Rexosphere Pneumonia AI

Explainable Pneumonia Detection Using Two-Phase Fine-Tuned DenseNet121 for BioFusion Hackathon 2026.

Web App

The Stitch-generated visual direction has been integrated as a plain HTML/CSS/JS frontend:

• index.html defines the app shell, sidebar navigation, topbar, and all product screens.
• assets/styles.css contains the unified responsive light/dark theme.
• assets/app.js handles navigation, upload preview, diagnosis API calls, result rendering, Grad-CAM controls, and frontend configuration.
• backend/app.py serves the frontend and exposes the API.
• backend/inference.py loads the trained DenseNet121 model, runs inference, and generates Grad-CAM images.
• config/model_metrics.json stores exported notebook performance metrics used by /api/performance.

Screens

• Upload & Diagnose
• Diagnosis Result
• Grad-CAM Explainability
• Model Performance
• Training Method
• Dataset
• Robustness & Quality
• Research
• Settings/About

Backend integration

The frontend is connected to:

POST /api/diagnose
GET /api/performance
GET /api/health

Diagnosis requests are sent as multipart/form-data with one field:

image

Supported response fields are:

{
  "predictionClass": "Normal or Pneumonia",
  "confidence": 0.94,
  "threshold": 0.25,
  "inferenceTimeMs": 123,
  "recommendation": "Clinical review recommended.",
  "gradcamOverlayUrl": "/path/to/overlay.png",
  "gradcamHeatmapUrl": "/path/to/heatmap.png"
}

No backend predictions are faked in the UI. If the trained model file is missing, /api/diagnose returns 503 and the upload screen shows a clear backend error.

Model file

Export the trained notebook model and place one of these files in models/:

models/Pneumonia_DenseNet121_Final.pth
models/best_phase2.pth

The backend expects a PyTorch DenseNet121 state dict matching:

torchvision.models.densenet121(weights=IMAGENET1K_V1)
classifier replaced with Linear(1024, 2)
classes: Normal, Pneumonia

Run locally

cd /home/ifaz/Coding/BioFusion
python -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
.venv/bin/uvicorn backend.app:app --host 127.0.0.1 --port 8000

Then open:

http://127.0.0.1:8000

If port 8000 is already in use:

.venv/bin/uvicorn backend.app:app --host 127.0.0.1 --port 8010

Then open:

http://127.0.0.1:8010

How to use the app

1. Start the FastAPI app with Uvicorn.
2. Open the app URL in a browser.
3. Go to Upload & Diagnose.
4. Select a chest X-ray image.
5. Click Run Diagnosis.
6. Review Diagnosis Result for class, confidence, threshold, inference time, recommendation, and disclaimer.
7. Open Grad-CAM to inspect the original image, overlay, or heatmap. Use the opacity slider for overlay review.
8. Open Model Performance, Dataset, Training Method, Robustness & Quality, and Research for project documentation.

Diagnosis and Grad-CAM require a real trained .pth file in models/. Without it, the rest of the app still runs and diagnosis requests fail safely with an explicit model-missing message.