🫁 Chest X-ray Pneumonia Detector
VGG16 transfer learning + Grad-CAM explanation. Runs 100% in your browser: your image is never uploaded.
1 · Chest X-ray
Drop an X-ray here or click to choose
JPG or PNG · stays on your device
Or try an example from the Kaggle test set:
Loading model…
2 · Result
Upload an X-ray or pick an example to see the prediction.
About the model
VGG16 (ImageNet weights) fine-tuned on the Kaggle Chest X-Ray Images (Pneumonia) dataset, rebuilt from the Knowledge Doctor tutorial with a proper validation split, class weighting, fine-tuning and honest evaluation. Inference runs here via ONNX Runtime Web (WebAssembly).
| Official test set (624 images, threshold 0.5) | |
|---|---|
| Accuracy | 89.9% |
| Sensitivity (pneumonia caught) | 97.9% (382 of 390) |
| Specificity (normal cleared) | 76.5% (179 of 234) |
| Precision · F1 · ROC-AUC | 87.4% · 0.924 · 0.965 |
It is tuned to miss few pneumonia cases, so it raises false alarms on roughly 1 in 4 normal X-rays.
Confusion matrix and ROC curve
Limitations
- Pediatric patients only (ages 1–5, one hospital in Guangzhou).
- Not validated on adults, other scanners, or other diseases; it will give an answer for any image, even a non-X-ray.
- It cannot tell bacterial from viral pneumonia.
- Grad-CAM is a visual aid, not proof of clinical reasoning.