Scholarlink

We present a novel architecture, SwinFuseNet, to tackle the challenges of precisely identifying brain tumours from medical imaging. Our methodology, which combines convolutional neural networks (CNNs) with Swin Transformers, efficiently addresses differences in lesion sizes, kinds, and minor colour and shape resemblances to surrounding tissue. The Dense Cross-Multiplication Link (DCML) module is one of the key components of SwinFuseNet. It integrates semantic feature information across many scales in a stepwise manner using a Dense Cross-Multiplication Link approach. This improves feature expression and reduces the influence of background noise. Furthermore, in the encoder-decoder architecture, we extract local detailed features using ELSA transformer blocks. Channel squeeze and spatial excitation blocks are added to the encoder to further strengthen it and extract more relevant spatial and channel-wise feature representations. We test SwinFuseNet on 1251 brain pictures from the BraTS 2021 datasets, and with an average Hausdorff distance of 3.20 mm and Dice score of 89.75%, we obtain impressive segmentation results. By combining the capabilities of Swin Transformer with improved local self-attention and feature fusion, SwinFuseNet significantly improves brain tumour segmentation from MRI data, beyond the capabilities of current 3D techniques. Our automated methodology offers a dependable means of tackling issues associated with lesion heterogeneity and minute visual fluctuations in medical imaging.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Precision Brain Tumour Segmentation in MRI Using SwinFuseNet: A Deep Learning Approach Integrating CNNs, Swin Transformers, and DCML

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Vikash Verma,
  • Pritaj Yadav

摘要

We present a novel architecture, SwinFuseNet, to tackle the challenges of precisely identifying brain tumours from medical imaging. Our methodology, which combines convolutional neural networks (CNNs) with Swin Transformers, efficiently addresses differences in lesion sizes, kinds, and minor colour and shape resemblances to surrounding tissue. The Dense Cross-Multiplication Link (DCML) module is one of the key components of SwinFuseNet. It integrates semantic feature information across many scales in a stepwise manner using a Dense Cross-Multiplication Link approach. This improves feature expression and reduces the influence of background noise. Furthermore, in the encoder-decoder architecture, we extract local detailed features using ELSA transformer blocks. Channel squeeze and spatial excitation blocks are added to the encoder to further strengthen it and extract more relevant spatial and channel-wise feature representations. We test SwinFuseNet on 1251 brain pictures from the BraTS 2021 datasets, and with an average Hausdorff distance of 3.20 mm and Dice score of 89.75%, we obtain impressive segmentation results. By combining the capabilities of Swin Transformer with improved local self-attention and feature fusion, SwinFuseNet significantly improves brain tumour segmentation from MRI data, beyond the capabilities of current 3D techniques. Our automated methodology offers a dependable means of tackling issues associated with lesion heterogeneity and minute visual fluctuations in medical imaging.