Abstract
Purpose :
This study aims to evaluate the layer information fusion options in optical coherence tomography angiography (OCTA) for deep learning classification of diabetic retinopathy (DR).
Methods :
OCTA images were acquired from normal eyes, diabetic eyes with no diabetic retinopathy (NoDR), and diabetic eyes with non-proliferative DR (NPDR). For each eye, three en-face projections from the superficial, deep, and choriocapillaris layers were generated. Deep learning models based on VGG architecture were used to accomplish the classification task. Individual OCTA layers were trained separately first. Then, to evaluate the layer fusion effect, three OCTA layers were utilized through the early, intermediate, and late fusion architectures. Saliency maps were employed to determine which areas of the image were engaged in the deep learning model.
Results :
For individual OCTA layer classification, the superficial OCTA achieved the better performance, with 87.74% accuracy, 80% sensitivity, and 91.08% specificity, to differentiate control, NoDR, and NPDR. For multi-layer fusion strategy, the intermediate fusion architecture performs the best. The early fusion model achieved an average of 86.77% accuracy, 79.37% sensitivity, and 90.63% specificity, the intermediate fusion model achieved an average of 93.18% accuracy, 87.45% sensitivity, 94.55% specificity, and the late fusion model achieved an average of 91.67% accuracy, 85.03% sensitivity, 93.5% specificity for the same categorization task. A typical example of an NPDR subject is shown in Fig 1E. The regions with red to brown color were the regions the model used to make decisions.
Conclusions :
With superficial, deep, and choriocapillaris OCTA layers involved, the intermediate fusion model provides the best performance of deep learning classification of DR.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.