Chest X-Ray Classification with TensorFlow and React

The challenge

Chest X-rays are among the most common diagnostic images taken anywhere, and reading them is skilled work. The research question here was how much of the classification — what kind of finding is present, and how pronounced — a model could learn from images alone.

What we built

A model, and a web application to put it in front of someone.

• A TensorFlow 2 classification model, trained to distinguish finding type and severity from a chest X-ray image.
• A React web application where an image is uploaded and the predicted result is displayed — the difference between a model that exists in a notebook and one a non-engineer can actually try.

That second half is worth stating plainly. A large share of machine-learning work never reaches anyone who could evaluate it, because running it requires the environment it was built in. Wrapping a model in something a domain expert can open and use is often what turns a promising result into a decision.

The dataset was closed and provided by the client. We did not source, collect or hold patient data independently.

About that 91%

The model reached 91% accuracy on the client's dataset.

One caveat belongs with any multi-class figure, and we would rather state it than have a reader assume we missed it: raw accuracy flatters imbalanced classes. If some findings are far more common than others in the data, a model can score well overall while performing poorly on the rarer categories — which, in medical imaging, are frequently the ones that matter most. Per-class performance is the question that follows, and it is a different question from the headline number.

91% on a closed dataset is evidence that an approach is worth pursuing. It is not evidence of how the model behaves on images from another source, another scanner, or another population.

What this is and isn't

This was research. 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 against independent data, evidence, and whatever regulatory pathway 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 classification and the applications that make model output usable.

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