From Fractals to Efficient Net: A Hybrid Interpretable Model Outperforming CNN Baselines in Brain Tumor MRI Analysis

Abstract

Brain tumor classification using MRI is critical in neuro-oncology, but traditional Convolutional Neural Networks (CNNs) often lack clinical interpretability and fail to adequately capture complex morphological tumor patterns. This study introduces a novel, hybrid deep learning framework that combines geometric fractal analysis with an EfficientNet-B3 architecture to overcome these limitations. The proposed method extracts multi-scale fractal features—such as fractal dimension and lacunarity—to quantify spatial structural heterogeneity, fusing them with deep visual features from a transfer-learned EfficientNet-B3 model. Evaluated on a comprehensive dataset of 7,023 MRI images across four classes (glioma, meningioma, pituitary tumor, and no tumor), the hybrid model achieves a superior classification accuracy of 95.7%. This significantly outperforms baseline models, including standalone EfficientNet-B3 (94.5%), DenseNet-121 (93.8%), and ResNet-50 (92.1%). Furthermore, the model incorporates SHAP and Grad-CAM for enhanced explainability, validating that the network accurately targets clinically relevant tumor regions. The integration of fractal geometry with EfficientNet offers a highly accurate, computationally efficient, and interpretable solution for automated brain tumor diagnosis, successfully bridging the gap between AI performance and clinical trust.