An explainable multi-stage framework for brain tumor classification using hybrid feature fusion and EfficientNetB5 model

Accurate recognition and classification of severity level of brain tumor (BT) is essential for clinical decision making. Manual assessment of brain tumor is a laborious and time consuming task for Experts. Existing deep learning (DL) models have been shown promising classification accuracy with limited features diversity. This study presents a HyFusion-Net model, a novel and explainable DL framework, fuses handcrafted features (SIFT, LBP, HoG, Canny) with deep features derived from DenseNet201, achieving superior accuracy with an EfficientNetB5 classifier. For clinical interpretation, an occlusion map-based explainability is integrated to highlight tumor regions driving classification. In addition, the proposed HyFusion-Net model provides robust performance against MRI variability and ensure a transparent diagnostic rate. Evaluated on a total number of 7023 MRI images using various statistical metrics (AUC, precision, F1-score, and accuracy), the proposed HyFusion-Net outperforms other DL methods while providing heatmaps aligned with radiological markers. Experimental results endorse the suitability of the proposed HyFusion-Net model for clinical use and proves the ability of the model to helps radiologists to understand the classification decision.