An interpretable machine learning framework using multi-phase computed tomography for differentiation of adrenal lipid-poor adenomas and pheochromocytomas

Abstract

Objective To develop and validate an interpretable machine learning (ML) framework that differentiates adrenal lipid-poor adenomas (LPAs) from pheochromocytomas (PHEOs) using radiomic features derived from multi-phase contrast-enhanced computed tomography (CECT). Methods This retrospective study included 229 patients with pathologically confirmed adrenal tumours (132 LPAs, 97 PHEOs). Lesions were stratified into washout (absolute percentage washout [APW] > 60%) and non-washout (APW ≤ 60%) cohorts according to established criteria. We trained decision tree (DT), support vector machine (SVM), and logistic regression (LR) models using a predefined set of clinically relevant CT features. The inherently interpretable DT model was further dissected to revealr key discriminative features and its underlying decision logic. Model performance was evaluated using the area under the curve (AUC) with 95% confidence intervals. Results The DT model achieved excellent diagnostic performance, with an AUC of 0.977 (95% CI: 0.934–1.000) in the washout group and 0.983 (95% CI: 0.962–1.000) in the non-washout group. Feature importance analysis identified cystic degeneration as the most influential discriminator, followed by enhancement potential (EP) and baseline attenuation (CTu). The DT’s hierarchical decision rules provided clear and clinically transparent pathways. Conclusion This interpretable ML framework accurately distinguishes LPAs from PHEOs. The DT model strikes an optimal balance between high diagnostic accuracy and inherent interpretability, offering a practical tool that may enhance clinical confidence when managing indeterminate adrenal lesions.