medscan — Hackathon-ready Anomaly Detector

Fast, dirty, and useful: medscan is a lightweight Gradio web UI that loads best-effort PyTorch checkpoints for CT / X-ray / Ultrasound and shows live anomaly detection with a left-side Anomaly Log and a masked preview.

This README is written in a hackathon style — short, actionable, and focused on getting you running fast.

What you'll find

• backend/app.py — Gradio app & robust model loader
• backend/repair_checkpoints.py — helper to remap/checkpoint-fix
• backend/models/ — trained checkpoints (if present). Fixed copies use *_fixed.pth
• result-images/ — demo output + example masked previews (look here first)
• userImage/ — example baseline images used by the masking routine

Quick demo (2-minute setup)

Open a PowerShell terminal in the repository root and run:

# create virtualenv (Windows)
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r backend\requirements.txt
# launch the web UI
python backend\app.py

After launch, open the printed URL (http://localhost:) and upload an image. The left column shows the Anomaly Log and a masked preview (from userImage/ baseline if present). The result-images/ folder contains example outputs.

Notes about models

• Model files (.pth/.pt) are expected in backend/models/. The loader prefers files named like ct_model.pth, cnn_chestxray.pth, ultrasound_model.pth.
• If checkpoint keys don't match the app's simple architecture, run python backend/repair_checkpoints.py to create <name>_fixed.pth copies — the app prefers _fixed.pth when present.
• Large model files should be stored with Git LFS or hosted outside the repo (recommended).

How masking works (quick)

• The app compares the uploaded image to a baseline (from userImage/) using a simple per-pixel abs-diff on a 128x128 tensor.
• Differences above a threshold produce a mask that overlays red where changes occur — example previews are in result-images/.

Troubleshooting

• If you see Git warnings about line endings or the virtualenv being tracked, ensure .gitignore includes .venv/ and venv/ and then run:

# stop tracking an already-tracked venv without deleting files
git rm -r --cached .venv
git add .gitignore
git commit -m "Ignore virtualenv"

• If backend/models/ isn't uploaded because *.pth is ignored, either whitelist the folder in .gitignore or use git lfs to track large files.

Screenshots / examples

Two example images are included in result-images/ for quick reference:

• result-images/example_comparison.png — left: baseline, center: new scan, right: detected changes mask (thresholded diff)
• result-images/ui_screenshot.png — screenshot of the running Gradio UI showing modality, prediction, and confidence

Detection Output

Baseline, New Scan, and Detected Changes:

Web UI Screenshot

!