Scholarlink

Retinal procedures such as epiretinal membrane (ERM) peeling and macular hole repair demand precise instrument positioning to avoid contact with the retina, as even the smallest forces can cause sight impairment. Intraoperative Optical Coherence Tomography (iOCT) delivers depth details and real-time views of tissue and tool interactions. Accurate segmentation of iOCT images would permit estimation of the distance between surgical instruments and retinal layers, and the incorporation of safety alerts. Although pre-operative OCT (pOCT) segmentation is tractable, iOCT segmentation presents challenges due to the lack of annotation, the presence of instruments, shadows, tissue variations, and low signal-to-noise ratio (SNR). To overcome the problems, we pioneered the application of few-shot segmentation (FSS) frameworks that transfer information from well-annotated pOCT data to iOCT segmentation with only a handful of labeled examples. We discussed four FSS methods and evaluated them on three iOCT datasets collected from different surgical scenarios on real patients.

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

Cross Domain Few Shot Learning for Intra-Operative OCT Segmentation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Minghan Zhao,
  • Gongyu Zhang,
  • Adriana Namour,
  • Charalampos Komninos,
  • Lyndon da Cruz,
  • Sebastien Ourselin,
  • Christos Bergeles

摘要

Retinal procedures such as epiretinal membrane (ERM) peeling and macular hole repair demand precise instrument positioning to avoid contact with the retina, as even the smallest forces can cause sight impairment. Intraoperative Optical Coherence Tomography (iOCT) delivers depth details and real-time views of tissue and tool interactions. Accurate segmentation of iOCT images would permit estimation of the distance between surgical instruments and retinal layers, and the incorporation of safety alerts. Although pre-operative OCT (pOCT) segmentation is tractable, iOCT segmentation presents challenges due to the lack of annotation, the presence of instruments, shadows, tissue variations, and low signal-to-noise ratio (SNR). To overcome the problems, we pioneered the application of few-shot segmentation (FSS) frameworks that transfer information from well-annotated pOCT data to iOCT segmentation with only a handful of labeled examples. We discussed four FSS methods and evaluated them on three iOCT datasets collected from different surgical scenarios on real patients.