Classification of adrenocortical carcinoma, pheochromocytomas, and adrenal adenomas based on contrast-enhanced CT images using machine learning

BACKGROUND: Differential diagnosis of adrenocortical carcinoma (ACC), pheochromocytomas, and adenomas based on contrast-enhanced computed tomography (CECT) remains a challenging task due to significant overlap in their radiological characteristics. Existing classification methods based on standard morphological criteria demonstrate limited accuracy, which may lead to misdiagnoses and inappropriate treatment strategies. The implementation of radiomics and machine learning methods has the potential to improve diagnostic accuracy; however, multiclass models covering all three types of lesions remain insufficiently studied.

AIM: Development and evaluation of a machine learning model for multiclass classification of adrenal lesions into three categories (adenomas, ACC, and pheochromocytomas) based on CECT data using texture features.

METHODS: A retrospective study was conducted, including 196 patients with histologically verified adrenal tumors: 28 cases of ACC, 125 pheochromocytomas, and 43 adenomas. CT images were processed using PyRadiomics to extract 106 texture features for each CT phase. To reduce the impact of scanner variability, data harmonization was performed using singular value decomposition (SVD). The XGBoost model was trained using stratified k-fold cross-validation. Model performance was evaluated using macro-averaged accuracy, class-specific accuracy, F1-score, and ROC-AUC metrics.

RESULTS: The model achieved an average accuracy of 0,833 ± 0,083, a macro-averaged F1-score of 0,784 ± 0,096, and a ROC-AUC of 0,899 ± 0,061. The classification accuracy for pheochromocytomas and adenomas was 0,842 ± 0,112 and 0,872 ± 0,089, respectively, while for ACC, it was 0.647 ± 0.098. Analysis of the most informative features indicated that parameters describing homogeneity and intensity across different contrast phases were significant for classification.

CONCLUSION: Radiomics and machine learning methods enable high accuracy in multiclass classification of adrenal lesions based on CECT data. However, the diagnostic accuracy for ACC remains lower, which may be related to tumor heterogeneity and sample size limitations. The results highlight the need for further research, including the integration of clinical data to improve diagnostic accuracy.