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

Background: The vitreomacular interface (VMI) encompasses a group of retinal disorders that significantly impact vision, requiring accurate classification for effective management. This study aims to compare the effectiveness of an expert-designed custom deep learning (DL) model and a code free Auto Machine Learning (ML) model in classifying optical coherence tomography (OCT) images of VMI disorders. Materials and Methods: A balanced dataset of OCT images across five classes-normal, epiretinal membrane (ERM), idiopathic full-thickness macular hole (FTMH), lamellar macular hole (LMH), and vitreomacular traction (VMT)-was used. The expert-designed model combined ResNet-50 and EfficientNet-B0 architectures with Monte Carlo cross-validation. The AutoML model was created on Google Vertex AI, which handled data processing, model selection, and hyperparameter tuning automatically. Performance was evaluated using average precision, precision, and recall metrics. Results: The expert-designed model achieved an overall balanced accuracy of 95.97% and a Matthews Correlation Coefficient (MCC) of 94.65%. Both models attained 100% precision and recall for normal cases. For FTMH, the expert model reached perfect precision and recall, while the AutoML model scored 97.8% average precision, and 97.4% recall. In VMT detection, the AutoML model showed 99.5% average precision with a slightly lower recall of 94.7% compared to the expert model's 95%. For ERM, the expert model achieved 95% recall, while the AutoML model had higher precision at 93.9% but a lower recall of 79.5%. In LMH classification, the expert model exhibited 95% precision, compared to 72.3% for the AutoML model, with similar recall for both (88% and 87.2%, respectively). Conclusions: While the AutoML model demonstrated strong performance, the expert-designed model achieved superior accuracy across certain classes. AutoML platforms, although accessible to healthcare professionals, may require further advancements to match the performance of expert-designed models in clinical applications.