Multiclass Tumor Segmentation From Brain MRIs Using GARU-Net: Gelu Activated Attention Aware Res-3DUNET for Adaptive Feature Pooling

The process of segmenting brain tumor images is of paramount importance in the supplementary diagnosis of diseases, devising treatment plans, and aiding in surgical navigation. To achieve precise segmentation of brain tumor images, this presents a comprehensive structure for automating the segmentation of 3-D brain tumors. The model proposed combines the deep residual network and U-Net model with attention guidance and is referred to as GARU-Net. The residual network is used as an encoder to solve the problem of vanishing gradient, and the decoder of the U-Net model is employed in the proposed architecture. Additionally, the U-Net decoder side is amplified with an attention mechanism that de-emphasizes healthy tissues and highlights malignant tissues, leading to improved generalization and reduced computational resources. The proposed architecture has demonstrated excellent results, with an average dice score of 0.860, 0.908, and 0.824 for the tumor core, whole tumor, and enhancing tumor, respectively, on the BraTS 2020 dataset. The research suggests that the proposed approach could improve brain tumor segmentation using multimodal MRI data, contributing to a better understanding and diagnosis of brain diseases.