Comparison of deep learning architectures for colon cancer mutation detection

Colorectal cancer is responsible of the death of hundred of thousands of people worldwide each year. The histopathological features of the tumor are generally identified from the analysis of tissue taken from a biopsy providing information for selecting the adequate treatment. With the advent of digital pathology, slides of tissue are increasingly available as Whole Slide Images (WSI) allowing their analysis using artificial intelligence algorithms. Additionally, genetic sequencing of the tumor can also be performed to identify specific mutations (e.g. KRAS) and their subtypes (e.g. G12C). While sequencing can take time and is costly, it provides key information on the specific type of cancer. In this paper, we target the identification of key genes and variants using only stained histopathology images using deep neural networks. Predicting gene information only from histopathology images could be easily used routinely and would allow saving the time and cost of tumor sequencing. Experiments performed on 45 colorectal cancer WSI from 45 patients revealed that deep models can correctly classify 90% of patients with the KRAS G12C mutation. Furthermore, we also used CAM and LIME methods to explore the interpretability of the results highlighting which parts of the images were used by the models to predict the presence of specific mutations.