SwinNetPlus: A Novel Deep Learning Approach Using Swin Transformer for Enhancing Brain Tumor Segmentation

Abstract

SwinNetPlus is a new architecture that uses Swin Transformers and Convolutional Neural Networks (CNNs) for ccurate segmentation of brain tumors in 3D MRI scans. This model deals with the problem of lesion sizes changing and with the fact that tumor and healthy tissue may look similar to the naked eye. It introduces two main building blocks: the Spatial Local Attention Module (SLAM) which pays close attention to small features and the Dense Cross-Multiplication Link (DCML) which helps keep the meaning of the features consistent by blending them together and eliminating distractions. The encoder relies on ELSA transformer blocks to pull out specific details from the images and the channel squeeze-and-excitation blocks enhance how features are represented across channels. To combine context from the whole image with spatial information, SwinNetPlus uses an encoder-decoder structure. On the BraTS 2021 dataset of 1,251 multimodal brain MRI scans, it scored 92.69% Dice and 5.34 mm Hausdorff distance which is better than many other 3D segmentation methods. These findings prove that SwinNetPlus is able to outline tumors accurately when images are not ideal.