Brain Tumor MRI Segmentation Method Based on Improved Res-UNet

Automatic segmentation of MRI images is crucial for diagnosis and evaluation of brain tumors. However, significant variability in brain tumor shape, uneven spatial distribution, and intricate boundaries bring challenges, which lead information loss and decreased accuracy during segmentation. To solve these problems, an improved Res-UNet network employing attention-guided and scale-aware strategies is proposed. First, a module that employs attention mechanisms and features fusion is incorporated to catch relatively important contextual information. Secondly, a module designed to retrieve hidden contextual information and dynamically aggregate multi-scale features is integrated into the bottom layer of the network, which facilitates feature acquisition and enhancement at multiple scales. Finally, the results show that the method achieves a Dice similarity coefficient of 92.24% in whole tumor region, which is an improvement of about 4% compared to the pre-improved Res-UNet network.