The pitfalls of fixed-ratio data splitting in radiomics model performance evaluation.

Abstract

Over the past decade, radiomics has seen exponential growth, with over ten thousand publications in PubMed and a steady increase in related studies in journals like Abdominal Radiology. Despite the potential of radiomics, a major challenge lies in validating radiomics models, as most studies rely on single-center datasets with fixed-ratio splits, which can lead to variability in performance due to randomness in data splitting. Therefore, researchers should adopt more robust cross-validation methods rather than relying solely on the fixed-ratio holdout method to ensure robust and reliable radiomics model performance evaluation.