A FastAPI dashboard around the pretrained TorchXRayVision DenseNet121 chest X-ray classifier. It shows the model's per-pathology scores, bands them for display, and can ask a local LLM (Ollama) to rephrase the scores in plain language.
Not a medical device. This is experimental research and learning software. The model was not trained or validated by this project, its scores are uncalibrated, and nothing here may be used for diagnosis or treatment decisions.
- It is an application layer: upload handling, inference wrapper, score banding, an optional LLM explanation step, a web dashboard and a CLI.
- It is not a new model. Classification is done entirely by the
TorchXRayVision weights
densenet121-res224-all, used unchanged. No training, fine-tuning or evaluation was performed in this repository, so no accuracy figures are reported here. See the TorchXRayVision paper and repository for the model's own evaluation.
- Multi-label pathology scores for a chest X-ray (top 5 shown)
- Display banding per score: LOW (≤ 40%), MODERATE (40-70%), HIGH (> 70%), and an overall status from the highest score. These cut-offs are presentation choices, not clinically derived thresholds.
- Optional plain-language explanation from a local Ollama model, with fixed non-diagnostic wording when Ollama is not running
- Web dashboard (vanilla JS + Chart.js) and a command-line tool
- Uploads are validated, analysed and deleted; nothing is stored or served back
Requires Python 3.10+. The first run downloads the model weights (TorchXRayVision caches them in your user directory).
git clone https://github.com/projectbin07-stack/xray-analysis-aiml.git
cd xray-analysis-aiml
python -m venv .venv
# Windows: .venv\Scripts\activate Linux/macOS: source .venv/bin/activate
pip install -r requirements.txt
python app.pyOpen http://127.0.0.1:8001 and upload a JPG or PNG chest X-ray. The server
binds to localhost only; set HOST=0.0.0.0 to expose it on your network.
Command line, without the web server:
python -m core.xray_engine --image path/to/xray.png --top-k 5Install Ollama, run ollama pull llama3, and keep it
running. Without it the dashboard still works and shows a fixed explanation.
| Variable | Default | Purpose |
|---|---|---|
HOST |
127.0.0.1 |
Bind address |
PORT |
8001 |
Port |
OLLAMA_URL |
http://localhost:11434/api/generate |
Local LLM endpoint |
OLLAMA_MODEL |
llama3 |
Model used for the explanation |
OLLAMA_TIMEOUT_S |
15 |
Timeout before the fixed explanation is used |
LOG_LEVEL |
INFO |
Python logging level |
POST /analyze with multipart field image (jpg, png or bmp, up to 15 MB).
The numbers below only illustrate the response shape.
{
"predictions": [{"disease": "Effusion", "confidence": 82.0, "risk": "HIGH"}],
"status": "CRITICAL",
"explanation": {"summary": "...", "meaning": "...", "recommendation": "..."},
"confidence_scores": [82.0],
"labels": ["Effusion"]
}Errors are returned as {"error": "..."} with a 4xx/5xx status.
GET /health returns {"status": "ok", "model_loaded": true|false}.
upload ──> validate (type, size) ──> temp file ──> XRayInferenceEngine
│ grayscale, normalise,
│ centre-crop, 224x224
▼
DenseNet121 (TorchXRayVision)
│ sigmoid scores
▼
core/report.py: banding + status ──> JSON ──> dashboard
│
└─> optional Ollama rephrasing
| Path | Role |
|---|---|
app.py |
FastAPI routes, upload validation, Ollama call |
core/xray_engine.py |
Model loading, preprocessing, inference, CLI |
core/report.py |
Score banding and LLM-output parsing (no PyTorch dependency) |
static/, templates/ |
Dashboard |
tests/ |
pytest suite |
pip install -r requirements-dev.txt
ruff check . && ruff format --check .
pytestCI runs these on Python 3.10 and 3.13. Tests cover score banding, LLM-output
parsing, upload validation, upload deletion and error handling, with the model
replaced by a fake. CI does not install PyTorch or run the real network, so
core/xray_engine.py is verified only by manual runs.
- Scores are raw sigmoid outputs of a pretrained multi-dataset model; they are not calibrated probabilities and the display bands are arbitrary.
- No check that the uploaded image is actually a frontal chest X-ray; any image produces scores.
- DICOM is not supported (JPG/PNG/BMP only) and images are centre-cropped to a square, which can cut off anatomy.
- No localisation or saliency maps: the "high contrast" toggle is a display filter, not model output.
- LLM explanations can be wrong or overconfident; the prompt forbids diagnosis but output is not otherwise checked.
- Grad-CAM overlays for the top predictions
- CI job that runs the real model on a public sample image
- Input sanity check (reject non-X-ray images)
- DICOM input
![]() |
![]() |
The screenshots were taken with an earlier version of the interface; a panel they show with illustrative "model performance" figures has since been removed because those numbers were not measurements.
MIT for the code in this repository - see LICENSE. TorchXRayVision and its model weights are distributed under their own terms (Apache-2.0 for the library; check the upstream repository for dataset-related conditions).
See CONTRIBUTING.md and SECURITY.md. Never post identifiable patient images in issues or pull requests.
Navaneeth KG


