Risk Prediction in Cancer Imaging Using Enriched Radiomics Features

Abstract

Diagnostic imaging has gained prominence as potential biomarkers for early detection and diagnosis in a diverse array of disorders including cancer. However, existing methods routinely face challenges due to costs and technical performance in detecting small tumors and predicting risk, and heterogeneity. Radiomics, which extracts summary features from images, is widely used in cancer imaging research but has yet to be demonstrated in a systematic fashion to be a reliable tool in clinical settings. We introduce a novel class of enriched radiomics features that combine traditional summary statistics with the probability distribution of pixel-level features extracted from enhancement pattern mapping (EPM), images derived from MRI scans of the liver. Using these enriched features, we develop a classification model for predicting clinical diagnosis, LIRADS score, and longitudinal tumor progression from baseline images. The approach is evaluated against existing methods using a liver cancer dataset. Classification based on enriched radiomics features illustrates AUC greater than 0.8 for all tasks, that is significantly greater than AUC based on classical radiomics summary measures. Further, significant longitudinal changes in the liver EPM images are shown to be associated with change in LIRADS score at future visits, implying that changes in the liver EPM images are indicative of tumor progression. The proposed approach can be generalized to various types of cancer imaging applications, and can serve as a promising alternative to classical radiomics analysis.