README.md

metadata

title: GazeAlign
emoji: 👁️
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
license: mit

GazeAlign

Gaze-supervised medical image classification.

How to use

1. Upload an image — JPG / PNG / BMP / TIFF / WEBP or a DICOM (.dcm).
2. Add fixations (a radiologist-style scanpath) in either of two ways:
   • Click on the image to drop fixation points, or
   • Upload a fixation table (.csv / .xlsx / .xls) and map its ID / X / Y / Time columns — X/Y may be raw pixels or normalised [0,1].
3. Run to get the predicted class with per-class probabilities, plus the learned gaze-conditioned attention mask/overlay.

Model weights

The demo loads the checkpoint declared by the cxr preset in configs/presets.yaml (default checkpoints/best_model_CXR.pth). The weights are not committed to the GitHub repo (too large); add them to this Space — e.g. track checkpoints/*.pth with Git LFS, or download them in a startup step — so checkpoints/best_model_CXR.pth exists at launch.

See the GitHub repository for training code, evaluation scripts, and the paper.