Code-free automated machine learning for OCT-based classification of vitreoretinal interface diseases

Differentiation of vitreoretinal interface disorders on optical coherence tomography (OCT) relies on expert interpretation and can be challenging in borderline cases. Automated machine learning (AutoML) platforms may enable clinician-driven artificial intelligence development without coding expertise. This study evaluated the performance of a code-free AutoML approach for OCT-based classification.

In this cross-sectional image classification study, 434 OCT B-scans from publicly available datasets were manually labeled into four categories: epiretinal membrane (ERM), lamellar macular hole (LMH), full-thickness macular hole (MH), and normal retina. Images were uploaded to a cloud-based AutoML platform (Google Cloud Vertex AI), which automatically performed data splitting (80% training, 10% validation, 10% test), model training, and optimization. Performance was assessed using precision, recall, average precision, and confusion matrix analysis.

The model achieved an overall average precision of 0.988, with precision and recall of 97.6%. MH and normal retina were classified with perfect precision and recall (100%). ERM showed high precision (100%) with slightly reduced recall (92.9%), while LMH demonstrated complete recall (100%) with lower precision (83.3%). Misclassifications were limited to anatomically related entities.

Code-free AutoML enables accurate OCT-based classification of vitreoretinal interface disorders using a clinician-driven workflow. This approach may facilitate broader adoption of artificial intelligence in ophthalmology and support rapid clinical research prototyping.