Imaging Features by Machine Learning for Quantification of Optic Disc Changes and Impact on Choroidal Thickness in Young Myopic Patients

Frontiers in Medicine

article

Dandan Sun¹  Ying Fan¹  Xun Xu¹  Yuchen Du⁷  Qiuying Chen¹  Luyao Ye¹  Huai Chen⁷  Menghan Li¹  Jiangnan He¹  Jianfeng Zhu¹  Lisheng Wang⁷ 

[1]Department of Ophthalmology, Shanghai General Hospital, Shanghai Jiao Tong University School of Medicine

[2]National Clinical Research Center for Eye Diseases

[3]Shanghai Key Laboratory of Ocular Fundus Disease

[4]Shanghai Engineering Center for Visual Science and Photo Medicine

[5]Shanghai Engineering Center for Precise Diagnosis and Treatment of Eye Diseases

[6]Department of Preventative Ophthalmology, Shanghai Eye Disease Prevention and Treatment Center, Shanghai Eye Hospital

[7]Department of Automation, Institute of Image Processing and Pattern Recognition, Shanghai Jiao Tong University

关键词: myopia;    machine learning;    radiomics;    optic disc;    choroidal thickness;   

DOI  :  10.3389/fmed.2021.657566

学科分类：社会科学、人文和艺术（综合）

来源: Frontiers

【 摘 要 】

Purpose: To construct quantifiable models of imaging features by machine learning describing early changes of optic disc and peripapillary region, and to explore their performance as early indicators for choroidal thickness (ChT) in young myopic patients. Methods: Eight hundred and ninety six subjects were enrolled. Imaging features were extracted from fundus photographs. Macular ChT (mChT) and peripapillary ChT (pChT) were measured on swept-source optical coherence tomography scans. All participants were divided randomly into training (70%) and test (30%) sets. Imaging features correlated with ChT were selected by LASSO regression and combined into new indicators of optic disc (IODs) for mChT (IOD_mChT) and for pChT (IOD_pChT) by multivariate regression models in the training set. The performance of IODs was evaluated in the test set. Results: A significant correlation between IOD_mChT and mChT ( r = 0.650, R 2 = 0.423, P < 0.001) was found in the test set. IOD_mChT was negatively associated with axial length (AL) ( r = −0.562, P < 0.001) and peripapillary atrophy (PPA) area ( r = −0.738, P < 0.001) and positively associated with ovality index ( r = 0.503, P < 0.001) and torsion angle ( r = 0.242, P < 0.001) in the test set. Every 1 × 10 μm decrease in IOD_mChT was associated with an 8.87 μm decrease in mChT. A significant correlation between IOD_pChT and pChT ( r = 0.576, R 2 = 0.331, P < 0.001) was found in the test set. IOD_pChT was negatively associated with AL ( r = −0.478, P < 0.001) and PPA area ( r = −0.651, P < 0.001) and positively associated with ovality index ( r = 0.285, P < 0.001) and torsion angle ( r = 0.180, P < 0.001) in the test set. Every 1 × 10 μm decrease in IOD_pChT was associated with a 9.64 μm decrease in pChT. Conclusions: The study introduced a machine learning approach to acquire imaging information of early changes of optic disc and peripapillary region and constructed quantitative models significantly correlated with choroidal thickness. The objective models from fundus photographs represented a new approach that offset limitations of human annotation and could be applied in other areas of fundus diseases.

【 授权许可】

CC BY