Multi-model radiomics and machine learning for differentiating lipid-poor adrenal adenomas from metastases using...

Radiomics based on automatic segmentation of CT images has emerged as a highly promising approach for differentiating adrenal adenomas from metastases in clinical practice; however, its preoperative diagnostic value has not been fully evaluated in previously developed methodologies.

To fully elucidate the diagnostic value of radiomics based on automatic segmentation techniques in differentiating adrenal adenomas from metastases through a retrospective analysis of clinical and contrast-enhanced CT (CECT) data.

A retrospective analysis was conducted on the clinical and imaging data of 416 patients with adrenal masses larger than 10 mm, who had clinically indicated contrast-enhanced CT (CECT) examinations at our hospital between January 2020 and June 2024. Adrenal lesions were segmented automatically using 3D Slicer, and radiomic features were extracted from the segmented arterial and venous phase images using PyRadiomics. Feature selection and dimensionality reduction were performed using mutual information (MI), minimum redundancy maximum relevance (MRMR), LASSO, and Pearson correlation analysis. Clinical and imaging features were then incorporated into an XGBoost machine learning model, and model performance was evaluated using Area Under Curve (AUC), accuracy, precision, sensitivity, specificity, and F1 score. SHAP analysis was used to interpret the model's predictions and identify the most influential features.

This study included 221 adenomas and 195 metastases. Significant differences were observed between the two groups in terms of age, lesion size, and contrast washout rate (P < 0.001). After feature extraction, selection, and dimensionality reduction, 15 arterial phase features, 6 venous phase features, and 18 combined features were used for model training. The AUC values of the XGBoost model for the arterial phase, venous phase, combined arterial and venous phase data, and combined arterial, venous phase, and clinical indicators were 0.81, 0.81, 0.88, and 0.92, respectively. Five-fold cross-validation showed that the average scores of XGBoost were 0.868, 0.823, 0.897, and 0.89, respectively. SHAP summary plot for each sample under different features were used to illustrate the interpretability of the model.

A machine learning model, combining multimodal CT radiomics and automatic segmentation technology, enables machine-based clinical features extraction, improves the differentiation between adrenal adenomas and metastases, and provides a reliable foundation for accurate diagnosis and treatment planning.