Computed Tomography Image-Based Deep Survival Regression for Metastatic Colorectal Cancer Using a Non-proportional Hazards Model

Abstract

With more than 1,800,000 cases and over 862,000 deaths per year, metastatic colorectal cancer is the second leading cause of cancer related deaths in modern societies. The estimated patient survival is one of the main factors for therapy adjustment. While proportional hazard models are a key instrument for survival analysis within the last centuries, the assumption of hazard proportionality might be overly restrictive and their applicability to complex data remains difficult. Especially the integration of image data comes at the cost of a careful pre-selection of hand-crafted features only. With the rise of deep learning, directly differentiable models for survival analysis have been developed. While some inherit the difficulties of the proportionality assumption, others are restricted to scalar data input. Computed-tomography-based survival analysis remains a hardly researched topic at all. We propose a deep model for computed-tomography-based survival analysis providing a hazard probability output representation comparable to Cox regression without relying on the hazard proportionality assumption. The model is evaluated on multiple datasets, including metastatic colorectal cancer computed tomography imaging data, and significantly reduces the average prediction error compared to the Cox proportional hazards model.