based training. The models were trained based upon the tumor location and gBRCA status annotated by a breast cancer pathologist. Four models were trained using tiles cropped to 5x,10x,20x and 40x magnification. This was done because it was assumed that high magnification would provide more information than low magnification.
Results: A trained models was validated using an external dataset that contained 17 mutants and 48 wilds. The AUCs (95 % CI) for DL models using 40x, 20x and 10x magnification tiles were respectively 0.766 (0.760-0.769), 0.750 (0.737-0.761), and 0.550 (0.526-0.575). For the magnification slides, the AUCs (95 % CI) were respectively 0.774 (0.642-0.905), 0.804 (0.760-0.931), 0.828 (0.691-0.966), The study also determined the impact of histological grade on the accuracy of prediction.
Source:
https://www.frontiersin.org/articles/10.3389/fgene.2021.661109/full
Leave a Reply