Scholarlink

<p>The appearance of stains in digital pathological images is prone to be affected by variations in imaging protocols, dyes, scanners, and illumination conditions. This inconsistency will prevent robustness and generalization of computer-aided diagnostic algorithms. Thus, many researchers have proposed efficient methods to normalize different stained images, among which the CycleGAN method prevails. In practice, we have found that this method can cause the background region to be mistaken for the foreground region or can result in the nucleus being considered cytoplasm, a phenomenon we refer to as Stain Region Inversion (SRI). To address the problem and improve its structure-preserving performance in stain normalization tasks, this paper proposes a novel stain normalization method called Structure-Preserving Self-Attention CycleGAN (SPSA-CycleGAN), which enhances the performance of CycleGAN in processing histological and cytological images. We demonstrate how to utilize multi-head self-attention to capture local features and use grey-scaled images to address the issue of SRI, enhancing the pixel-level structure-preserving capability of the original CycleGAN model. Our method is then verified in five experiments and compared with six other state-of-the-art stain normalization methods. The experimental results demonstrated that our SPSA-CycleGAN has better or comparable performance compared to all the other methods. Code available at: <a href="https://github.com/Smile-We/SPSA-CycleGAN">https://github.com/Smile-We/SPSA-CycleGAN</a></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

Stain normalization of pathological images using self-attention based cycleGAN with grey scale consistency loss

  • 期刊论文

  • 发表日期：

    2025年8月20日

查看全文
  • Zheng Chen,
  • Peng Jiang,
  • Wensi Duan,
  • Lang Wang,
  • Cheng Li,
  • Junfeng Wang,
  • Juan Liu

摘要

The appearance of stains in digital pathological images is prone to be affected by variations in imaging protocols, dyes, scanners, and illumination conditions. This inconsistency will prevent robustness and generalization of computer-aided diagnostic algorithms. Thus, many researchers have proposed efficient methods to normalize different stained images, among which the CycleGAN method prevails. In practice, we have found that this method can cause the background region to be mistaken for the foreground region or can result in the nucleus being considered cytoplasm, a phenomenon we refer to as Stain Region Inversion (SRI). To address the problem and improve its structure-preserving performance in stain normalization tasks, this paper proposes a novel stain normalization method called Structure-Preserving Self-Attention CycleGAN (SPSA-CycleGAN), which enhances the performance of CycleGAN in processing histological and cytological images. We demonstrate how to utilize multi-head self-attention to capture local features and use grey-scaled images to address the issue of SRI, enhancing the pixel-level structure-preserving capability of the original CycleGAN model. Our method is then verified in five experiments and compared with six other state-of-the-art stain normalization methods. The experimental results demonstrated that our SPSA-CycleGAN has better or comparable performance compared to all the other methods. Code available at: https://github.com/Smile-We/SPSA-CycleGAN