Accurate Brain Tumor Segmentation Based on an Improved U-Net with Feature Aggregation

Recently, image segmentation methods with high accuracy have become increasingly important in the diagnosis and treatment of brain tumors. Since segmentation models based on the U-Net architecture often have limited sizes for the sensing fields of image features, the contextual information thus cannot be fully used for segmentation. This paper proposes a new model that can effectively fuse multi-scale image features for the segmentation of brain tumor images. Firstly, an improved convolution module is utilized to replace the original convolutional component for the extraction of image features, which increases the depth and width of the network, and enhances its feature learning ability in complex tasks. Secondly, several improved downsampling and upsampling modules that have an inverted bottleneck structure with residuals are introduced in the sampling process, which maintains the semantic richness across various scales. Finally, a feature aggregation attention module is included in the bottom of the U-shape, which enables the network to pay more attention to effective regions and suppresses the noise. Experimental results show that the proposed model achieves excellent segmentation performance on the brain tumor dataset BraTS 2021, and the Dice coefficients of whole tumor, tumor core, and enhancing tumor are 0.923, 0.890, and 0.861 respectively, which are improved by 1.6%, 1.3%, and 2.2% compared to the base network of the U-Net architecture. The proposed model can achieve an average value of 0.891 for the Dice coefficients and an average value of 5.172 for the Hausdorff distances. The results of comparison suggest that the proposed approach can outperform a few other state-of-the-art approaches for brain tumor segmentation.