Bridging the Rural Healthcare Gap: A Cascaded Edge-Cloud Architecture for Automated Retinal Screening

2026-05-15 · Source: cs.CV updates on arXiv.org · Field: Health & Wellbeing — Medical Devices & Health Technology, Clinical Care & Medical Practice · Depth: Advanced, extended

Summary

A two-tier edge-cloud cascade architecture has been developed for automated Diabetic Retinopathy (DR) screening, specifically targeting rural regions with limited infrastructure. This system, evaluated on the APTOS 2019 Blindness Detection dataset, employs a lightweight MobileNetV3-small model as Tier 1 on a local clinic device for binary triage, classifying images as Referable DR (Classes 2–4) or Non-referable DR (Classes 0–1). Only images flagged as referable are then forwarded to Tier 2, a cloud-based RETFound-DINOv2 model, for ordinal severity grading. The cascade achieved 98.99% sensitivity and 84.37% specificity for Tier 1 triage. It reduced cloud calls by 50.48% by forwarding only 49.52% of test images to Tier 2. In the deployed 4-class output space, the cascade maintained 80.49% accuracy and a 0.8167 quadratic weighted kappa, closely matching the cloud-only baseline's 80.76% accuracy and 0.8184 kappa, demonstrating significant cloud utilization reduction with minimal performance impact.

Key takeaway

For AI Architects and AI Engineers designing healthcare solutions for low-resource settings, this cascaded edge-cloud architecture offers a practical blueprint. Your teams should consider implementing a local triage model to significantly reduce cloud infrastructure costs and bandwidth dependency, especially for high-resolution medical imaging. This approach allows for maintaining high diagnostic accuracy for critical cases while optimizing resource utilization, making advanced screening more accessible in rural or underserved areas.

Key insights

An edge-cloud cascade reduces cloud usage for DR screening while maintaining high diagnostic performance.

Principles

• Prioritize sensitivity in edge triage to avoid missed referable cases.
• Match model complexity to task difficulty across tiers.
• Optimize for local processing to mitigate bandwidth and latency issues.

Method

A two-tier cascade uses a local MobileNetV3-small for binary triage (referable/non-referable) and a cloud-based RETFound-DINOv2 for ordinal grading, forwarding only referable cases to the cloud.

In practice

• Implement MobileNetV3-small for initial DR triage.
• Use RETFound-DINOv2 for detailed cloud-based grading.
• Configure Tier 1 with a high-sensitivity threshold.

Topics

• Diabetic Retinopathy Screening
• Edge-Cloud Architecture
• Automated Retinal Screening
• MobileNetV3-small
• RETFound-DINOv2

Code references

• ClementPla/FundusDRGradingSpecific

Best for: AI Architect, AI Engineer, Computer Vision Engineer, AI Scientist, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Edge AI — Edge computing with integrated AI acceleration is crucial for real-time, deterministic healthcare applications.
• MPD$^2$-Router: Mask-aware Multi-expert Prior-regularized Dual-head Deferral Router in Glaucoma Screening and Diagnosis — MPD$^2$-Router optimizes glaucoma screening by routing complex cases to available human experts, balancing cost and diagnostic accuracy.
• Intelligent Healthcare Imaging Platform: A VLM-Based Framework for Automated Medical Image Analysis and Clinical Report Generation — A VLM-based framework automates medical image analysis and clinical report generation across multiple imaging modalities.
• StreamPPG: Low-Latency rPPG Estimation via Consistent Privileged Learning — StreamPPG leverages consistent privileged learning to enable low-latency, accurate remote photoplethysmography estimation suitable for real-time edge device deployment.
• A Dual Edge Spatial Jacobian Image Graph for Interpretable Diabetic Retinopathy Grading — Interpretable DR grading benefits from graph representations linking spatial lesion-vessel geometry and embedding-biomarker sensitivity.
• JASMINE — Early skin cancer detection in rural areas can be significantly improved with low-cost, offline, edge AI devices.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.