文献速递眼底病与葡萄膜炎

高度近视中国人扫频光源OCT设备脉络膜厚度测量比较

Comparison between choroidal thickness measurements from swept-source optical coherence tomography devices in highly myopic Chinese adults.
PubMed 文献 · Br J Ophthalmol · 2026-09-11

本研究纳入97名高度近视中国成人,分别用DREAM OCT和Triton OCT成像,采用厂商软件和深度学习模型测量ETDRS网格脉络膜厚度。厂商软件下两设备全局厚度差异达61.04-71.22 µm(p<0.001),ICC仅0.63;深度学习模型下全局差异仅2.26 µm,ICC达0.98,设备间一致性显著更优。

要点

  • 厂商软件下DREAM OCT测得的脉络膜厚度显著厚于Triton OCT,全局差异61.04-71.22 µm。
  • 深度学习模型将设备间全局厚度平均差异缩小至2.26 µm,各ETDRS分区ICC均>0.90。
  • 深度学习模型设备间全局ICC为0.98,显著优于厂商软件的0.63(p<0.05)。

核心结论

  • 厂商自带软件在DREAM OCT与Triton OCT间产生系统性脉络膜厚度测量差异,深度学习模型显著降低设备间变异性。

研究设计与样本

  • 本研究为设备比较研究,纳入97名高度近视中国成人,分别使用DREAM OCT和Triton OCT成像,采用厂商软件和深度学习分割模型测量ETDRS网格脉络膜厚度并进行眼放大校正。

主要发现

  • 厂商软件下DREAM OCT全局ETDRS脉络膜厚度为221.48±66.88 µm,Triton OCT为156.90±57.64 µm,平均差异61.04-71.22 µm(均p<0.001),全局ICC=0.63(95%CI 0.56-0.68)。
  • 深度学习模型下DREAM OCT全局厚度为173.47±59.54 µm,Triton OCT为171.21±56.62 µm,全局平均差异2.26 µm,一致性界限-17.38至21.90 µm。
  • 深度学习模型各ETDRS分区ICC均>0.90,全局ICC=0.98,显著优于厂商软件的0.63(p<0.05)。

关键数值与推荐

  • 厂商软件下两设备全局ETDRS脉络膜厚度平均差异为61.04-71.22 µm(均p<0.001),全局ICC=0.63(95%CI 0.56-0.68)。
  • 深度学习模型下全局平均差异2.26 µm,一致性界限-17.38至21.90 µm,各ETDRS分区ICC>0.90,全局ICC=0.98。

临床怎么用

  • 在高度近视人群中比较不同扫频光源OCT平台的脉络膜厚度时,厂商软件测量值存在系统性偏移,不宜直接互换。
  • 采用深度学习分割模型可显著提高跨设备脉络膜厚度测量的一致性,可能更适合多平台随访比较。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

BACKGROUND/AIMS: Reproducibility of choroidal thickness (CT) measurements is imperative when comparing choroidal imaging from different optical coherence tomography (OCT) platforms particularly in high myopia. This study compared CT measurements from commercial swept-source OCT devices in highly myopic Chinese adults using vendor-provided software and a deep learning segmentation model. METHODS: 97 participants underwent imaging with DREAM OCT and Triton OCT. CT was measured across the Early Treatment Diabetic Retinopathy Study (ETDRS) grid using vendor-provided software and a deep learning model, with ocular magnification correction. Agreement was evaluated using mean differences with 95% CIs, intraclass correlation coefficients (ICCs) and Bland-Altman plots. RESULTS: With vendor-provided software, DREAM OCT yielded thicker CT than Triton OCT (global ETDRS: 221.48±66.88 µm vs 156.90±57.64 µm), with mean differences of 61.04-71.22 µm (all p<0.001). Global agreement was moderate (ICC=0.63; 95% CI 0.56 to 0.68). Deep learning-derived CT measurements showed much smaller inter-device differences (DREAM: 173.47±59.54 µm; Triton: 171.21±56.62 µm), yielding a mean global difference of 2.26 µm with limits of agreement from -17.38 to 21.90 µm. Sectoral CT differences were similarly minimised (ICCs>0.90 across all ETDRS regions). The deep learning model demonstrated significantly superior inter-device agreement (global ICC=0.98 vs 0.63; p<0.05). CONCLUSIONS: Vendor-provided software produced systematic differences between DREAM OCT and Triton OCT, whereas the deep learning model produced CT measurements with significantly reduced variability. When evaluating CT measurements from different platforms, deep learning-based approaches may provide more comparable results than vendor-provided software.

来源:Br J Ophthalmol,查看原文。版权归原作者所有。

引用

  1. Sreng S, Tan H, Cheong KX, et al. Comparison between choroidal thickness measurements from swept-source optical coherence tomography devices in highly myopic Chinese adults. Br J Ophthalmol. 2026. PMID: 42728039. doi:10.1136/bjo-2026-329740

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。