ABSTRACT Radiomics, the practice of mining quantitative features from medical imaging, has gained increasing popularity in diagnosing and treating cancer. However, clinical deployment and adoption of radiomics tools in real‐world practice remain limited. To increase trust in radiomics models, one of the main remaining challenges is ensuring fairness by addressing algorithmic bias and promoting equitable outcomes. This study aims to identify key fairness challenges in radiomics and demonstrate how state‐of‐the‐art mitigation techniques can be applied to address them. Moreover, we aim to provide new guidelines for both the selection of mitigation techniques and procedural recommendations for future directions in fairness. This study evaluates the fairness of radiomics models using six machine learning algorithms applied to four publicly available oncology datasets: Liver, Gastrointestinal, Desmoid‐type fibromatosis, and Liposarcoma. Fairness was assessed using Equal Opportunity Difference (EOD) with respect to two protected attributes: age and sex. We employed three bias mitigation techniques—Disparate Impact Remover (DIR), Reweighting, and Reject Option Classification (ROC)—to examine their effect on fairness metrics and model performance. DIR, Reweighting, and ROC each showed dataset‐ and attribute‐specific improvements in fairness. For example, DIR notably reduced sex‐related bias in the Liposarcoma dataset, while ROC offered the best fairness–performance trade‐off in the Liver dataset. Reweighting was effective in mitigating age‐related bias in Liposarcoma. In the Desmoid dataset, applying ROC to Extra Trees marginally improved fairness; however, age‐related bias persisted. Overall, the applied techniques reduced disparities in true positive rates between groups, enhancing fairness. Our findings show that while radiomics models can be accurate, they often exhibit bias with respect to demographic groups. Carefully addressing this trade‐off is crucial for developing radiomics models that are accurate but also trustworthy for clinical deployment.
Pujadas et al. (Sun,) studied this question.