Automated Machine Learning for Referable Diabetic Retinopathy Image Classification from Ultrawide Field Images

Abstract

Objective: To develop and evaluate the diagnostic performance of an automated machine learning (AutoML) model for the detection of referable diabetic retinopathy (refDR) in ultrawide field (UWF) retinal images from local Philippine retinal image datasets.

Methods: A Google AutoML Vision model was trained using 2000 UWF images with a 50/50 ratio of refDR/non-refDR. Images were labeled according to the Early Treatment Diabetic Retinopathy Study (ETDRS) severity grading. RefDR was defined as moderate nonproliferative DR or worse. The dataset was split with 80% for training, 10% for validation, and 10% for testing. Two sets of published UWF image sets were used for external validation. Sensitivity and specificity were calculated in accordance with United States Food and Drug Administration (US FDA) performance requirements of 0.85 and 0.825, respectively.

Results: The area under the precision-recall curve was 0.998. External validation against two datasets showed a sensitivity/specificity of 0.88/0.83 (95% CI 0.80-0.94/0.74-0.89) and 0.83/0.80 (95% CI 0.74-0.89/0.72-0.86), respectively. Positive and negative predictive values were 0.81/0.89 (95% CI 0.73-0.89/0.82-0.94) and 0.75/0.86 (95% CI 0.66-0.83/0.79-0.91), respectively.

Conclusions: The pilot performance of the custom AutoML model constructed using local Philippine data approaches US FDA requirements for the diagnosis of referable DR. The ease of use and intuitiveness of the platform, combined with its performance, support the potential of no-code AI in the detection of refDR.