Deep Learning-Assisted Retinal Vessel Segmentation for Diabetic Retinopathy Screening
Okonkwo, J., Adeyemi, F., Nwosu, C., Eze, I.. Deep Learning-Assisted Retinal Vessel Segmentation for Diabetic Retinopathy Screening. IJBE.
Vol.1, No.1. Mar 2025. https://doi.org/10.58922/ijbe.2025.0102
Article
Recommended articles
Cited by 18
Metrics
Highlights
- Diabetic retinopathy (DR) affects approximately 100 million people worldwide, with timely screening critical for preventing vision loss.
- RetinalNet-Lite achieves 98.1% sensitivity and 97.8% specificity for DR screening using a lightweight CNN deployable on mobile devices without cloud connectivity.
- The architecture employs depthwise separable convolutions and knowledge distillation from a larger teacher network, achieving 94% parameter reduction with <1.5% accuracy penalty.
Abstract
Diabetic retinopathy (DR) affects approximately 100 million people worldwide, with timely screening critical for preventing vision loss. RetinalNet-Lite achieves 98.1% sensitivity and 97.8% specificity for DR screening using a lightweight CNN deployable on mobile devices without cloud connectivity. The architecture employs depthwise separable convolutions and knowledge distillation from a larger teacher network, achieving 94% parameter reduction with <1.5% accuracy penalty. Validated prospectively on 12,000 retinal images across three clinical sites in Nigeria, Ghana, and Kenya, RetinalNet-Lite demonstrates consistent performance across population demographics, camera models, and lighting conditions representative of resource-limited settings.