Quality gate before grading
Blur, illumination, evenness, contrast and field-of-view are scored first. Ungradable images get spoken recapture instructions instead of a false result.
Screening support · never a diagnosis
Explainable diabetic retinopathy screening for the last mile
Built on a MATLAB image-processing and deep-learning pipeline: judge the photograph, enhance it, show the lesions, grade it on the international scale — and hand doubtful eyes to a human.
Evidence you can point at
Microaneurysms, haemorrhages and exudates are circled on the image, with a Grad-CAM++ map for the predicted grade.
Calibrated, and it can abstain
Temperature-scaled confidence and entropy-based uncertainty route doubtful cases to an ophthalmologist rather than guessing.
Runs at the camp, not the cloud
The whole pipeline executes on the device so screening continues without connectivity; records sync later.
Scaled with a Simulink twin
A queueing simulation sizes cameras, edge compute, bandwidth and specialist hours for 100,000+ patients a year.
Auditable by design
Original images are preserved unmodified and every screening keeps its evidence trail for review.