Application of artificial intelligence-based automated choroidal image segmentation and quantitative analysis in non-myopic children
摘要
本研究旨在利用扫频光学相干断层扫描(swept-source optical coherence tomography,SS-OCT)结合人工智能(artificial intelligence,AI)辅助的脉络膜分割算法,观察5~12岁未近视儿童黄斑区脉络膜参数特征,并探讨黄斑中心凹下脉络膜厚度(subfoveal choroidal thickness,SFCT)的相关影响因素.研究采用横断面研究方法,纳入83例未近视儿童右眼数据,通过SS-OCT进行黄斑区扫描,并基于U-Net架构的深度学习算法自动量化分析不同区域的脉络膜厚度(choroidal thickness,CT)、脉络膜血管管腔面积(luminal area,LA)、基质面积(stromal area,SA)、总面积(total choroidal area,TCA)及血管指数(choroidal vascularity index,CVI).结果显示,早期治疗糖尿病视网膜病变研究(early treatment diabetic retinopathy study,ETDRS)内、外环各象限区域的CT、LA、SA、TCA和CVI均存在显著差异,鼻侧脉络膜最薄、血流最少.单因素回归分析显示SFCT与角膜K值、瞳孔直径呈正相关,与眼轴长度呈负相关;多因素回归分析进一步表明,SFCT仅与瞳孔直径和眼轴长度独立相关,与其他参数无显著关联.结论认为,AI辅助脉络膜分析技术可高效量化儿童脉络膜参数,SFCT与瞳孔直径和眼轴长度显著相关,本研究为开发基于SS-OCT的近视风险智能预警系统提供了新思路,凸显了人工智能在近视防控领域的应用潜力.
This study aimed to quantitatively characterize macular choroidal parameters and investigate factors influencing subfoveal choroidal thickness(SFCT)in non-myopic children,leveraging the combined power of swept-source optical coherence tomography(SS-OCT)and an artificial intelligence(AI)-based automated segmentation algorithm.In this cross-sectional study,data from the right eyes of 83 non-myopic children aged 5 to 12 years were analyzed.Macular regions were scanned using SS-OCT,and a deep learning algorithm utilizing a U-Net architecture was employed for the fully automated segmentation of the choroid.This AI system enabled the precise quantification of a comprehensive set of choroidal parameters-including choroidal thickness(CT),luminal area(LA),stromal area(SA),total choroidal area(TCA),and choroidal vascularity index(CVI)—across the nine subfields of the standardized Early Treatment Diabetic Retinopathy Study(ETDRS)grid.Ocular biometric parameters,such as axial length(AL)and pupil diameter(PD),were concurrently measured.Our results revealed significant regional variations in CT,LA,SA,TCA,and CVI across the ETDRS inner and outer rings,with the nasal quadrant consistently exhibiting the thinnest choroid and the least blood flow.Univariate regression analysis indicated that SFCT was positively correlated with corneal keratometry(K-value)and pupil diameter,and negatively correlated with axial length.Multivariate regression analysis further refined these findings,demonstrating that SFCT was independently associated only with pupil diameter(positive correlation,β=0.04,P<0.001)and axial length(negative correlation,β=-0.061,P<0.001),with no significant independent relationships to other ocular parameters.In conclusion,AI-assisted choroidal image analysis proves to be an efficient and robust tool for quantifying intricate choroidal parameters in children.The independent correlations of SFCT with pupil diameter and axial length offer novel insights into the ocular physiology of non-myopic children.This work not only delineates normative choroidal topography in this specific population but also provides a foundational methodology and critical data for the future development of an SS-OCT-based intelligent early warning system for assessing myopia risk,thereby underscoring the significant potential of artificial intelligence in the realm of myopia prevention and control.