Swinnetplus: A Novel Deep Learning Approach Using Swin Transformer For Enhancing Brain Tumor Segmentation
TL;DR
The suggested SwinNet Plus model, the new hybrid architecture of deep learning that combines the complementary advantages of Swin Transformers and Convolutional Neural Networks, is suggested to strengthen accurate brain tumor localization and becomes an attractive instrument in clinical determination support systems.
Abstract
Abstract The segmentation of brain tumors based on three-dimensional (3D) magnetic resonance imaging (MRI) is a basic but difficult challenge when analyzing medical images due to the large differences in size, shapes, and appearances of tumors, and the high visual similarity of tumor with the normal tissues. To overcome such difficulties, the paper suggests SwinNet Plus, the new hybrid architecture of deep learning that combines the complementary advantages of Swin Transformers and Convolutional Neural Networks (CNNs) to strengthen accurate brain tumor localization. The SwinNetPlus is made to capture both large-scale contextual features and small-scale spatial features with an encoder-decoder architecture. The encoder uses the Swin Transformer based on Enhanced Local Self-Attention (ELSA) to model long-range contextual features and maintain local structural features. In order to further improve feature discriminability, the Swin Transformer adds a channel squeeze-and-excitation block to an adaptive recalibration channel-blockwise responses to multimodal MRI input. Moreover, two special modules are provided to enhance the accuracy in segmentation. The Spatial Local Attention Module (SLAM) has a focus on small local details and tumor edges, which allows precise definition small and irregular tumor areas. The Dense Cross-Multiplication Link is an enhancement of a cross-multiplication link that encourages consistent multiscale fusion, implying that encoder features and decoder features are densely integrated to eliminate irrelevant activation and noise transmission. The suggested SwinNetPlus model is highly tested on BraTS 2021 dataset, with 1,251 multimodal 3D brain MRI scans. The experimental outcomes also indicate that SwinNetPlus has a Dice similarity coefficient of 92.69 percent and Hausdorff distance of 5.34 mm which is superior to the performance of some current methods of 3D brain tumor segmentation. These findings underscore the strength and efficacy of SwinNetPlus in the correct segmentation of brain tumors under magneticity tough circumstances, thus it becomes an attractive instrument in clinical determination support systems.