Abstract
Purpose :
Diabetic retinopathy (DR) is a leading cause of vision loss in adults. Neovascularization (NV) and nonperfusion (NP) are angiographic features of DR. This study evaluated the association of quantitative imaging biomarkers including NV and NP with disease progression and advanced disease and whether biomarkers could be identified using artificial intelligence (AI).
Methods :
A retrospective study was performed on ultrawide-field fluorescein angiography (UWF FA) images obtained at the UM Kellogg Eye Center after IRB approval. Demographic & clinical data were acquired. Trained, masked graders segmented regions of NP and NV. Optos research software determined the surface area in mm2 of the segmented regions of NP and NV. AI identification was performed with 5-fold cross-validation using an ensemble of 3 convolutional neural networks (CNN): ResNet152V2, EfficientNetB6, and InceptionResNetV2.
Results :
651 eyes from 363 patients were included: 76 eyes with no DR, 92 mild NPDR, 144 moderate NPDR, 101 severe NPDR, 220 PDR, & 18 unknown. Mean age was 59.4 ± 13.7 yrs. Mean HbA1c was 8.1% (SD=2.1%). Mean visual acuity was logMAR 0.35 (SD=0.30). Mean NP area was 80.8mm2 (SD=63.3). We identified a threshold NP area of 77.48 mm2, above which patients have an increased risk of developing PDR (sensitivity 59.5%, specificity 73.6%). These patients received a total of 2556 anti-VEGF injections with a mean follow up of 915 ± 714 days. 46 eyes received anti-VEGF treatment (mean 8.6 ± 8.0 injections) and had more than 1 FA (mean 30.4 ± 14.7 months of follow up). 19 eyes received no anti-VEGF and had more than 1 FA (mean 26.0 ± 17.2 months of follow up). The mean area of NP was 23.3 mm2 in the anti-VEGF group compared to 30.4 mm2 in the no anti-VEGF group. Using AI to identify NV, the area under the ROC curve (AUC) was 0.96 for the model-averaged ensemble predictor. The AUCs for each individual CNN in the ensemble were 0.90 for InceptionResNetV2, 0.92 for EfficientNetB6, and 0.94 for ResNet152V2. The average precision was 0.87 for the model-averaged ensemble predictor. Individual CNN average precisions were 0.76 for InceptionResNetV2, 0.79 for EfficientNetB6, and 0.83 for ResNet152V2.
Conclusions :
Quantifiable biomarkers on UWF FA can predict progression of PDR. Treatment with anti-VEGF could reduce NP area. CNNs can accurately classify NV and distinguish this biomarker from other FA features.
This abstract was presented at the 2023 ARVO Imaging in the Eye Conference, held in New Orleans, LA, April 21-22, 2023.