Mutational Analysis and Deep Learning Classification of Uterine and Cervical Cancers

Abstract

We analyzed tumor mutations of 7 uterine and 2 cervical cancers with the goal of developing a Deep Learning (DL) software tool that can automatically classify tumors based on their somatic mutations. The data were obtained from the AACR Genie Project, that has a collection of more than 120,000 tumor samples for more than 750 cancer types. We performed a thorough analysis of the mutational data of tumors of the uterus and uterine cervix, selecting tumors with 3 or more mutations and cancer types with more than 15 cases. For each cancer type we then selected the top 12 most mutated genes among their neoplasms. In the introduction section we summarize our analysis of these nine diseases and in the methods section we present a convolutional neural network (CNN) that yields an overall classification accuracy of 94.3% and 89.2% on the train and test datasets, respectively. We hope this tool can be added to the existing arsenal of histological and immunohistochemical techniques in cases when a precise diagnosis cannot be clearly determined. Each cancer type has a unique somatic mutational profile that can be used to disambiguate two candidate malignancies with similar histologic features.