Reproducibility and interpretability in radiomics: a critical assessment

Radiomics aims to improve clinical decision making through the use of radiological imaging. However, the field is challenged by reproducibility issues due to variability in imaging and subsequent statistical analysis, which particularly affects the interpretability of the model. In fact, radiomics extracts many highly correlated features that, combined with the small sample sizes often found in radiomics studies, result in high-dimensional datasets. These datasets, which are characterized by containing more features than samples, have different statistical properties than other datasets, thereby complicating their training by machine learning and deep learning methods. This review critically examines the challenges of both reproducibility issues and interpretability, beginning with an overview of the radiomics pipeline, followed by a discussion of the imaging and statistical reproducibility issues. It further highlights how limited model interpretability hinders clinical translation. The discussion concludes that these challenges could be mitigated by following best practices and by creating large, representative, and publicly available datasets.