DMA-Net: A Deep Multi-Attention Network for Brain Tumor Classification on MRI Images

Abstract:
Accurate and impartial classification of brain tumors is crucial for early diagnosis and treatment. Deep learning technology can effectively represent and utilize the visual content of images, playing a key role in more precise medical image analysis. In this study, we proposed an advanced deep multi-attention network (DMA-Net) to address the challenge of brain tumor classification on MRI images. The proposed DMA-Net employs a novel multi-layer attention mechanism, integrating efficient channel attention (ECA), spatial attention (SPA), depthwise separable convolutions with vision transformer (DWSC-ViT), and group attention (GA) to enhance the processing of complex neural imaging. Our approach is characterized by its ability to effectively balance detailed image representation and computational efficiency, maintaining excellent classification accuracy even with a substantial reduction in model parameters. With the integration of these new modules, DMA-Net outperformed other state-of-the-art (SOTA) mobile networks, achieving a classification accuracy of 96.70% and an AUC of 0.9971 in brain tumor classification. Furthermore, DMA-Net is the most compact model among the SOTA networks, demonstrating significant potential for deployment on mobile devices to enhance application efficiency.