Detail-aware semantic segmentation network for brain tumor MRI images combining multi-frequency directional filtering and lifting wavelets

Abstract

Brain tumors segmentation in Magnetic Resonance Imaging (MRI) images poses significant challenges owing to the uncertain location and size of the tumors, the difficulty in describing their boundaries, and the fuzzy demarcation of diseased tissues. Although U-Net and its recent variants have emerged as leading models for semantic segmentation in medical imaging, they still face structural limitations. These limitations cause the erosion of detail information during downsampling and poor performance in segmenting small lesions when handling targets of varying sizes, indicating a lack of detail handling capability. To counteract these issues, we designed a segmentation model that enhances detail features using frequency information. To reduce the loss of feature information during downsampling, we developed a downsampling module based on lifting wavelets. By lifting wavelets to group and integrate features according to frequency from high to low, we reduce feature resolution while enhancing information transmission and minimizing feature information loss. In our designed multi-frequency directional filtering edge feature extraction module, we extract low-frequency and high-frequency features and construct a dual-channel multi-directional filtering combination. This combination extracts directional information from low-frequency and high-frequency features separately, increasing the multi-angle directional information of the features and enriching the detailed information such as direction and position within the features. On the BraTS2018, BraTS2020, and BraTS2024 brain tumor datasets, our model demonstrated optimal results compared to 14 other advanced models. The average Dice Similarity Coefficients are 78.48 %, 79.80 %, and 74.35 %, while the 95th percentile Hausdorff Distances are 5.75, 6.60, and 7.72. Our code link is https://github.com/Eric-H8/BraTS_Seg_Model.