An edge-AI fundus camera that triages diabetic retinopathy at the point of care — the problem, the system, the model architecture, and the seven papers that back it up.
Build an edge-AI layer for standard fundus cameras so a technician, not a specialist, can triage diabetic retinopathy at the point of capture — clearing routine cases instantly, prioritizing sight-threatening ones, and getting measurably better with every patient scanned — so India's ~1-per-1.26-million retina specialists can reach far more people than manual grading ever could.
Diabetic Retinopathy is a leading cause of preventable blindness worldwide — and screening is bottlenecked not by the disease, but by how few specialists exist to grade it.
This project proposes a proprietary, AI-driven software ecosystem built directly into standard fundus cameras — a "smart camera" that acts as an automated point-of-care triage system, operable by technicians and allied health professionals rather than specialists alone.
The system does three things no manually-graded camp can do at scale: it filters patients in real time so specialists only see cases that need them, it guides technicians toward better image quality at the point of capture, and it learns continuously — improving across every device in the network without ever centralizing raw patient images.
The goal is not to replace the ophthalmologist's judgment, but to multiply it — turning one specialist's limited time into the deciding factor for hundreds of patients instead of dozens.
Shift the diagnostic burden of primary screening from the specialist to an edge-computing AI framework, fully integrated into the camera hardware and workflow.
The technician only needs to capture the image. The software instantly analyzes the retina for all four primary lesion types — microaneurysms, hemorrhages, hard exudates, soft exudates — and generates an automated severity report.
Patients are categorized in real time. Healthy individuals are cleared on the spot; high-risk cases are instantly queued for an on-site or remote ophthalmologist, sorted by clinical urgency rather than capture order.
The on-device CNN incrementally updates as new data arrives. Periodic federated synchronization across units shares model updates only — never raw patient images — so the whole network gets smarter together.
Every flagged case is shown to the doctor with a visual heatmap over the exact lesion the model detected — the clinician verifies in seconds rather than trusting a black box.
Filtering out routine cases redirects specialist time almost entirely toward high-value clinical decisions.
Camps process patients at far higher rates — no critical case is lost to fatigue or backlog, even in crowds of thousands.
Edge-optimized AI runs on standard hardware, no constant cloud connectivity or expensive infrastructure needed.
The federated continuous-learning loop compounds accuracy across every deployed camera — a static competitor can't catch up passively.
A prioritization tool, never a replacement for the doctor's final call. This is what makes it clinically trustworthy and adoptable.
If roughly 12–18% of screened patients turn out to have any DR, and only about 4% are vision-threatening, then in a purely manual workflow, 80–88% of a specialist's review time is spent clearing patients who didn't need specialist attention at all.
Reclaiming that time isn't a marginal efficiency gain — given India's roughly 1-specialist-per-1.26-million-population ratio, it may be the only realistic way to close the access gap without waiting a decade to train enough new specialists.
Every patient moved from "undetected vision-threatening DR" to "flagged and referred in time" is a case of preventable blindness avoided. The mechanism is faster, more consistent triage and referral — not an AI "diagnosing" anyone.
The camera concept isn't separate from the pipeline underway — it's the deployment target for it. The diagnostic core is a multi-stage pipeline:
Preprocessing for consistency across camera hardware and lighting. Lesion segmentation (Attention U-Net) to localize microaneurysms, hemorrhages, exudates. Binary triage (EfficientNet-B4) to separate healthy from at-risk at high sensitivity. Ordinal severity grading (DenseNet121) for prioritized routing.
In parallel, an ensemble of five CNNs — VGG16, VGG19, EfficientNetB5, EfficientNetB7, EfficientNetV2S — is combined via hard voting, soft voting, and a stacked-generalization (1D-CNN) meta-model, then validated on external datasets and eventually real hospital-sourced images.
Train and validate the segmentation and ensemble classification pipeline on public benchmark datasets — APTOS, Messidor-2, IDRiD.
Deploy the trained pipeline on a single fundus camera at one SSSIHL-affiliated medical camp, running alongside manual grading to validate against ophthalmologist ground truth.
Add the guided-capture assistant, explainability overlays, and auto-generated referral slips once the diagnostic core is validated.
Extend to multiple camera units across camps, enabling the privacy-preserving federated learning loop.
Expand to hospital partners and additional camp sites, with live dashboards feeding aggregate statistics back for institutional and public-health reporting.
Five of seven papers explicitly tested cross-dataset or external generalization. All five confirm it's achievable — accuracy just doesn't come for free once the data source changes.
Not "I need compute" — but "here's what comparable published work used to achieve real, tested generalization, and my request is sized to the low-to-middle end of that range."
"My request — a single GPU with at least 16GB VRAM, for approximately 80–100 GPU-hours to cover 25 training runs (5 architectures × 5-fold cross-validation) plus hyperparameter tuning and external generalization testing on Messidor-2 and IDRiD — is scoped conservatively relative to these benchmarks: single-GPU, not a cluster, and a shorter 25-epoch early-stopping budget than CMAC-Net's up to 2,000 epochs."
High-70s to high-80s% in-distribution (LLNet: 82.7–87.2%), dropping to the low-to-mid 70s% under real domain shift (LLNet: 72.1%). Not diagnostic-grade — and not meant to be. This is a triage filter that separates clear cases from ones needing a specialist. The doctor makes every final call.
Generalization testing is built in from the start. 5 of 7 reference papers explicitly tested cross-dataset generalization and all 5 confirm it's achievable. We follow the same protocol — APTOS → Messidor-2/IDRiD → real hospital images.
Ensembling covers individual blind spots. The literature backs this — the multi-task transformer paper found a simpler radiomics pipeline outperformed a single transformer under domain shift (88.3% vs. 83.8%).
These papers validate that the class of approach works, but none target an edge-deployable, technician-operable, self-learning triage system for Indian medical camps. We build on proven blocks (Attention U-Net, EfficientNet/DenseNet) — the real contribution is the integration into an actual screening ecosystem.
One GPU, 16GB+ VRAM, ~80–100 GPU-hours for 25 training configurations plus tuning and generalization testing — scoped below CMAC-Net's single-3090 setup and well below LLNet's two-GPU setup.
Two safeguards: learning is human-in-the-loop, so only doctor-verified corrections feed back into training; and automated drift monitoring flags any device whose confidence or flag-rate suddenly shifts.
The opposite. The system never renders a final diagnosis, only prioritizes and routes. Every flagged case still goes to a doctor. The value is multiplying how many patients one doctor's limited time can meaningfully reach.
The rollout degrades gracefully — even just segmentation and binary triage on APTOS alone produces a working, presentable proof of concept. Generalization testing and the ensemble/federated layers sit on top of that working core.