Interpretability and Repeatability of Radiomic Features: Applied on In Vivo Tumor Models

Radiomic features are typically used in machine learning models and are proven to generate reliable results when predicting tumor grade and responses to treatment. However, the inherent non-biological-interpretability of the radiomic features strongly hinders their clinical application. Therefore, it is of pivotal importance to elucidate the biological meaning behind the given radiomic features. In this article, an innovative approach is proposed where dedicated in vivo experiments are used to correlate biological meaning to specific radiomic features. As a proof of concept, the radiomic features extracted from the computed tomography (CT) scans of three widely used and well-characterized murine tumor models (CT26, 4T1, and EMT6) were analyzed and compared using an exploratory factor analysis (EFA). The results revealed that on the basis of the features, a distinction could be made between the different tumor models. Furthermore, the effect of an inflammatory response on the radiomic features was investigated. Lastly, the repeatability of radiomic features upon modulation of the tumor microenvironment (TME) was analyzed. The features exhibited a high repeatability level over the course of time, displaying consistency between the different experiments. Altogether, these encouraging results support the feasibility of the proposed approach to pave the way for the use of radiomics in routine clinical practice.