An interpretable machine learning model for identifying granulation patterns in somatotroph tumors: A multi-center radiomics study

Background: To develop non-invasive, interpretable machine learning (ML) models using radiomic and clinical -features to distinguish sparsely granulated (SGST) and densely granulated (DGST) somatotroph tumors by leveraging preoperative multiparametric magnetic resonance imaging (MRI) and clinical data. Methods: We retrospectively analyzed 201 patients (107 DGST, 94 SGST) with surgically treated somatotroph tumors across four institutions (156 in the training cohort, 45 in the external validation cohort). From preoperative contrast-enhanced T1 and T2-weighted MRI scans, 3004 radiomic features were extracted using Pyradiomics. Feature selection involved least absolute shrinkage and selection operator (LASSO), and the Boruta algorithm. Six ML-algorithms were evaluated, with the support vector machine (SVM) model selected for tumor subtype classification. Shapley Additive Explanations (SHAP) enhanced model interpretability by ranking feature importance. Results: The SVM model, integrating eight radiomic features and five clinical factors, achieved areas under the curve (AUCs) of 0.828 in the internal training cohort and 0.820 in the external validation cohort. SHAP analysis identified key radiomic and clinical predictors, enhancing the model's transparency and clinical applicability. Conclusion: This multi-center study validates an interpretable, radiomics-based SVM model with high accuracy and generalizability for preoperative classification of somatotroph tumor granulation patterns. By offering a non-invasive tool to predict tumor subtypes, this approach enhances personalized treatment planning and holds translational potential for improving acromegaly management.