Tan Ou, Liu Keke, Chen Aiyin, Choi Dongseok, Chan Jonathan C H, Choy Bonnie N K, Shih Kendrick C, Wong Jasper K W, Ng Alex L K, Cheung Janice J C, Ni Michael Y, Lai Jimmy S M, Leung Gabriel M, Wong Ian Y H, Huang David
Casey Eye Institute, Oregon Health & Science University, Portland, Oregon.
Department of Ophthalmology, Duke University School of Medicine, Durham, North Carolina.
Ophthalmol Sci. 2025 Jun 9;5(6):100849. doi: 10.1016/j.xops.2025.100849. eCollection 2025 Nov-Dec.
Use generative deep learning (DL) models to estimate baseline reference nerve fiber layer thickness (NFLT) profiles, taking into account individual ocular characteristics.
A cross-sectional study.
Six hundred eighty-six individuals from the Hong Kong FAMILY cohort and 75 individuals from the Casey Eye Institute (CEI) cohort.
Healthy eyes were selected from the Hong Kong FAMILY and CEI cohorts. Circumpapillary NFLT profiles and vascular patterns were measured by a spectral-domain OCT. Generative DL models were trained using the FAMILY data to reconstruct the individualized baseline NFLT, a customized normal reference based on each eye's own vascular pattern, axial length (AL), spherical equivalent (SE) refractive error, disc size, and demographic information. Two DL models were developed. The MAG model used actual AL and SE, while the REG model estimated AL and SE using vascular patterns as input. For comparison, a multiple linear regression (MLR) was trained to estimate baseline NFLT using AL and demographic information. Fivefold cross-validation was used to assess performance.
The prediction error: root-mean-square of the difference between the actual NFLT profile and the predicted individualized baseline.
A total of 1152 healthy eyes from 686 participants in the Hong Kong Family cohort were divided into 4 subgroups: high myopia (SE <-6 diopters [D]), low myopia (SE = -6 D ∼ -1 D), emmetropia (SE = -1D∼1D), and hyperopia (SE >1D). Compared with the population means, both DL models significantly reduced the prediction error for overall and quadrant NFLT and decreased the false-positive rate of identifying abnormal NFLT thinning in both myopia groups (from 13.0%-27.0% to 6.3%∼9.4%). Both DL models significantly reduced prediction error for the NFLT profiles compared with both the population mean and the MLR-adjusted NFLT. The reductions in prediction errors for NFLT profile and overall NFLT value were independently validated using the CEI data.
Generative DL models (a type of artificial intelligence) can construct individualized NFLT baseline profiles using the vascular pattern derived from the same OCT scans. The individualized baseline reduced the prediction error of the NFLT profile in healthy eyes and may improve the accuracy of identifying abnormal NFLT thinning, especially in myopic eyes.
Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
利用生成式深度学习（DL）模型，在考虑个体眼部特征的情况下，估计基线参考神经纤维层厚度（NFLT）轮廓。
一项横断面研究。
来自香港家庭队列的686名个体以及来自凯西眼科研究所（CEI）队列的75名个体。
从香港家庭队列和CEI队列中选取健康眼睛。通过光谱域光学相干断层扫描（OCT）测量视乳头周围NFLT轮廓和血管模式。使用家庭队列数据训练生成式DL模型，以重建个性化的基线NFLT，即基于每只眼睛自身的血管模式、眼轴长度（AL）、等效球镜（SE）屈光不正、视盘大小和人口统计学信息定制的正常参考值。开发了两种DL模型。MAG模型使用实际的AL和SE，而REG模型使用血管模式作为输入来估计AL和SE。为作比较，训练了一个多元线性回归（MLR）模型，使用AL和人口统计学信息来估计基线NFLT。采用五折交叉验证来评估性能。
预测误差：实际NFLT轮廓与预测的个性化基线之间差异的均方根。
香港家庭队列中686名参与者的1152只健康眼睛被分为4个亚组：高度近视（SE<-6屈光度[D]）、低度近视（SE=-6 D～-1 D）、正视（SE=-1D～1D）和远视（SE>1D）。与总体均值相比，两种DL模型均显著降低了整体和象限NFLT的预测误差，并降低了两个近视组中识别异常NFLT变薄的假阳性率（从13.0%-27.0%降至6.3%～9.4%）。与总体均值和MLR调整后的NFLT相比，两种DL模型均显著降低了NFLT轮廓的预测误差。使用CEI数据独立验证了NFLT轮廓和总体NFLT值预测误差的降低情况。
生成式DL模型（一种人工智能）可以使用来自相同OCT扫描的血管模式构建个性化的NFLT基线轮廓。个性化基线降低了健康眼睛中NFLT轮廓的预测误差，并可能提高识别异常NFLT变薄的准确性，尤其是在近视眼中。
在本文末尾的脚注和披露中可能会发现专有或商业披露信息。
