Abstract
Purpose :
Proliferative diabetic retinopathy (PDR) is the primary pathological determinant to induce vision loss in diabetic patients. Laser treatment remains a pivotal strategy for restraining the onset of it. Typically, the timing of laser relies on ophthalmologists experience and lacks objective reference. Therefore, we developed a deep-learning-based auto-segmentation method to quantify non-perfusion areas (NPA) in fundus fluorescence angiography images, and analyzed the predicting value of NPA parameters in PDR risk.
Methods :
An U-net-based method was enhanced through three modules: adaptive encoder feature fusion, multilayer deep supervised loss, and atrous spatial pyramid pooling, enabling the integration of multi-scale features and contextual information. Biochemical and NPA parameters were collected from a cohort of 155 DR eyes in 124 patients. The value of HbA1c, total cholesterol, non-perfusion index (NPI), and NPA/optic disc area (ODA) in predicting retinal neovascularization was assessed through random forest analysis, and cut-off points were determined by Youden index and restricted cubic spline analysis.
Results :
The presented method exhibited superior performance in NPA identification compared to existing methods, including CE-net, deeplab, ConvNeXt, etc, with an AUC of 0.975, accuracy of 0.943, sensitivity of 0.879, specificity of 0.945, and Dice of 0.568. In random forest predicting model, the 5-fold cross-validation demonstrated a neovascularization predicting performance with an AUC value of 0.92, with NPI and NPA/ODA emerged as the most pivotal factors. Using cut-off values of NPA parameters alone (NPI=4.50% and NPA/ODA=16.14), the sensitivity values to stratify groups with/without neovascularization reached 0.924 and 0.909, respectively, with specificities of 0.718 and 0.788.
Conclusions :
The auto-segmentation method demonstrated precise detection of NPA in DR. Notably, NPA holds more crucial significance than biochemical parameters in assessing retinal ischemia. Cut-off values established based-on NPA parameters could serve as effective metrics to predict the presence of new vessels in DR eyes. This study offers strong evidence supporting the value and feasibility of establishing reference ranges for NPA parameters, which enable a quantitative assessment of PDR risk, and guide clinical determination of timing of laser treatment.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.