MedVision-V0 Demo

Live demo of YongchengYAO/MedVision-V0-7B, a Qwen2.5-VL-7B fine-tune for quantitative medical image analysis.

Examples — pick a curated benchmark case; its exact prompt is prefilled and editable. Hit Run to watch the model reason live, with the prediction (orange) and ground truth (green) overlaid and full metrics (IoU / MAE / MRE) computed against the stored GT.

Upload your own — choose a task (detection, T/L, A/D distance, A/D angle), fill the info panel to assemble an editable prompt, upload a 2D image with its pixel spacing, and Run.

Images are fed at 512×512; the model perceives 504×504 internally, and the prompt states 504×504 with scaledPS-rescaled spacing so image_size × pixel_size equals the true physical extent.

Figure axis ranges — the demo's overlay figures are rendered from the 512×512 gallery images, so the array-index axis ticks span ~0–512. The published project-page case figures are rendered from the full native-resolution NIfTI slice (e.g. ~0–1800 for a tall cephalogram), so their tick ranges differ. This is a data/pipeline resolution difference, not a rendering discrepancy — the imshow + rotation + aspect setup is identical.

⚠️ Research / education only — NOT for clinical decision-making.

Local dev

pip install -r requirements.txt
python -m pytest -q     # GPU-free unit tests
python app.py           # launch UI (model load needs a GPU)