To develop and validate a 3D hybrid deep learning model (3D C-Vit) for accurate grading of pediatric brain tumors, emphasizing its performance and interpretability for clinical use.
Key Findings:
3D C-Vit model achieved AUC of 91.36%, accuracy of 86.53%, and F1-score of 89.29 on the test set, significantly outperforming traditional clinical models.
Ablation studies showed significant performance improvements from CAEFF, MSFE, and MHSA modules.
3D C-Vit outperformed traditional clinical models and existing radiomics models in all assessment metrics.
Interpretation:
The 3D C-Vit model effectively combines local feature extraction and global modeling, enhancing grading accuracy and providing reliable interpretability for clinicians, which aids in informed decision-making.
Limitations:
Retrospective nature may introduce bias, potentially affecting the generalizability of the findings.
Limited to specific MRI sequences and pediatric population, which may not represent all cases.
Conclusion:
The 3D C-Vit model is a promising tool for automatic grading of pediatric brain tumors, significantly improving upon traditional methods and aiding in personalized treatment planning.
