Interpretability and performance of a 3D C-vit model for accurate grading of pediatric brain tumors

Background Accurate preoperative grading of pediatric brain tumors is crucial for formulating individualized treatment plans. Traditional methods rely on subjective experience, while existing deep learning models have limitations in capturing long-distance dependencies and local details. This study aims to develop and validate an innovative 3D hybrid deep learning model (3D C-Vit) for pediatric brain tumor grading and analyze its performance and interpretability. Methods This retrospective study included 340 cases of pediatric brain tumors (143 low-grade cases and 197 high-grade cases). Tumor regions were independently annotated by two senior radiologists with consistency achieved. The data were divided into training, validation, and test sets in a ratio of 70:15:15. The model input included five MRI sequences: CE-T1WI, T1WI, T2WI, FLAIR, and ADC. The proposed 3D C-Vit model integrates the Channel Attention-Enhanced Feature Fusion (CAEFF) module, Multi-Scale Feature Extraction (MSFE) module, and Multi-Head Self-Attention (MHSA) mechanism. Model performance was evaluated using AUC, accuracy (ACC), precision, recall, and F1-score. Chi-square test and LASSO regression were used for feature selection and interpretability analysis. Results The 3D C-Vit model performed optimally on the test set: AUC was 91.36%, ACC was 86.53%, and F1-score was 89.29. Ablation experiments confirmed that CAEFF, MSFE, and MHSA modules increased ACC by 6.92%, 11.67%, and 1.64%, respectively, and AUC by 6.79%, 11.14%, and 1.66%, respectively. Among the radiomics models, LASSO regression screened out 59 key features. The 3D C-Vit model was significantly superior to the clinical model (ACC 69.23%, AUC 79.09%) and the best radiomics models (SVM, ACC 77.55%, AUC 86.14%) in all assessment metrics. Conclusion The 3D C-Vit model proposed in this study can effectively and automatically grade pediatric brain tumor, and its performance significantly surpasses traditional clinical methods and existing radiomics models. The model combines the local feature extraction capability of CNN with the global modeling advantage of Transformer and effectively improves the grading accuracy through the innovative CAEFF and MSFE modules. Its high accuracy and interpretability provide clinicians with a reliable preoperative tumor grading tool, which is helpful for quickly formulating precise individualized treatment plans.