Improving DR Diagnosis Accuracy with Deep Learning Using Adjudicated Grading

Globally, diabetic retinopathy, or DR, is one of the main causes of blindness. The effectiveness of deep learning models in diagnosing DR from retinal fundus images is hindered by inconsistencies in training data quality. This study addresses this issue by integrating adjudicated grading, a process involving independent assessments of image sets by trained readers with discrepancies resolved by an expert adjudicator. This approach is evaluated using the Messidor 2 dataset, comprising 1,200 retinal images with adjudicated grades for DR and diabetic macular edema (DME). The results show significant improvements in model accuracy,achieving an area under the receiver operating characteristic curve (AUC-ROC) of 0.95 for DR diagnosis and 0.91 for DME detection. This research highlights the potential of combining adjudicated grading and deep learning models to enhance the precision of DR diagnosis and screening, with broad implications for public health and healthcare delivery.