Interpretable machine learning and deep learning model for discriminating pheochromocytoma from adrenocortical adenoma based on CT: A multicenter study

Background Differentiating pheochromocytoma (PHEO) from adrenocortical adenoma (ACA) is vital to avoid intraoperative hypertensive crises or redundant surgeries. Purpose To identify the optimal diagnostic strategy by benchmarking various machine learning (ML) and deep learning (DL) architectures for the preoperative differentiation of PHEO from ACA. Methods We retrospectively enrolled 401 patients from two centers. Center 1 (n = 331) was divided into training (n = 232), validation (n = 49), and internal test (n = 50) sets; Center 2 (n = 70) served as an external validation cohort. From dual-phase CT, 1,781 radiomics features were extracted. We benchmarked eight ML classifiers against a SqueezeNet-based DL model. Performance was quantified using AUC and decision curve analysis. Shapley Additive Explanations (SHAP) were applied to interpret the best-performing model and construct a clinical nomogram. Results The Logistic Regression (LR)-based integrated model outperformed other classifiers and the DL baseline (AUC: 0.850, 95% CI: 0.838–0.869)), achieving an AUC of 0.917 (95% CI: 0.820–0.990) with an accuracy of 86.4% in the internal test set and 0.909 (95% CI: 0.7995–1.000) in external validation. Subgroup analysis showed high diagnostic stability, with an AUC of 0.933 (95% CI: 0.821–1.000) for tumors ≥ 4 cm. SHAP analysis identified dependence non-uniformity as a key diagnostic driver. The resulting nomogram provided personalized, transparent risk estimation. Conclusion The SHAP-interpreted integrated LR model is a robust, non-invasive tool for distinguishing PHEO from ACA. By combining radiomic and clinical data, this transparent framework supports precise surgical planning and individualized management.