Edge-cloud cascade cuts cloud costs 50% for retinal screening
98.99% sensitivity at the edge, 50% fewer cloud calls, 80% accuracy retained.
A new paper from Nishi Doshi and Shrey Shah tackles the rural healthcare gap in diabetic retinopathy (DR) screening by proposing a cascaded edge-cloud architecture. Their system uses a two-tier design on the APTOS 2019 dataset: Tier 1 runs a lightweight MobileNetV3-small model on a local clinic device to perform binary triage (referable vs non-referable DR). At a tuned high-sensitivity threshold, Tier 1 achieves 98.99% sensitivity and 84.37% specificity, ensuring few false negatives. Only the 49.52% of images flagged as referable are sent to the cloud (Tier 2), reducing cloud transmission and processing costs by 50.48% compared to sending all images.
Tier 2 uses the larger RETFoundDINOv2 model for ordinal severity grading (classes 0-4). The cascade achieves 80.49% accuracy and 0.8167 quadratic weighted kappa, nearly matching the cloud-only baseline (80.76% accuracy, 0.8184 kappa) while cutting cloud usage in half. This architecture addresses key barriers in rural settings—high latency, limited bandwidth, and high data costs—making automated retinal screening feasible in low-resource clinics. The work underscores how lightweight edge models combined with selective cloud offloading can bridge healthcare gaps without sacrificing accuracy.
- Tier 1 (MobileNetV3-small) triages images with 98.99% sensitivity at the edge, catching nearly all referable cases.
- Cascade forwards only 49.52% of images to the cloud, cutting cloud calls by 50.48% vs a full cloud pipeline.
- Tier 2 (RETFoundDINOv2) retains 80.49% accuracy for 4-class severity grading, just 0.27% below the cloud-only baseline.
Why It Matters
Enables affordable, automated retinal screening in rural clinics with poor connectivity, preventing blindness from diabetic retinopathy.