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

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Pachychoroid diseases, including central serous chorioretinopathy and pachychoroid pigment epitheliopathy, are characterized by specific choroidal changes.
• Accurate differentiation between pachychoroid and healthy choroids is crucial for diagnosis and management.
• Enhanced depth-optical coherence tomography (EDI-OCT) provides detailed imaging of the choroid.

Purpose of the Study:

• To introduce and validate a novel machine learning approach for distinguishing pachychoroid from healthy choroids using EDI-OCT data.
• To identify the most significant choroidal features for this differentiation.

Main Methods:

• Utilized EDI-OCT images from 103 patients with pachychoroid conditions and 103 age/sex-matched healthy controls.
• Extracted choroidal features: choroidal thickness (CT), choroidal area (CA), Haller layer thickness (HT), Sattler-choriocapillaris thickness (SCT), and choroidal vascular index (CVI).
• Calculated Haller ratio (HR = HT/CT) and applied multivariate TwoStep cluster analysis to differentiate groups.

Main Results:

• The machine learning model achieved high sensitivity (1.000) and specificity (0.986) in differentiating pachychoroid from healthy choroids.
• A combination of CT, HR, and CVI yielded the highest correct classification rate (0.993).
• Relative variable importance identified HR (1.0) and CVI (0.87) as the most critical features, followed by CT (0.70).

Conclusions:

• An unsupervised machine learning approach effectively differentiates pachychoroid from healthy choroids using EDI-OCT.
• The Haller ratio and choroidal thickness are the most valuable factors for clinical differentiation.
• Developed clinical criteria based on HR and CT achieved good sensitivity and specificity with high inter-rater reliability.