Comparison of Radiomics Models Built Through Machine Learning in a Multicentric Context With Independent Testing: Identical Data, Similar Algorithms, Different Methodologies

Abstract

Machine learning techniques are becoming increasingly popular in radiomics studies. They can handle high dimensional sets of radiomics features with higher robustness than usual statistical analyses, by capturing complex interactions between features themselves and between feature combinations and clinical endpoints under investigation in order to build efficient prognostic/predictive models. However, there is no “one fits all” solution and deciding which algorithm is the most accurate for a given application is not always straightforward. In this paper, to keep a realistic perspective on various emerging clinical applications based on radiomics, we performed an evaluation of the popular random forest classifier for predicting local failure in cervix cancer exploiting identical data, but relying on different methodologies to select and combine features of interest. The main objective was to demonstrate various challenges of model building and tuning for radiomics applications. The results obtained in the present work could provide general guidelines to assist in the practical development of radiomics-based models.