A Deep Learning Segmentation Model for Detection of Active Proliferative Diabetic Retinopathy

• DOI: 10.1007/s40123-025-01127-w
• Publication Date: 2025-03-30T04:52:44Z
• Authors: Sebastian Dinesen; Marianne G. Schou; Christoffer V. He...; Yousif Subhi; Thiusius R. Savar...; Tunde Peto; Jakob K. H. Andersen; Jakob Grauslund

Existing deep learning (DL) algorithms lack the capability to accurately identify patients in immediate need of treatment for proliferative diabetic retinopathy (PDR). We aimed develop a DL segmentation model detect active PDR six-field retinal images by annotation new vessels and preretinal hemorrhages. identified classified at level 4 International Clinical Diabetic Retinopathy Disease Severity Scale collected Island Funen from 2009 2019 as part Danish screening program (DR). A certified grader (grader 1) manually dichotomized into or inactive PDR, were then reassessed two independent graders. In cases disagreement, final classification decision was made collaboration between 1 one secondary Overall, 637 PDR. applied our pre-established annotate nine lesion types before training algorithm. The segmentations hemorrhages corrected any inaccuracies After pre-segmentation phases divided (70%), validation (10%), testing (20%) datasets. added 301 with dataset. included 199 individuals. dataset had 1381 vessel hemorrhage lesions, while 123 lesions 374 lesions. system demonstrated sensitivity 90% specificity 70% annotation-assisted negative predictive value 94%, positive 57%. Our achieved excellent acceptable distinguishing