A Robust ConvNeXt-Based Framework for Efficient, Generalizable, and Explainable Brain Tumor Classification on MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Accurate brain tumor classification from MRI is crucial but challenging due to data variability and model generalization issues.
• Current deep learning models often lack robust validation and interpretability, limiting clinical reliability.
• Existing methods struggle with diverse tumor appearances and inter-dataset variability.

Purpose of the Study:

• To develop and evaluate a robust brain tumor classification framework using the ConvNeXt Base architecture.
• To assess the model's performance across multiple independent MRI datasets for glioma, meningioma, and pituitary tumors.
• To ensure the model's reliability, generalizability, and clinical applicability through rigorous validation and interpretability analysis.

Main Methods:

• Utilized the ConvNeXt Base architecture for brain tumor classification on three independent MRI datasets.
• Evaluated performance using comprehensive metrics (accuracy, AUC, F1-score, etc.) and statistical validation (Friedman test, Holm-Bonferroni).
• Assessed model interpretability using Grad-CAM++ and Gradient SHAP, alongside computational efficiency analysis.

Main Results:

• ConvNeXt Base achieved near-perfect classification performance (accuracy >99.6%, AUC ≈1.0) across all datasets.
• Statistical analyses confirmed significant and reproducible performance gains over other architectures.
• Explainability methods confirmed predictions are based on tumor-relevant regions, with favorable inference speed and resource usage.

Conclusions:

• ConvNeXt Base offers a reliable, generalizable, and explainable solution for MRI-based brain tumor classification.
• The model's high diagnostic accuracy and statistical robustness support its integration into clinical workflows.
• The framework's computational efficiency and interpretability enhance its potential for real-world clinical applications.