Get in touch
Computer vision · Explainable AI

RadiantCare

Research prototype For research use only

Chest X-ray AI that highlights the image regions behind its prediction and writes a plain-language report

A research prototype built on a public dataset of chest X-rays. A deep learning model scores 14 findings, and the whole system runs as a web app and as an API that other software can call. The point is trust: the main result comes with a heatmap of the image regions behind it, and a report in plain words.

Code on GitHub
Read what the demo shows
  1. 0:04 Upload a chest X-ray. Upload a chest X-ray, add optional clinical context, then click “Analyze”.
  2. 0:16 Heatmap, scores and report. A heatmap (Grad-CAM) highlights the image regions behind the result, and can be compared with the original X-ray. The model scores 14 findings, highest first, and an AI language model writes the report from the scores and the context.
  3. 0:34 Ask follow-up questions. Ask a follow-up question about the case, and the AI answers in plain words from the same scores and report.
  4. 0:52 The API. Everything is also available as a documented API that other software can call.
14 findings
scored on each chest X-ray
86,000
X-rays used to train the model
≈0.2 s
for the model to score one X-ray on an ordinary processor, no graphics card needed

The problem

Reading chest X-rays takes specialist time, and an AI model that only outputs a label is hard to trust. People need to see why it flagged something, and get the result in plain words.

What I built

  1. A deep learning model (DenseNet-121) gives a probability for each of 14 findings on an uploaded chest X-ray. It was trained on about 86,000 de-identified X-rays from the public Grand X-Ray Slam dataset on Kaggle, labelled with the standard CheXpert list of 14 findings.
  2. A heatmap (Grad-CAM) highlights the parts of the image that drove the top prediction, so a person can check what the model based its result on. It switches on and off with one click.
  3. An AI language model (GPT-4.1-mini) writes a short report from the scores and the optional clinical context, drawing on a local library of medical reference documents (retrieval-augmented generation), and answers follow-up questions in a chat.
  4. It runs as a web app and as a documented API, so other software can send an X-ray and get the results back. Docker packaging means it installs the same way on any server.
RadiantCare results: Grad-CAM heatmap, model predictions and the AI report
Heatmap, scores and report on one screen.

The result

Tested on 10,779 X-rays held back from training, the model reaches a mean AUROC of 0.92 across the 14 findings (0.83 to 0.97 per finding). AUROC measures how well a model ranks X-rays that have a finding above X-rays that do not: 0.5 is a coin toss, 1.0 is perfect.

Caveat: the split was by image, not by patient, so the same patient can appear in both training and test data, which likely makes the score optimistic. It has no clinical validation and is not a diagnostic tool.

How a business would use it

The same approach works wherever people look at images to make a decision. A model takes the first pass in a fraction of a second, a heatmap shows what each result was based on so your staff can check it instead of trusting a black box, and it runs as a service your existing software can call, not as a one-off experiment.

  • Spotting defects on a production line
  • Checking site, shelf or vehicle photos
  • Reading and sorting scanned documents
  • Any model that has to explain itself to be trusted
How a project like this runs for your business
Screens

Inside RadiantCare

Real screens from the running app. Tap or click any image to see it full size.

Upload screen with an X-ray and optional clinical context
Upload an X-ray and add optional clinical context.
The 14 findings with their probabilities
The model scores 14 findings, highest first.
Chat answering a follow-up question about the case
Follow-up questions, answered from the same scores.
Interactive API documentation listing the endpoints
A documented API, so your own software can send an image and get the result back.
Related services

Does your team check photos or scans by eye?

Product photos, site and shelf pictures, scanned forms: tell me what your team checks by eye, and I will tell you whether a model can take the first pass and what a first fixed-price step would cost.

Send a message

Free first conversation, no commitment.