Retinal and choroidal microvasculature and structural analysis in OCTA for refractive amblyopia diagnosis using machine learning.

Objective: To evaluate the features of retinal and choroidal microcirculation and structure in patients with amblyopia compared to healthy adolescents of the same age (>10 years old). To classify and diagnose amblyopia using machine learning techniques on optical coherence tomographic angiography (OCTA) images.

Methods: Nineteen adolescents aged 11-17 with hyperopic refractive amblyopia and 22 age-matched healthy controls underwent 12 × 12 mm macular OCTA scans. The eyes were classified into three groups: amblyopic, contralateral non-amblyopic, and control. Retinal thickness (RT), choroidal thickness (ChT), and perfusion densities in the superficial capillary plexus (SCP) and deep capillary plexus (DCP) were measured across nine regions. A combination of statistical analysis and machine learning, including cross-validation and Random Forest classification, was used to enhance the diagnostic accuracy and classify amblyopic and normal eyes.

Results: Retinal thickness was significantly higher in the amblyopic eyes compared to the control group in multiple regions, including the central (p < 0.001), nasal (p < 0.01), and temporal zones(p < 0.01). Choroidal thickness was also greater in the amblyopic eyes, particularly in the central and nasal regions (p < 0.05). However, no significant differences were observed in the perfusion densities of SCP and DCP. The machine learning classification model incorporating cross-validation achieved an accuracy of 92%, with Random Forest demonstrating improved classification and feature importance analysis.

Conclusion: The results indicate that eyes with refractive amblyopia have notably thicker retinal and choroidal layers, particularly in the central and nasal regions. Combining OCTA data with machine learning creates a strong diagnostic framework for detecting changes in the retina and choroid associated with refractive amblyopia. Utilizing sophisticated classification methods, like Random Forest and cross-validation, improves diagnostic precision and presents new possibilities for automated clinical evaluation.