Classification and mutation prediction based on histopathology H&E images in liver cancer using deep learning

Abstract

Hepatocellular carcinoma (HCC) is the most common subtype of liver cancer, and assessing its histopathological grade requires visual inspection by an experienced pathologist. In this study, the histopathological H&E images from the Genomic Data Commons Databases were used to train a neural network (inception V3) for automatic classification. According to the evaluation of our model by the Matthews correlation coefficient, the performance level was close to the ability of a 5-year experience pathologist, with 96.0% accuracy for benign and malignant classification, and 89.6% accuracy for well, moderate, and poor tumor differentiation. Furthermore, the model was trained to predict the ten most common and prognostic mutated genes in HCC. We found that four of them, including CTNNB1, FMN2, TP53, and ZFX4, could be predicted from histopathology images, with external AUCs from 0.71 to 0.89. The findings demonstrated that convolutional neural networks could be used to assist pathologists in the classification and detection of gene mutation in liver cancer.

Genomic pan-cancer classification using image-based deep learning

“…Globally, more than 18 million new cancer cases are diagnosed resulting to 9.6 million deaths in 2018 [3] . Genetic mutations have been shown to be associated with different types of cancer [4] , [5] , [6] . Cancer classification based on genetic mutations can be readily achieved through increased usage of high-throughput sequencing techniques.…”

Identifying transcriptomic correlates of histology using deep learning

“…For example, [12] trained a CNN to predict the mutational status of just 10 genes (the most frequently mutated genes in lung cancer, rather than a genome-wide study). Similar mutation predictors were developed for hepatocellular carcinoma [20] and prostate cancer [21]. But the focus of these studies was obtaining mutation predictors, rather than correlating the mutational status with visual histological features.…”

“…Machine Learning has also been used for discovering new clinical-pathological relationships by correlating histo-morphological features of cancers with their clinical evolution, by enabling analysis of huge amounts of data. Thus, relationships between morphological features and somatic mutations [12,20,21], as well as correlations with prognosis have been tentatively addressed. For example, [22] developed a prognosis predictor for lung cancer using a set of predefined image features.…”

Identifying Transcriptomic Correlates of Histology using Deep Learning

“…For example, [Coudray, 2018] trained a CNN to predict the mutational status of just 10 genes (the most frequently mutated genes in lung cancer, rather than a genome-wide study). Similar mutation predictors were developed for hepatocellular carcinoma [Chen, 2020] and prostate cancer [Schaumberg, 2018]. But the focus of these studies was obtaining mutation predictors, rather than correlating the mutational status with visual histological features.…”