AI-powered pneumonia detection from chest X-rays. 7 deep learning models. Weighted ensemble inference. Built end-to-end in Python.
Live Demo: https://yusra-shah--pneumo-scan-serve.modal.run
A doctor logs in, registers a patient, and uploads a chest X-ray. PneumoScan runs the image through a 7-model ensemble and returns a diagnosis. Every model votes independently. Results include confidence score, severity, subtype, and a Grad-CAM heatmap showing which lung regions influenced the prediction. A PDF report is generated and saved per patient in MongoDB.
All results come from real PyTorch inference on trained weights.
Trained on 15,000 chest X-rays.Weighted soft-vote ensemble. Weights stored in float16 for deployment.
| Model | Role | Val Acc | Test Acc | Weight |
|---|---|---|---|---|
| DenseNet-121 | Anchor — CheXNet architecture | 98.87% | 98.67% | 0.25 |
| InceptionV3 | Multi-scale feature detection | 98.80% | 98.73% | 0.10 |
| AttentionCNN | Grad-CAM explainability | 98.67% | 98.13% | 0.00 |
| EfficientNet-B4 | Efficiency and accuracy balance | 98.40% | 98.20% | 0.20 |
| ViT-B/16 | Global pattern detection | 98.13% | 97.73% | 0.17 |
| ResNet-50 | General-purpose feature extractor | 98.47% | 97.67% | 0.18 |
| MobileNetV3 | Speed model, runs first | 97.13% | 97.80% | 0.10 |
Ensemble accuracy: 98.4%
- Doctor authentication with bcrypt hashed passwords and MongoDB session management
- Patient registration, search, and full medical history tracking
- Real-time ensemble inference with per-model vote breakdown
- Grad-CAM heatmap overlay on the original X-ray
- Severity grading: None / Mild / Moderate / Severe
- Subtype classification: Bacterial or Viral
- PDF report generation via ReportLab
- MongoDB Atlas with transactions, nested documents, and audit logging
This project was built with MongoDB Atlas as a graded Advanced DBMS requirement. The following features are demonstrated:
Transactions — scan results are written atomically. The scan document, model votes, and audit log entry are all committed in a single MongoDB transaction. If any write fails, the entire operation is rolled back.
Nested documents — the scans collection stores the full ensemble result as a nested structure. Each document contains a result subdocument (diagnosis, confidence, severity, subtype) and a model_votes array with per-model predictions embedded directly.
Audit logging — every login, patient registration, and scan submission is recorded to a dedicated audit_log collection with timestamp, doctor ID, action type, and relevant metadata.
Schema design — four collections: doctors, patients, scans, audit_log. All MongoDB logic is isolated to database/db.py. No GUI file imports pymongo directly.
| Layer | Technology |
|---|---|
| GUI | PySide6 6.6.1 |
| Deep learning | PyTorch 2.1.0, timm 0.9.7 |
| Explainability | grad-cam 1.4.8, OpenCV |
| Database | MongoDB Atlas |
| Reports | ReportLab 4.3.0 |
| Deployment | Modal serverless |
| Auth | bcrypt |
15,000 balanced chest X-ray images. 7,500 Normal, 7,500 Pneumonia.
Sources: Kaggle Kermany dataset, NIH ChestX-ray14, Albumentations augmentation pipeline.
Split: 80% train / 10% val / 10% test, stratified.
Python version: 3.12
MongoDB: A MongoDB Atlas URI is required. Create a free cluster at mongodb.com/atlas, then add your connection string to config.yaml:
mongodb_uri: "mongodb+srv://<user>:<password>@<cluster>.mongodb.net/"Model weights: The 7 trained .pth files are included in the submission folder under weights/lung/. Place them there before running. They are not in the GitHub repository due to file size.
git clone https://github.com/Yusra-Shah/Pneumo-Scan.git
cd Pneumo-Scan
python -m venv nemo_env
nemo_env\Scripts\activate
pip install -r requirements.txt
python main.py- @sinyaakumarii
Pneumo/
core/
inference/engine.py Ensemble inference engine
report_generator.py PDF report generation
gradcam.py Grad-CAM heatmap logic
gui/
main_window.py
dashboard.py
scan_panel.py
patients_panel.py
models_panel.py
styles.py
database/ MongoDB Atlas integration (all pymongo logic isolated here)
weights/lung/ Trained .pth files (float16, not included in repo)
outputs/
reports/ Generated PDFs
heatmaps/ Saved Grad-CAM overlays
training/ Training scripts and augmentation pipeline
modal_app.py Modal deployment
main.py Entry point
Built by Yusra Shah and Sinya Kumari — Sukkur IBA University, 2026