Retinal Image Analysis with Unsupervised Lesion Grouping

The challenge

This was a research project, not a product build — and that distinction shapes everything below.

Diabetic retinopathy is assessed from images of the retina, and the assessment depends on what is visible in them: where the affected regions are, and how much of the retina they cover. Doing that by eye is slow and hard to make consistent — two people looking at the same image can reasonably disagree on extent. "How much" is exactly the kind of question software measures better than people, provided the regions have been identified first.

The research question was whether that identification and measurement could be done reliably enough to be worth pursuing.

What we built

An analysis pipeline, and a web application to see its output.

• Region identification — locating affected areas within a retinal image.
• Area measurement — quantifying how much of the retina those regions cover, turning a subjective impression into a number.
• Unsupervised grouping — clustering the identified regions rather than classifying them against pre-labelled categories.
• A web application presenting the predicted result, so output was readable without running anything locally.

The dataset was proprietary and provided by the client. We did not source, collect or hold patient data independently — the client owned the data and the domain expertise, and we built the analysis around both.

Why unsupervised is the interesting choice

Supervised learning needs labelled examples of every category you want to recognise, and in medical imaging those labels are expensive: they require expert time, and experts disagree.

Clustering sidesteps that by grouping regions on their own similarity instead of against a fixed answer key. The trade-off is real — the groups it finds are not guaranteed to mean anything clinically useful, and interpreting them is a domain question rather than a modelling one. That is why this kind of work belongs alongside people who know the domain, not instead of them.

About that 92%

The system reached 92% accuracy on the client's dataset.

Both halves of that sentence matter. 92% is a strong research result. "On the client's dataset" is the boundary around it — a figure from one closed, proprietary dataset is evidence that an approach is worth pursuing, not evidence of how it behaves on data it has never seen. Generalisation is a separate question, answered by validation on independent data, and this project did not answer it.

We state it that way because research results quoted without their dataset are close to meaningless, and anyone technical enough to hire us for this work knows it.

What this is and isn't

It is a research system that performed well on the data it was given.

It is not a screening tool, not a diagnostic instrument, and makes no clinical claim. Whether an approach like this becomes any of those depends on validation, evidence and the regulatory pathway that applies — determinations for the product owner, not for the software vendor who built the prototype.

Interested in imaging or ML research?

We build AI and machine-learning systems as custom software, including image analysis pipelines and the applications that make their output usable — research prototypes included, where the goal is finding out whether something works before committing to build it properly.

If you have a problem in this space, get a free consultation.