👁️ Inside AI-Assisted Diabetic Retinopathy Screening

About this simulation

NHS Diabetic Eye Screening photographs the retina of every registered diabetic patient once a year. A convolutional neural network — usually a large network pretrained on natural images and then transfer-learned onto labelled fundus photographs — grades each image on the International Clinical Diabetic Retinopathy (ICDR) severity scale, from no retinopathy to sight-threatening proliferative disease.

How it works

• Fundus photograph — the camera on the left captures a wide-angle photo of the retina: blood vessels, optic disc, macula and, as disease progresses, microaneurysms, haemorrhages and exudates.
• Preprocessing — contrast-limited adaptive histogram equalisation (CLAHE) and colour normalisation make faint lesions easier for the network to detect before inference.
• Grad-CAM — gradient-weighted class activation mapping highlights which regions of the image most influenced the network's decision, giving clinicians an explainable overlay rather than a black-box score.
• ICDR grading — the model outputs a probability over five severity classes; the highest scoring class becomes the referral grade.

Controls

• ICDR severity grade — choose a ground-truth grade to see how lesion density and heatmap coverage change on the retina.
• Grad-CAM heatmap — fade the explainability overlay in or out over the fundus photo.
• CLAHE preprocessing — toggle contrast-enhanced imaging on and off.
• Run AI screening — animates the scan beam sweeping the retina and updates the grading panel, mirroring a real inference pass.
• Auto-rotate camera — slowly orbits the scene; drag to rotate manually, scroll to zoom.

Did you know?

Diabetic retinopathy is the leading cause of preventable blindness in working-age adults in the UK. Automated grading systems such as this have been validated to safely triage the large majority of "no referable disease" images, letting human graders focus on the borderline and vision-threatening cases.