Analysis of Information Compression of Medical Images for Survival Models

Abstract

In this paper, we investigate the methods for extraction of significant information from medical breast cancer images for survival analysis. In breast cancer diagnostics as well as in the analysis of other medical images it is still common to employ manual assessment by medical staff, which needs a lot of prior professional knowledge. Unfortunately, such feature engineering is difficult to manage, reproduce and it also depends on each specific task and expert experience. In this study, we analyze how high-level features can be used for survival models. We create unsupervised learning models for information compression to bottlenecks via convolutional neural networks (CNN) and autoencoders (AE) to obtain the informative covariates. Then we use these image-related covariates in Cox proportional hazards regression. We demonstrate that unsupervised methods allow extracting meaningful covariates that are significant for survival analysis without using explicit feature engineering or image labeling. We run the experiments on TCGA (The Cancer Genome Atlass) breast cancer dataset.