Xu Zhiqiang, Liu Anran, Su Binbin, Wu Minhui, Zhang Bin, Chen Guanyan, Lu Fan, Hu Liang, Mao Xinjie
National Clinical Research Center for Ocular Diseases, Eye Hospital, Wenzhou Medical University, Wenzhou, Zhejiang, People's Republic of China.
National Engineering Research Center of Ophthalmology and Optometry, Eye Hospital, Wenzhou Medical University, Wenzhou, People's Republic of China.
Transl Vis Sci Technol. 2025 Aug 1;14(8):16. doi: 10.1167/tvst.14.8.16.
The purpose of this study was to develop a standardized artificial intelligence (AI) system integrating corneal topography images and numerical parameters for optimizing orthokeratology (OK) lens fitting.
Retrospective data from 1153 patients (2341 eyes) with Euclid OK lenses were analyzed. Five hundred nineteen eyes (393 patients) with treatment zone decentration ≤1 mm were included for model training. A device-agnostic corneal topography reconstruction pipeline generated standardized tangential curvature maps. A hybrid model combined deep learning (ResNet for image features) and machine learning (using numerical parameters) to predict alignment curve (AC) and cylinder power (CP), with numerical regression for AC and classification regression for CP. Multitask learning addressed AC-CP biomechanical coupling.
Numerical parameter-based models achieved optimal axial AC prediction (mean absolute error [MAE] = 0.290, R2 = 0.917), and CP prediction (accuracy [ACC] = 0.798, area under the curve [AUC] = 0.791). The image-based deep learning model using baseline corneal topography alone attained acceptable AC prediction (MAE = 0.248, R2 = 0.850), yet demonstrated suboptimal CP classification accuracy (ACC = 0.674, AUC = 0.621). Hybrid modeling achieved breakthrough performance in AC prediction (MAE = 0.136, R2 = 0.973) and superior CP classification (ACC = 0.898, AUC = 0.896).
This system standardizes corneal topography across devices, addressing a critical barrier to generalizability in existing AI models, significantly enhancing fitting precision and generalizability for myopia control applications.
The device-agnostic design in the present study allows seamless integration into diverse clinical settings. The hybrid AI framework achieves near-expert accuracy, offering a scalable solution to access to high-quality OK lens fitting.
本研究旨在开发一种标准化人工智能（AI）系统，该系统整合角膜地形图图像和数值参数，以优化角膜塑形术（OK）镜片的验配。
分析了1153例佩戴欧几里得OK镜片患者（2341只眼）的回顾性数据。纳入519只眼（393例患者），其治疗区偏心度≤1mm用于模型训练。一个与设备无关的角膜地形图重建管道生成标准化的切向曲率图。一个混合模型结合深度学习（用于图像特征的ResNet）和机器学习（使用数值参数）来预测对齐曲线（AC）和柱镜度数（CP），AC采用数值回归，CP采用分类回归。多任务学习解决了AC-CP生物力学耦合问题。
基于数值参数的模型在轴向AC预测方面达到最优（平均绝对误差[MAE]=0.290，R2=0.917），在CP预测方面也达到最优（准确率[ACC]=0.798，曲线下面积[AUC]=0.791）。仅使用基线角膜地形图的基于图像的深度学习模型在AC预测方面达到了可接受的水平（MAE=0.248，R2=0.850），但在CP分类准确率方面表现欠佳（ACC=0.674，AUC=0.621）。混合建模在AC预测方面取得了突破性表现（MAE=0.136，R2=0.973），在CP分类方面表现优异（ACC=0.898，AUC=0.896）。
该系统对不同设备的角膜地形图进行了标准化，解决了现有AI模型在可推广性方面的关键障碍，显著提高了近视控制应用中验配的精度和可推广性。
本研究中与设备无关的设计允许无缝集成到各种临床环境中。混合AI框架实现了近乎专家级的准确率，为获得高质量的OK镜片验配提供了一种可扩展的解决方案。
