10th International Congress on Information and Communication Technology

Authors - Shahd Tarek, Ali Hamdi
Abstract - Brain tumors represent one of the most critical health challenges due to their complexity and high mortality rates, necessitating early and precise diagnosis to improve patient outcomes. Traditional MRI interpretation methods rely heavily on manual analysis, which is timeconsuming, error-prone, and inconsistent. To address these limitations, this study introduces a novel deep attentional framework that integrates multiple Convolutional Neural Network (CNN) base models—EfficientNet- B0, ResNet50, and VGG16—within a Multi-Head Attention (MHA) mechanism for robust brain tumor classification. Convolutional features extracted from these CNNs are fed into the MHA as Query (Q), Key (K), and Value (V) inputs, enabling the model to focus on the most distinguishing features within MRI images. By leveraging complementary feature maps from diverse CNN architectures, the MHA mechanism generates more refined, attentive representations, significantly improving classification accuracy. The proposed approach classifies MRI images into four categories: pituitary tumor, meningioma, glioma, and no tumor. A dataset of 7,023 labeled MRI images was curated from public repositories, including Figshare, SARTAJ, and Br35H, with preprocessing steps to standardize dimensions and remove margins. Experimental results demonstrate the superior performance of individual CNNs—VGG16 achieving 97.25% accuracy, ResNet50 98.02%, and EfficientNet-B0 93.21%. Moreover, the ensemble model integrating VGG16, EfficientNet-B0, and ResNet50 achieves the highest accuracy of 98.70%, surpassing other ensemble configurations such as ResNet50 + VGG16 + EfficientNet-B0 (96.95%) and VGG16 + ResNet50 + EfficientNet-B0 (95.96%). These findings underscore the effectiveness of multi-level attention in refining predictions and provide a reliable, automated tool to assist radiologists. The proposed framework highlights the transformative potential of deep learning in medical imaging, streamlining clinical workflows, and enhancing healthcare outcomes.