AI-assisted pediatric pneumonia X-Ray screening, but make it human-approved.
PneumoX helps screen pediatric chest X-rays for pneumonia using deep learning, then hands the final call to an actual medical professional, because AI should assist, not replace. Human-in-the-loop like always.
Built by Team 9 HerAI 💜 Hustler: Badariandini F, Salma NS · Hipster: Salsa DRP · Hacker : Riana DL
Reading chest X-rays for pneumonia is slow and heavily dependent on how experienced the radiologist reading it is. When you've got a stack of scans and limited specialists, screening turns into a bottleneck — and kids who need urgent care can end up waiting longer than they should.
Core question: how do we make pneumonia screening faster and more consistent without cutting the expert out of the loop?
- Upload a chest X-ray
- AI screens it — DenseNet121 (transfer learning + fine-tuning) flags pneumonia vs. normal with a confidence score
- A human reviews it — radiologist / medical staff confirms, overrides, or flags the result
- Feedback gets logged — every review becomes training data for the next model iteration
No auto-diagnosis. No black box. Just a faster first pass that's always double-checked.
| Who | What they get |
|---|---|
| Medical staff / radiologists | Reviews & validates predictions, gives feedback |
| MCU / clinic teams | Faster screening, helps prioritize urgent cases |
| Hospitals | Monitor system performance & screening quality |
| Data analysts / devs | Use reviewer feedback to keep improving the model |
(currently scoped to pediatric patients, ages 1–5)
- 91.03% accuracy, AUC 0.96
- Model: DenseNet121 transfer learning, gradual fine-tuning, with class imbalance handling
- Dataset: pediatric chest X-rays via Kaggle (originally from Guangzhou Women and Children's Medical Center)
Model / AI: Python · TensorFlow · Keras · NumPy · scikit-learn · OpenCV · matplotlib · Kaggle
Prototype: Google Colab
Web app: Python · Streamlit · VS Code · GitHub
- Trained on pediatric data only — not validated for adults
- 91.03% is solid, not perfect — some cases can still be misclassified, so it's a screening aid, not a final diagnosis
- Login is currently a demo/hardcoded setup, not production-grade auth (no real user DB, hashing, or session management yet)
This is an assistive tool. The clinical call always stays with a qualified professional!!
- Cross-dataset validation (adult patients, other hospitals, etc etc) for broader generalization
- Mobile app version for remote clinics/facilities
- Role-based login for reviewers/radiologists
- Grad-CAM heatmaps so you can actually see what the model's looking at (explainability ftw)