Comparing Auto-Machine Learning and Expert-Designed Models in Diagnosing Vitreomacular Interface Disorders

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Vitreomacular interface (VMI) disorders significantly impact vision and require precise classification.
• Accurate diagnosis of VMI disorders is crucial for effective patient management.
• Optical coherence tomography (OCT) is a key imaging modality for VMI assessment.

Purpose of the Study:

• To compare the diagnostic performance of a custom deep learning (DL) model against an automated machine learning (AutoML) model.
• To evaluate the effectiveness of these AI models in classifying various VMI disorders using OCT images.
• To determine the superiority of expert-designed AI versus automated AI in a clinical context.

Main Methods:

• A balanced dataset of OCT images was curated, including normal cases and five VMI disorder classes: epiretinal membrane (ERM), idiopathic full-thickness macular hole (FTMH), lamellar macular hole (LMH), and vitreomacular traction (VMT).
• An expert-designed DL model integrating ResNet-50 and EfficientNet-B0 architectures with Monte Carlo cross-validation was developed.
• A code-free AutoML model was implemented on Google Vertex AI for automated data processing, model selection, and hyperparameter tuning.

Main Results:

• The expert-designed DL model achieved 95.97% balanced accuracy and 94.65% MCC, outperforming AutoML in classifying FTMH, ERM, and LMH.
• Both models demonstrated perfect precision and recall for normal OCT images.
• AutoML showed strong performance in VMT detection (99.5% precision) but lagged in LMH classification (72.3% precision).

Conclusions:

• Expert-designed deep learning models demonstrate superior accuracy for specific vitreomacular interface disorder classifications compared to current AutoML platforms.
• While AutoML offers accessibility for healthcare professionals, further advancements are needed to match expert-driven AI performance in clinical OCT image analysis.
• The findings highlight the potential of tailored AI solutions for improving the diagnostic accuracy of retinal disorders.