Educational demonstration — not a medical device, no real patient data is used, and the generated scores are not clinically validated.

About the AI

What model this demo uses, what it was trained on, and — just as important — what it can't do.

This application is a university prototype and is not a certified medical device.

The model

TorchXRayVision (open source), weights densenet121-res224-all — a DenseNet121 convolutional network. It runs entirely on this machine (CPU, or GPU/MPS if available) — no external API call, no data leaves your computer.

Reference: "TorchXRayVision: A library of chest X-ray datasets and models" — arXiv:2111.00595

Training data

The "all" weights are jointly trained on 7 public chest X-ray datasets combined — well over 800,000 images in total:

  • NIH ChestX-ray14NIH Clinical Center — ~112,000 images
  • PadChestHospital San Juan, Spain — ~160,000 images
  • CheXpertStanford — ~224,000 images
  • MIMIC-CXRMIT / Beth Israel — ~377,000 images
  • NIH (Google-relabeled)Radiologist re-annotation, 2019 paper
  • Open-IIndiana University — ~7,500 images
  • RSNA Pneumonia ChallengeKaggle / RSNA

Supported pathologies (18)

Only these — the app never shows a pathology the model doesn't actually predict:

AtelectasisConsolidationInfiltrationPneumothoraxEdemaEmphysemaFibrosisEffusionPneumoniaPleural_ThickeningCardiomegalyNoduleMassHerniaLung LesionFractureLung OpacityEnlarged Cardiomediastinum

Limitations

  • Chest X-rays only. The model has never seen a skull, limb, spine, or abdomen image. Running it on anything other than a chest X-ray produces a meaningless score, not a valid "no finding".
  • Reported accuracy varies a lot by pathology and by dataset. Per the library's own published benchmarks: on the NIH test set, AUC ranges from 0.91 (Hernia) and 0.88 (Cardiomegaly) down to 0.68 (Infiltration) and 0.69 (Nodule). On CheXpert it's roughly 0.91–0.94; on PadChest, 0.69–0.97. These are research-split numbers from the library's authors — not a validation of this specific demo deployment.
  • Scores are not calibrated clinical probabilities. A high score is not a diagnosis; a low score does not rule anything out. Both false positives and false negatives are expected and common — this is exactly why the radiologist review step exists.
  • No clinical validation of this application. Nobody has evaluated this specific pipeline (preprocessing + model + priority thresholds + UI) against real clinical outcomes. It is a teaching tool illustrating how such a system could be built, not evidence that it works safely in practice.

Built for an academic case study. Not affiliated with the TorchXRayVision authors or any dataset provider listed above.