Chest X-ray Reader

How it works

A model is the easy part. Here's the rest.

Train & track

PyTorch training with patient-level splits, logged to MLflow. Weights versioned on Hugging Face and pinned by SHA-256.

Gate before release

CI runs tests and a quality gate: a new model can't ship if its AUROC drops against the baseline.

Export to the browser

CI exports the weights to ONNX and checks the output matches PyTorch and Grad-CAM before this page is published.

Guard the input

Colour photos, tiny images and odd shapes are flagged before anyone trusts the answer.

Explain

Grad-CAM shows where the evidence came from — and when the model is looking at the wrong thing.

Serve & monitor

The same model ships as a FastAPI Docker image with Prometheus, Grafana and drift detection for server deployments.

Model card

What this model is — and isn't.

Model: DenseNet121 initialised from ImageNet, multi-label sigmoid head. 224 × 224 input, ImageNet normalisation. Exported to ONNX for in-browser inference; Grad-CAM baked into the graph.
Training data: Stanford CheXpert: 159,329 frontal and lateral radiographs after dropping uncertain labels (U-Ignore). Labels were extracted from reports by an NLP labeller, so they're noisy.
Training: BCE with per-class positive weights, Adam 1e-4, 10 epochs, batch 32, one Colab T4 GPU (~4.5 h). Tracked in MLflow.
Evaluation: Mean AUROC 0.879 on the 234-image CheXpert validation set (radiologist consensus labels).
Atelectasis 0.818Cardiomegaly 0.855Consolidation 0.897Edema 0.906Pleural Effusion 0.919
Intended use: Research, education and demonstrating MLOps practice. Not a medical device; never for diagnosis or treatment decisions.
Limitations: Single-hospital training data, mostly adult inpatients. The checkpoint was selected on the validation set, so reported AUROC is likely optimistic. Performance drops on images from other hospitals, and Grad-CAM can highlight tubes or markers.
Versioning: Weights on Hugging Face, pinned by SHA-256 (7d88ee05…); the server refuses to start with any other file.

FAQ

Questions people ask.

The model

Is this a medical device?

No. It's a research and portfolio project. It hasn't been clinically validated and must not be used for diagnosis.

How accurate is it?

Mean AUROC 0.879 on the 234-image CheXpert validation set. That number is optimistic because the best epoch was picked on the same set — the next version fixes this with a held-out patient split and bootstrap confidence intervals. Expect lower scores on images from other hospitals.

What was it trained on?

159,329 chest radiographs from Stanford's CheXpert dataset (rows with uncertain labels were dropped), using a DenseNet121 pretrained on ImageNet, 10 epochs on one Colab T4 GPU.

What is the heatmap?

Grad-CAM: it highlights the regions that pushed the model's score up. It shows where the evidence came from, not whether the model is right — it can latch onto tubes, markers or text.

Your data

Where does my image go?

Nowhere. The model is downloaded once and runs in your browser with ONNX Runtime Web. Open your browser's network tab while you drop an image — nothing is sent.

Why did I get an input warning?

Chest X-rays are grayscale, reasonably large and roughly square. Colour photos, tiny thumbnails and panoramas are flagged so nobody trusts an answer the model was never trained to give.

Engineering

What keeps the model honest?

CI runs unit and API tests, a quality gate that blocks releases if AUROC drops, SHA-256 pinning of the weights, and a parity check proving the browser model matches PyTorch and Grad-CAM before this page is published.

Can I run it as a server?

Yes — docker compose up starts a FastAPI service with Prometheus, Grafana and live drift detection. The code is on GitHub.

159,329 X-rays. One pipeline.

The project, flipped into view.

Drag a film in. Get a second look.