An interpretable

Purpose: To evaluate the potential of an interpretable radiomics model based on Materials and methods: A total of 177 patients with extra-adrenal malignancies and indeterminate adrenal nodules (74 metastases; 103 benign lesions) were included and randomly assigned to training and testing sets in a 7:3 ratio. Radiomics features were extracted separately from the CT and PET components of PET/CT examinations. Least Absolute Shrinkage and Selection Operator (LASSO) and multivariate logistic regression (LR) were used to identify independent predictive radiomics factors. Based on these features, single-modality CT, PET, and combined PET/CT radiomics models were constructed using four machine learning algorithms: Random Forest (RF), Support Vector Machine (SVM), LR, and Decision Tree (DT). The best-performing algorithm for each modality determined through cross-validation was selected to establish the final models. Model performance was assessed using the area under the receiver operating characteristic curve (AUC) and decision curve analysis (DCA). DeLong test was used to compare the AUCs between models. Internal validation of the best-performing radiomics model was conducted by bootstrapping to assess potential optimism. Shapley Additive Explanations (SHAP) was utilized to interpret the best-performing radiomics model. Results: The optimal algorithms identified were LR for the CT model and SVM for both the PET and integrated PET/CT models. In the testing set, the AUC values were 0.811 (95% CI: 0.694-0.928) for the CT model and 0.879 (95%CI: 0.789- 0.970) for the PET model. The combined PET/CT model integrating both CT and PET radiomics features achieved an AUC of 0.915 (95%CI: 0.834-0.997), which was significantly higher than that of the CT model alone (p < 0.05). DCA confirmed superior clinical utility of the combined PET/CT model across most threshold probabilities compared to the single-modality models. Bootstrap-corrected internal validation showed an optimism-corrected AUC of 0.919 (95% CI: 0.884-0.964), with minimal observed optimism (0.003, 95% CI: -0.002-0.007). SHAP analysis showed that a texture feature derived from the gray level size zone matrix of PET images was the most significant predictor of AMs. Conclusions: The interpretable radiomics model based on combined PET/CT data provides a non-invasive tool for predicting AMs in cancer patients with indeterminate adrenal nodules. By integrating features from both modalities, this approach significantly improves diagnostic performance and holds strong potential to support personalized treatment.