Multi-Branch Attention-Guided Network for Accurate Brain Tumor Segmentation from Multi-Modal MRI

Aiming at the problem of low segmentation accuracy due to complex shapes and blurred boundaries in MRI brain tumor segmentation, a multi-scale convolutional neural network algorithm model (3DMDE-UNet) is proposed. In the encoding stage, an enhanced multi-branch attention module is constructed to integrate channel and spatial dual mechanisms. The channel attention is integrated into the hierarchical scaling of multi-scale factors to construct the MBCA module; the spatial attention is combined with diversified receptive fields to construct the MBSA module to improve the adaptability of the model to complex structural changes; at the same time, the multiscale module is used to replace the single-branch convolution structure in downsampling to enhance the network's ability to capture multi-level features. In the decoding stage, a dual convolution module is introduced to improve the network's ability to handle fine boundaries. Experiments show that compared with 3D-UNet, the improved model verifies the effectiveness of the model on the BraTS 2019 and BraTS 2021 public datasets. The Dice coefficient, Hausdorff distance and accuracy of the tumor core, enhanced tumor and full tumor area have been effectively improved, which confirms that 3DMDE-UNet can effectively improve the segmentation accuracy of MRI brain tumors.