Classification of adrenocortical carcinoma, pheochromocytoma, and adrenal adenomas using contrast-enhanced computed tomography with machine learning and texture features: a cross-sectional study

BACKGROUND: Differential diagnosis of adrenocortical carcinoma, pheochromocytoma, and adrenal adenomas based on contrast-enhanced computed tomography remains challenging because of substantial overlap in their radiologic characteristics. Existing classification approaches based on conventional morphological criteria demonstrate limited accuracy, which may result in misdiagnosis and inappropriate treatment strategies.

AIM: This study aimed to develop a machine learning model for multiclass classification of adrenal lesions (adenomas, adrenocortical carcinoma, and pheochromocytoma) using contrast-enhanced computed tomography data with texture features.

METHODS: This was a single-center, cross-sectional study with retrospective computed tomography data acquisition and prospective re-analysis of imaging results. Contrast-enhanced computed tomography images were processed using PyRadiomics to extract texture features for each computed tomography phase. Data standardization was performed to reduce the impact of variability in scanning parameters. LightGBM, XGBoost, and CatBoost gradient boosting models were trained using stratified five-fold cross-validation. Diagnostic performance was assessed using recall, precision, F1-score, macro-averaged F1-score, specificity, balanced accuracy, and area under the receiver operating characteristic curve (AUC) for each diagnostic category.

RESULTS: The study included data from 425 patients with histologically verified adrenal tumors: 42 cases of adrenocortical carcinoma, 204 pheochromocytomas, and 179 adrenal adenomas. The developed machine learning models demonstrated high classification performance by cross-validation for adrenal adenomas (F1-score up to 0.916 for the XGBoost model) and pheochromocytomas (F1-score up to 0.855 for the XGBoost model), but substantially lower performance for adrenocortical carcinoma (F1-score up to 0.521 for the CatBoost model). The highest AUC values reached 0.971 for adenomas (LightGBM), 0.924 for pheochromocytomas (LightGBM), and 0.879 for adrenocortical carcinoma (CatBoost). Balanced accuracy reached up to 0.773, and the macro-averaged F1-score reached 0.747 (CatBoost model). Analysis of the most informative features showed that parameters reflecting texture homogeneity and intensity across different contrast-enhancement phases were most relevant for classification.

CONCLUSION: Radiomics and machine learning methods provide high diagnostic accuracy for multiclass classification of adrenal lesions on contrast-enhanced computed tomography for adrenal adenomas and pheochromocytomas. However, diagnostic performance for adrenocortical carcinoma remains limited, which may be related to tumor heterogeneity and the relatively small number of cases.