SWATT: Semi-Wavelet Attention Module for Enhancing Brain Tumor Image Segmentation

Abstract

Detecting brain tumors and their boundaries is crucial for timely diagnosis and reducing patient mortality. Specialists determine these borders, but tumor swelling, tissue differences, and variability in shape and size across patients often result in irregular and indeterminate tumor borders. In recent years, the encoder-decoder framework, encompassing renowned architectures like U-Net, has demonstrated favorable outcomes in medical image segmentation tasks to overcome these obstacles. However, enhancing the quality of tumor segmentation and delineating precise tumor boundaries are immense challenges for effective diagnosis and treatment planning. To address these challenges, we propose a new model called SWATT U-Net that utilizes a Semi-Wavelet Attention (SWATT) module within the U-Net architecture. SWATT incorporates a semi-wavelet layer and its inverse, enabling multi-resolution analysis of input images by scaling the original image to focus on specific parts. The SWATT U-Net model is trained on the BraTS2020 dataset. The results demonstrate an improvement, with the Mean Intersection over Union (mean-IoU) increasing from 82.82% to 87.16%. Moreover, SWATT was applied to PSPNET, ResNet, SegNet, and DeepLabV3 models, resulting in significant mean-IoU enhancements ranging from 14.53% to 20.88% compared to those without SWATT. Additionally, the statistical analysis using t-tests confirms the superiority of the modified models incorporating the SWATT attention mechanism, leading to improved segmentation results. Finally, the SWATT U-Net model's results were compared with those of various new methods for brain tumor segmentation on the BraTS2020 dataset. It was found that the recommended model outperformed the competing models in the mean-IoU metric. © 2025