Scholarlink

<p>Optical coherence tomography angiography (OCTA) technology can accurately depict the microvascular architecture of the retina, providing crucial evidence for the diagnosis of retinal disease. Nevertheless, OCTA images are typically accompanied by artifacts and low signal-to-noise ratios, exerting a severe impact on diagnostic efficiency. Deep learning-based segmentation algorithms are widely recognized for improving retinopathy diagnosis accuracy. This research investigates the application of the segment anything model (SAM) in OCTA retinal vessel segmentation and proposes a self-prompting SAM based on wavelets and spatial domains, named WS-SAM. Specifically, we innovatively designed three key modules: (1) A dual-domain encoder aims to extract multi-scale features through a joint encoding method in the spatial domain and frequency domain to effectively suppress noise interference. (2) A wavelet space fusion module aims to effectively suppress artifact interference and enhance detail texture features by adaptively fusing multi-scale frequency-domain features with spatial-domain features. (3) A Meta Self-Prompter is designed to automatically generate prompt information based on prototype learning algorithms and guide the model to focus on vascular structures via the prompt mechanism, thereby preventing segmentation fractures. Furthermore, we have established a novel dataset, namely OCTA-RV, to augment the data in the field of OCTA retinal vessel segmentation. The experimental results indicate that the Dice coefficients of the WS-SAM model on the datasets of OCTA500-3&#xa0;M, OCTA500-6&#xa0;M and OCTA-RV are 0.8754, 0.8949 and 0.7412, respectively, manifesting a remarkable competitiveness compared with contrastive models.</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

WS-SAM: self-prompting SAM with wavelet and spatial domain for OCTA retinal vessel segmentation

  • 期刊论文

  • 发表日期：

    2025年3月28日

查看全文
  • Zhaohui Zhang,
  • Fei Ma,
  • Hongjuan Liu,
  • Xiwei Dong,
  • Yanfei Guo,
  • Jing Meng

摘要

Optical coherence tomography angiography (OCTA) technology can accurately depict the microvascular architecture of the retina, providing crucial evidence for the diagnosis of retinal disease. Nevertheless, OCTA images are typically accompanied by artifacts and low signal-to-noise ratios, exerting a severe impact on diagnostic efficiency. Deep learning-based segmentation algorithms are widely recognized for improving retinopathy diagnosis accuracy. This research investigates the application of the segment anything model (SAM) in OCTA retinal vessel segmentation and proposes a self-prompting SAM based on wavelets and spatial domains, named WS-SAM. Specifically, we innovatively designed three key modules: (1) A dual-domain encoder aims to extract multi-scale features through a joint encoding method in the spatial domain and frequency domain to effectively suppress noise interference. (2) A wavelet space fusion module aims to effectively suppress artifact interference and enhance detail texture features by adaptively fusing multi-scale frequency-domain features with spatial-domain features. (3) A Meta Self-Prompter is designed to automatically generate prompt information based on prototype learning algorithms and guide the model to focus on vascular structures via the prompt mechanism, thereby preventing segmentation fractures. Furthermore, we have established a novel dataset, namely OCTA-RV, to augment the data in the field of OCTA retinal vessel segmentation. The experimental results indicate that the Dice coefficients of the WS-SAM model on the datasets of OCTA500-3 M, OCTA500-6 M and OCTA-RV are 0.8754, 0.8949 and 0.7412, respectively, manifesting a remarkable competitiveness compared with contrastive models.