Edge-Cloud Cascade for Automated Diabetic Retinopathy Screening
A new innovative edge-cloud system for automated retinal screening has been introduced to improve healthcare in rural areas. This two-tier framework employs a MobileNetV3-small model at local clinics to differentiate between referable and non-referable diabetic retinopathy (DR). For cases needing further evaluation, the RETFoundDINOv2 model in the cloud assesses image severity. In trials using the APTOS 2019 Blindness Detection dataset, which included 733 images, the system demonstrated remarkable sensitivity of 98.99% and specificity of 84.37%. This method effectively reduces latency, bandwidth, and data transfer costs by optimizing cloud processing.
Key facts
- Diabetic Retinopathy is a leading cause of preventable blindness.
- Rural regions lack specialists and infrastructure for early detection.
- Cloud-based deep learning faces challenges: high latency, limited bandwidth, high data transmission costs.
- Proposed two-tier edge-cloud cascade uses MobileNetV3-small for Tier 1.
- Tier 1 performs binary triage: Referable DR (Classes 2-4) vs Non-referable DR (Classes 0-1).
- Tier 2 uses RETFoundDINOv2 model in the cloud for ordinal severity grading.
- Only images flagged as referable by Tier 1 are sent to Tier 2.
- Tested on stratified APTOS 2019 dataset split of 733 images.
- Tier 1 achieved 98.99% sensitivity and 84.37% specificity.
Entities
—