RetinaScan AI — explainable diabetic retinopathy screening

The problem

Diabetic retinopathy is one of the leading causes of preventable blindness in working-age adults, and it is entirely manageable when caught early. The problem is catching it. Screening depends on a specialist reading a retinal photo and judging where it falls across five stages of severity, from no disease to sight-threatening proliferative disease. That is slow, it does not scale to the volume of screening diabetic populations actually need, and every "AI can do this instead" pitch runs into the same wall: a black-box model that outputs a number with no way to see why. Clinicians will not act on a diagnosis they cannot inspect, and a screening tool nobody trusts is a screening tool nobody uses.

Approach

We designed RetinaScan AI around a simple principle: a diagnosis is only useful if you can see the reasoning behind it. At the core is a ResNet50 convolutional network fine-tuned to classify retinal fundus images into the five clinically recognised DR severity stages.

Every prediction is paired with a Grad-CAM attribution map highlighting the regions of the retina the model weighted most heavily, rendered as an interactive overlay so a viewer can blend between the raw photo and the model's heatmap and see precisely where the evidence lives. On top of that sits a clinical explanation layer that translates the activated regions into plain-language findings — microaneurysms, hard exudates, haemorrhages, neovascularisation — each with its own description on demand, so the output reads like a report rather than a probability score.

The whole thing ships as a single-page diagnostic workstation: drop in a fundus photo and get the classification, the heatmap overlay and the clinical write-up in one pass, from a Flask API behind a React front end. Procedurally generated sample cases spanning normal, moderate and severe presentations let the full pipeline be demonstrated end to end without depending on real patient imagery.

Outcome

An inspectable alternative to black-box DR screening: one photo in, a severity read and a visual, plain-language explanation out. Rather than trusting a single number, a viewer can see which part of the retina drove the classification and what clinical feature it corresponds to — which is the gap between "the model said so" and "here is the evidence."