Automatic segmentation-based radiomics and deep learning combined with clinical parameters for precise differentiation of lipid-poor adrenal adenomas and metastases.

objectiveTo explore the application value of a combined model automatic segmentation-based radiomics and deep learning, integrated with clinical parameters, in differentiating adrenal lipid-poor adenomas and metastases.

methodsThis study included totally 535 patients with adrenal tumors. The internal cohort comprised 418 patients, who were randomly divided into a training set (n = 292) and an internal validation set (n = 126) at a 7:3 ratio. A clinical prediction model was developed using multivariate XGBoost algorithm. After automatic segmentation of all images, the volume of interest(VOI) of the lesions was obtained. Radiomic features were extracted from contrast-enhanced abdominal CT images and combined with deep learning signatures. Feature selection and dimensionality reduction were performed using statistical tests, Least Absolute Shrinkage and Selection Operator (LASSO), and Pearson correlation analysis. Traditional radiomics models, a ResNet50 deep-learning model based on convolutional neural networks, and combined prediction models integrating clinical features were constructed. Model performance was evaluated using metrics such as the area under the curve (AUC), sensitivity, and specificity. An external test set of 117 adrenal tumor patients was used to validate the predictive performance of each model. Finally, SHAP plots were applied to enhance model interpretability and quantify the impact of individual features.

resultsThe combined prediction model integrating radiomics, deep-learning features, and clinical parameters outperformed single-feature models, with AUCs of 0.996, 0.939, and 0.852 in the training, internal validation, and external test sets, respectively. This combined model offers superior diagnostic performance and greater clinical benefit.

conclusionThe combined prediction model, which integrates radiomics and deep learning features derived from automatic segmentation with clinical parameters, demonstrates strong predictive value for distinguishing between lipid-poor adrenal adenomas and metastases.