Open Collections

Using convolutional neural networks to predict NRG1-fusions in PDAC biopsy images

University of British Columbia

Pancreatic ductal adenocarcinoma (PDAC) is considered the most lethal common cancer, with the highest incidence-to-mortality ratio of any solid tumour. Molecular pathology studies and genomic analyses have improved our understandings of how PDAC develops and progresses, and there has been significant progress in treatment strategies for specific genomic alterations. One of these alterations is the NRG1 gene fusion, which has been found to be a rare, but potentially targetable oncogenic driver. To determine whether PDAC patients have an NRG1 gene fusion, we used convolutional neural networks (CNNs) to analyze digital whole slide images (WSIs) of cancer biopsies. In particular, we used histopathological H&E slides from the Personalized OncoGenomics program to train a deep CNN (VGG-16) framework that automatically classifies normal tissue, NRG1-fusion positive tumour tissue, and NRG1-fusion negative tumour tissue. We implemented the model in two-stages, where the first stage classifies normal from tumour tissue, and the second stage classifies the tumour tissue as being NRG1-fusion positive or negative. The model achieved accuracies of 86.5% and 76.0% for each stage, respectively, and an overall accuracy of 68.8%. Additionally, we found that PDAC cases with high expression of the NRG1 gene (93rd-98th percentile of TCGA PDAC cases) were being classified as NRG1-fusion positive, suggesting a possible correlation between NRG1 gene fusions and high parent gene expression. Finally, we attempted to understand the inner workings and decisions of our CNN model by analyzing internal feature maps. We found activation patterns that matched distinct histological features and compared them with a more traditional image segmentation approach. Overall, our findings demonstrate that deep CNNs have the potential to assist pathologists in detecting therapeutically actionable genomic markers.

Text

2020-09-11

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/75986

Bioinformatics

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/