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

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Proliferative diabetic retinopathy (PDR) detection requires advanced tools for timely treatment.
• Current deep learning (DL) algorithms struggle with accurate PDR identification.

Purpose of the Study:

• Develop a DL segmentation model for active PDR detection using six-field retinal images.
• Annotate new retinal vessels and preretinal hemorrhages to identify active PDR.

Main Methods:

• Utilized six-field retinal images from the Danish diabetic retinopathy screening program.
• Manually classified images into active or inactive PDR by certified graders.
• Applied a DL segmentation model to annotate lesions, focusing on new vessels and preretinal hemorrhages.

Main Results:

• The DL model achieved 90% sensitivity and 70% specificity in classifying active PDR.
• Negative predictive value was 94%, and positive predictive value was 57%.
• The model successfully annotated new vessels and preretinal hemorrhages in 637 active PDR images.

Conclusions:

• The developed DL segmentation model demonstrates high sensitivity for active PDR detection.
• The model shows acceptable specificity in distinguishing active from inactive PDR.
• This tool can aid clinicians in identifying patients requiring urgent PDR treatment.