Using Convolutional Neural Networks to Predict Colon Cancer Patients Survival

This paper aims to predict colon cancer patients’ survival by using deep learning to extract prognostic biomarkers from haematoxylin- and eosin (HE)-stained tissue slides. A deep convolutional neural network is trained by transfer learning using 100,000 HE images achieving a nine-class accuracy \({>}97\%\). This model is then used to segment digital whole slide images from a cohort of patients from the Cancer Genome Atlas (TCGA). The classification map produced is then used to quantify tumour–stroma ratio and tumour-infiltrating lymphocytes regions. These are then evaluated for their prognostic value for overall survival (OS) in a multivariate Cox proportional hazard model.