Grad-CAM++ Enhanced Hybrid CNN-Random Forest Model for Accurate and Transparent Brain Tumor Detection

Brain tumor detection from MRI scans significantly benefits from advanced deep learning techniques; however, conventional Convolutional Neural Networks (CNNs) often lack interpretability, limiting clinical trust. This research proposes an explainable hybrid framework integrating CNNbased feature extraction using ResNet50 with Random Forest (RF) classification, augmented by Grad-CAM++ for enhanced transparency. Evaluations performed on the BraTS 2020 dataset (2,500 MRI images categorized into glioma, meningioma, pituitary tumors, and healthy tissues) demonstrated the superior performance of the CNN-RF hybrid model, achieving an accuracy of 97.2%, precision of 96.8%, recall of 96.4%, and $F 1$-score of 96.6%. These results significantly outperformed traditional models such as VGG16 (accuracy: 92.4%), EfficientNet-B0 (accuracy: 94.5%), and CNN-only ResNet50 (accuracy: 95.8%). Additionally, ROCAUC analysis validated exceptional discriminatory power with an AUC score of 98.5%, highlighting the robust diagnostic potential of the hybrid model. The incorporation of GradCAM++ visualization markedly improved interpretability by effectively identifying clinically relevant tumor features, closely aligning with expert radiologist observations. Confusion matrix and training-validation accuracy analyses further confirmed stable model generalization and minimal misclassification. Overall, the developed CNN-RF framework combines high diagnostic accuracy with critical interpretability, promoting clinical acceptance and offering substantial potential for enhancing diagnostic decision-making in medical imaging practice.