Explainable Hybrid Deep Learning Framework Integrating MobileNetV2, EfficientNetV2B0, and KNN for MRI-Based Brain...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Magnetic resonance imaging (MRI) is crucial for noninvasive brain tumor assessment.
• Clinical adoption of AI in neuro-oncology requires both high accuracy and model transparency.

Purpose of the Study:

• To develop a lightweight and interpretable hybrid AI framework for brain tumor classification using MRI.
• To fuse features from MobileNetV2 and EfficientNetV2B0 using late fusion for enhanced diagnostic performance.
• To ensure clinical interpretability through visualization techniques like Grad-CAM and SHAP analysis.

Main Methods:

• A hybrid framework combining MobileNetV2 and EfficientNetV2B0 convolutional backbones with late fusion.
• Classification using a K-Nearest Neighbors (KNN) classifier (k=5, Euclidean distance, distance-based weighting).
• Dataset comprising 7,023 MRI images across four categories: Glioma, Meningioma, Pituitary, and Notumor, with a 64%/16%/20% train/validation/test split.

Main Results:

• Achieved 99.69% overall accuracy on the held-out test set.
• Class-wise ROC-AUC of 1.00 for all four diagnostic categories.
• High class-wise precision, recall, and F1 scores, supported by 5-fold cross-validation.

Conclusions:

• The dual-backbone late-fusion design with a KNN classifier provides strong, balanced performance for brain tumor classification.
• The framework offers clinically relevant interpretability via Grad-CAM and SHAP analyses.
• External validation is recommended to confirm the generalizability of the findings.