Computed Tomography-Based Machine Learning Differentiates Adrenal Pheochromocytoma From Lipid-Poor Adenoma

Abstract

Objectives: To assess the accuracy of computed tomography (CT)-based machine learning models for differentiating subclinical pheochromocytoma (sPHEO) from lipid-poor adenoma (LPA) in patients with adrenal incidentalomas.

Patients and methods: The study included 188 tumors in the 183 patients with LPA and 92 tumors in 86 patients with sPHEO. Pre-enhanced CT imaging features of the tumors were evaluated. Machine learning prediction models and scoring systems for differentiating sPHEO from LPA were built using logistic regression (LR), support vector machine (SVM) and random forest (RF) approaches.

Results: The LR model performed better than other models. The LR model (M1) including three CT features: CT_pre value, shape, and necrosis/cystic changes had an area under the receiver operating characteristic curve (AUC) of 0.917 and an accuracy of 0.864. The LR model (M2) including three CT features: CT_pre value, shape and homogeneity had an AUC of 0.888 and an accuracy of 0.832. The S2 scoring system (sensitivity: 0.859, specificity: 0.824) had comparable diagnostic value to S1 (sensitivity: 0.815; specificity: 0.910).

Conclusions: Our results indicated the potential of using a non-invasive imaging method such as CT-based machine learning models and scoring systems for predicting histology of adrenal incidentalomas. This approach may assist the diagnosis and personalized care of patients with adrenal tumors.

Keywords: adrenal incidentaloma; computed tomography; lipid-poor adenoma; logistic regression; machine learning; subclinical pheochromocytoma.