Multimodal clinical-radiomic-deep learning model for preoperative classification of pediatric pineal region tumors.

BACKGROUND: Accurate preoperative differentiation among germinomas, non-germinomatous germ cell tumors (NGGCTs), and pineal parenchymal tumors (PPTs) is clinically crucial for treatment strategies. OBJECTIVES: To develop and compare predictive models integrating clinical, radiomic, and deep learning (DL) features for the preoperative classification of pediatric pineal region tumors. METHODS: This retrospective study included 280 pediatric patients with pineal region tumors (94 germinomas, 119 NGGCTs, and 67 PPTs), randomly divided into training and testing cohorts (7:3 ratio). Clinical features included demographic, imaging, and laboratory variables. Radiomic and 2.5D DL features were extracted from multiparametric MRI (T1WI, T2WI, CE-T1WI). After z-score normalization and feature selection with Pearson correlation and least absolute shrinkage and selection operator (LASSO) regression, logistic regression classifiers were built to develop clinical, radiomics, DL, and combined clinic-radiomic-deep learning (CRDL) models. Class weights inversely related to sample frequencies were used to balance the classes during training. Model performance was evaluated using ROC analysis, with DeLong's test for AUC comparison and decision curve analysis (DCA) for clinical utility. A reader study assessed diagnostic improvement with model assistance. RESULTS: The CRDL model achieved the best overall diagnostic performance with a macro-average AUC of 0.912, statistically outperforming the other models. DCA further validated the clinical value of the CRDL model in preoperative tumor classification. Model-assisted interpretation improved diagnostic accuracy for both junior and senior radiologists. CONCLUSION: The proposed CRDL model showed promising performance for the noninvasive preoperative classification of pediatric pineal region tumors and may provide a potential decision-support tool for treatment planning.