Differentiating a pachychoroid and healthy choroid using an unsupervised machine learning approach

• Published in: Scientific reports Vol. 12; no. 1; pp. 16323 - 9
• Main Authors: Mirshahi, Reza, Naseripour, Masood, Shojaei, Ahmad, Heirani, Mohsen, Alemzadeh, Sayyed Amirpooya, Moodi, Farzan, Anvari, Pasha, Falavarjani, Khalil Ghasemi
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 29.09.2022; Nature Publishing Group; Nature Portfolio
• Subjects: 631/114/1305; 692/699/3161/3175; Cluster analysis; Computed tomography; Diabetic retinopathy; Disease; Humanities and Social Sciences; Learning algorithms; Machine learning; Medical imaging; multidisciplinary; Ophthalmology; Research centers; Retina; Science; Science (multidisciplinary); Software
• ISSN: 2045-2322; 2045-2322
• DOI: 10.1038/s41598-022-20749-9

The purpose of this study was to introduce a new machine learning approach for differentiation of a pachychoroid from a healthy choroid based on enhanced depth-optical coherence tomography (EDI-OCT) imaging. This study included EDI-OCT images of 103 eyes from 82 patients with central serous chorioretinopathy or pachychoroid pigment epitheliopathy, and 103 eyes from 103 age- and sex-matched healthy subjects. Choroidal features including choroidal thickness (CT), choroidal area (CA), Haller layer thickness (HT), Sattler-choriocapillaris thickness (SCT), and the choroidal vascular index (CVI) were extracted. The Haller ratio (HR) was obtained by dividing HT by CT. Multivariate TwoStep cluster analysis was performed with a preset number of two clusters based on a combination of different choroidal features. Clinical criteria were developed based on the results of the cluster analysis, and two independent skilled retina specialists graded a separate testing dataset based on the new clinical criteria. TwoStep cluster analysis achieved a sensitivity of 1.000 (95-CI: 0.938–1.000) and a specificity of 0.986 (95-CI: 0.919–1.000) in the differentiation of pachy- and healthy choroid. The best result for identification of pachychoroid was obtained for a combination of CT, HR, and CVI, with a correct classification rate of 0.993 (95-CI: 0.980–1.000). Based on the relative variable importance (RVI), the cluster analysis prioritized the choroidal features as follows: HR (RVI: 1.0), CVI (RVI: 0.87), CT (RVI: 0.70), CA (RVI: 0.59), and SCT (RVI: 0.27). After performing a receiver operating characteristic curve analysis on the cluster membership variable, a cutoff point of 389 µm and 0.79 was determined for CT and HR, respectively. Based on these clinical criteria, a sensitivity of 0.793 (95-CI: 0.611–0.904) and a specificity of 0.786 (95-CI: 0.600–0.900) and 0.821 (95-CI: 0.638–0.924) were achieved for each grader. Cohen's kappa of inter-rater reliability was 0.895. Based on an unsupervised machine learning approach, a combination of the Haller ratio and choroidal thickness is the most valuable factor in the differentiation of pachy- and healthy choroids in a clinical setting.