Scholarlink

<p>Leveraging the progress made in computer vision, the field of medical imaging has seen increasing interest in the use of transformers. Vision transformers have demonstrated superior feature representation capabilities compared to Convolutional Neural Networks (CNNs). In this paper, using similar intuition to DenseNets, we propose the Swin-DWA architecture. Our method introduces an additional averaging phase, termed Depth-Weighted Average (DWA), which computes a weighted average of the previous and current representations after each Window Multi-Head Self-Attention (W-MSA) and Shifted-Window Multi-Head Self-Attention (SW-MSA) operation. The learned DWA weights exhibit well-structured and resilient information flow patterns, suggesting the effective reuse of activations from distant layers. The model was evaluated using the AAPM-Mayo Clinic LDCT Grand Challenge Dataset and demonstrated superior performance compared to previous state-of-the-art (SOTA) models across various architectures, achieving a PSNR of 33.1883, an SSIM of 0.9148, and an RMSE of 8.9502. In terms of both visual quality and quantitative performance, our method outperforms sophisticated LDCT denoising techniques while also being superior in inference time.</p>

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

Swin-DWA: Enhancing Information Flow in Swin Transformer-Based Low-Dose CT Image Denoising via Depth-Weighted Averaging

  • 期刊论文

  • 发表日期：

    2026年2月5日

查看全文
  • Abdelkarim Cherhabil,
  • Lahcène Mitiche,
  • Amel Baha Houda Adamou-Mitiche

摘要

Leveraging the progress made in computer vision, the field of medical imaging has seen increasing interest in the use of transformers. Vision transformers have demonstrated superior feature representation capabilities compared to Convolutional Neural Networks (CNNs). In this paper, using similar intuition to DenseNets, we propose the Swin-DWA architecture. Our method introduces an additional averaging phase, termed Depth-Weighted Average (DWA), which computes a weighted average of the previous and current representations after each Window Multi-Head Self-Attention (W-MSA) and Shifted-Window Multi-Head Self-Attention (SW-MSA) operation. The learned DWA weights exhibit well-structured and resilient information flow patterns, suggesting the effective reuse of activations from distant layers. The model was evaluated using the AAPM-Mayo Clinic LDCT Grand Challenge Dataset and demonstrated superior performance compared to previous state-of-the-art (SOTA) models across various architectures, achieving a PSNR of 33.1883, an SSIM of 0.9148, and an RMSE of 8.9502. In terms of both visual quality and quantitative performance, our method outperforms sophisticated LDCT denoising techniques while also being superior in inference time.