Abstract
Purpose :
Previous reports suggest that the retinal vasculature of diabetic patients is altered before the onset of diabetic retinopathy (DR). However, the morphological details of the vascular change remain unknown. Previously, we reported a deep learning algorithm to perform semantic segmentation of retinal vessels from fundus photographs (Fukutsu et al. Ophthalmology Science. 2021). In this study, we sought to detect and visualize the alteration of retinal vasculature in diabetic patients without DR using a vascular distribution map created based on our algorithm.
Methods :
Fundus images (n=10460) from 5569 individuals who went through a regular health checkup were analyzed. Blurred photos and pictures with apparent DR findings such as dot hemorrhages were excluded. Our deep learning algorithm automatically generated retinal vascular images (arterioles and venules) from the fundus images. The retinal vascular images of 10460 eyes went through an additive average. Each pixel of the additionally averaged image was classified into 256 luminance levels, and a heat map based on the luminance level was defined as a vascular distribution (probability) map. The difference in vascular distribution by pixels was analyzed between pictures of individuals with diabetes (diabetic group, n=495) and those without diabetes (control group, n=9965).
Results :
Regarding the arterioles, the vascular distribution around the main stem vessels increased by 9727 pixels compared to the control group at 20 luminance levels (p<0.05), while the distribution around the peripheral vessels and macula decreased by 28141 pixels compared to the control group at 61 luminance levels in the diabetic group (p<0.05). As for the retinal venules, the vascular distribution of the diabetic group decreased by 111139 pixels at 61 luminance levels compared to the control group (p<0.05) in the widespread area, particularly in the peri-macular area.
Conclusions :
Retinal vascular changes before the onset of DR in diabetic patients were detected using a vascular distribution map. This method could be used for onset prediction of DR from fundus photographs.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.