Radiomics validation for non-invasive characterisation and surgical decision-making in adrenal disease

Abstract Background Adrenal neoplasia is increasingly detected, yet differentiating benign from potentially or overtly malignant lesions on conventional imaging remains challenging. Diagnostic uncertainty can lead to extended follow-up or surgery. Radiomics offers a non-invasive approach to lesion characterisation using quantitative imaging features. Method This retrospective study analysed preoperative CT scans from 134 patients who underwent 135 adrenalectomies for functional or nonfunctional adrenal tumours. Neoplasia and normal adrenal parenchyma were manually 3D-segmented, radiomic features extracted using PyRadiomics, and key variables selected via Minimum Redundancy Maximum Relevance. Multiple machine-learning (ML) models were trained on 91 cases to classify lesions as benign (B), potentially malignant (PM) or malignant (M). Endpoints included diagnostic accuracy (mean validation area under the receiver operating characteristic curve, AUC) of the best-performing model against expert prediction and final histopathology, and the study’s Radiomics Quality Score (RQS). Results The ExtraTreesClassifier model showed the highest diagnostic performance, with a test AUC of 0.7189 (B vs. PM/M). Using CT images without biochemical data, the expert surgeon’s predictive accuracy in the test set (n=44) exceeded that of ML (PPV 0.8125 vs. 0.5385, p0.027; NPV 0.7857 vs. 0.6129, p0.0317). The overall radiomics pipeline achieved a RQS of 16/36, the highest score in adrenal radiomics to date. Conclusion Radiomics shows potential for non-invasive characterisation of adrenal lesions and may support surgical decision-making. Nonetheless, important limitations including retrospective design, modest cohort size, and lack of standardised radiomics methods underscore the need for larger prospective validation and multimodal data integration.