Abstract 1632: Deep learning functional associations using histopathology images

Abstract

Abstract Histopathological images have critical value for understanding cancer subtype, local structure, and malignancy. Manual interpretation of such images can be time-consuming and error-prone, while also failing to make use of the power of image aggregation for coordinated data mining. Large-scale patient cohorts such as TCGA provide high-resolution H&E images that complement other -omics data, with the important advantage that images provide spatial information about heterogeneity and stromal cell interactions. We implemented a convolutional neural network (CNN) to classify H&E slides according to tissue type, cancer subtype, and genomic mutations. We modified the CNN architecture of Inception v3 according to our problem and trained it with tiled whole slide images from 31 solid cancers in TCGA, and evaluated our algorithm on holdout data. We first classified tumor/normal status from the images and determined AUC values for each cancer type. We then applied our method for tissue classification in a multilabel classification scenario. Furthermore, in few cancers (e.g. gastric cancer) we determined cancer subtypes using supervised learning. Finally we evaluated our model for classifying mutational status of cancer-related genes. To improve classification accuracy, we modified Inception architecture for multi-task classification by replacing the final fully-connected layer by a shared layer which provides the input to individual task layers corresponding to each gene mutation. Furthermore, we adopted image segmentation approaches (e.g.. auto-thresholding) to improve sample background removal, and used several approaches (e.g. resampling, cross entropy weighting) to rebalance the data. In some cases we achieved high classification results (e.g. AUC=0.7 for TP53 in lung adenocarcinoma). Our study suggests that histological images can be used to classify tumors according to their mutational status, and can provide biomarkers for molecular subtyping. Citation Format: Javad Noorbakhsh, Saman Farahmand, Mohammad Soltanieh Ha, Sandeep Namburi, Kourosh Zarringhalam, Jeff Chuang. Deep learning functional associations using histopathology images [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 1632.