Scholarlink

Purpose <p>To compare artificial intelligence (AI)-based annotations of hyperreflective material (HRM) and manual demarcation of macular neovascularization (MNV) on optical coherence tomography (OCT) volume scans in neovascular age-related macular degeneration (nAMD), and to assess the suitability of AI-driven OCT segmentation for longitudinal lesion monitoring.</p> Methods <p>In this retrospective study, 42 eyes from 36 patients (21 f, 15&#xa0;m; mean age baseline 76.6 y) with exudative nAMD were analyzed using longitudinal spectral-domain OCT data. Manual MNV demarcations on <i>en-face</i> OCT projections served as ground truth and were compared to AI-predicted HRM segmentations generated by a 3D nU-Net model on OCT scans. HRM and MNV lesion areas were quantified at multiple time points, and agreement between manual and AI-based measurements was evaluated using Pearson correlation, ordinary least squares regression and robust regression.</p> Results <p>A highly similar mean lesion growth was observed when comparing HRM/MNV lesion sizes in longitudinal measurements. Point-by-point comparison revealed a strong overall correlation (r = 0.78) between AI-predicted and manually annotated HRM areas with increasing significance with longer follow-up. However, two aspects were responsible for some AI measurements being larger than manual measurements: At baseline, AI measurements included hyperreflective subretinal fluid as HRM, which was resorbed after three anti-VEGF injections, and during longer-term follow-up, manually annotated MNV areas were occasionally smaller than those derived from AI-based HRM segmentation due to the manual underestimation of very thin HRM.</p> Conclusions <p>AI-based segmentation of HRM on OCT scans demonstrates strong overall agreement with manual MNV measurements, particularly on longitudinal assessments. Despite some AI-based overestimations occurring at baseline and some manual MNV underestimations during follow-up, measurements between both methods were highly comparable over 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

Automated measurement of macular neovascularization lesion size in nAMD using AI segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月17日

查看全文
  • Anna Vahldiek,
  • Lukas Heine,
  • Benja Vahldiek,
  • Jasper Schröter,
  • Jan-Niklas Wolf,
  • Michael Swora,
  • Lars Reissberg,
  • Laurenz Pauleikhoff,
  • Jens Kleesiek,
  • Daniel Pauleikhoff

摘要

Purpose

To compare artificial intelligence (AI)-based annotations of hyperreflective material (HRM) and manual demarcation of macular neovascularization (MNV) on optical coherence tomography (OCT) volume scans in neovascular age-related macular degeneration (nAMD), and to assess the suitability of AI-driven OCT segmentation for longitudinal lesion monitoring.

Methods

In this retrospective study, 42 eyes from 36 patients (21 f, 15 m; mean age baseline 76.6 y) with exudative nAMD were analyzed using longitudinal spectral-domain OCT data. Manual MNV demarcations on en-face OCT projections served as ground truth and were compared to AI-predicted HRM segmentations generated by a 3D nU-Net model on OCT scans. HRM and MNV lesion areas were quantified at multiple time points, and agreement between manual and AI-based measurements was evaluated using Pearson correlation, ordinary least squares regression and robust regression.

Results

A highly similar mean lesion growth was observed when comparing HRM/MNV lesion sizes in longitudinal measurements. Point-by-point comparison revealed a strong overall correlation (r = 0.78) between AI-predicted and manually annotated HRM areas with increasing significance with longer follow-up. However, two aspects were responsible for some AI measurements being larger than manual measurements: At baseline, AI measurements included hyperreflective subretinal fluid as HRM, which was resorbed after three anti-VEGF injections, and during longer-term follow-up, manually annotated MNV areas were occasionally smaller than those derived from AI-based HRM segmentation due to the manual underestimation of very thin HRM.

Conclusions

AI-based segmentation of HRM on OCT scans demonstrates strong overall agreement with manual MNV measurements, particularly on longitudinal assessments. Despite some AI-based overestimations occurring at baseline and some manual MNV underestimations during follow-up, measurements between both methods were highly comparable over time.