GazeSearch includes chest X-ray images derived from MIMIC-CXR, which is distributed under the PhysioNet Credentialed Health Data Use Agreement. To be granted access to this dataset you must already be credentialed for MIMIC-CXR on PhysioNet and agree to use the data only for the purposes permitted by that agreement.
Requests are reviewed manually. Please provide accurate information — false declarations will result in revoked access.

By requesting access I confirm that I will use GazeSearch only for research that complies with the PhysioNet Credentialed Health Data Use Agreement for MIMIC-CXR and will not redistribute the images.

GazeSearch: Radiology Findings Search Benchmark (WACV 2025)

Access is gated. This dataset contains images derived from MIMIC-CXR and is only shared with users credentialed under the PhysioNet Credentialed Health Data Use Agreement. Click Request access at the top of the page and confirm you hold an active MIMIC-CXR credentialed-access account; requests without proof of authorization will be denied.

GazeSearch is a curated visual search dataset for evaluating search algorithms on radiology findings. It is built from medical eye-tracking data (REFLACX and EGD) paired with chest X-ray images from MIMIC-CXR, and captures how radiologists visually search for specific findings.

Paper: GazeSearch: Radiology Findings Search Benchmark (WACV 2025). Original repository: https://github.com/uark-aicv/GazeSearch

gazesearch/
├── annotations/
│   └── finding_visual_search_coco_format_train_test_filtered_max_6_split_train_valid_test_2024-07-22.json
├── images/                      # 3,577 chest X-ray JPGs (from MIMIC-CXR)
└── checkpoints/
    ├── ckp_29999.pt             # ChestSearch model checkpoint (30k iterations)
    ├── M2F_R50.pkl              # Mask2Former ResNet-50 backbone weights
    └── M2F_R50_MSDeformAttnPixelDecoder.pkl  # MSD pixel decoder weights

Annotation format

The JSON file is a list of scanpath samples (4,875 entries over 2,328 unique images). Each entry has:

Field	Description
`name`	Image filename in `images/`
`subject`	Radiologist / subject ID
`task`	Target finding being searched (e.g. `lung opacity`, `cardiomegaly`, ...)
`condition`	Presence of the finding (`present`)
`bbox`	Ground-truth bounding box `[x, y, w, h]` of the finding
`X`, `Y`	Fixation coordinates (in the 224×224 patch space)
`T`	Fixation durations (seconds)
`length`	Number of fixations (max 6; add a center fixation to get length-7 paper setup)
`fixOnTarget`	Whether the scanpath ends on the target region
`correct`	Whether the radiologist's decision was correct
`split`	One of `train` (3,870) / `valid` (517) / `test` (488)

Tasks / findings

lung opacity, support devices, cardiomegaly, pleural effusion, atelectasis, edema, consolidation, enlarged cardiomediastinum, lung lesion, fracture, pneumonia, pneumothorax, pleural other.

Checkpoints

checkpoints/ckp_29999.pt is the trained ChestSearch scanpath-prediction baseline from the paper. M2F_R50.pkl and M2F_R50_MSDeformAttnPixelDecoder.pkl are the Mask2Former ResNet-50 backbone / pixel-decoder weights used to initialize training.

See the original repo for training/inference code and the src/demo_medical.ipynb demo notebook.

Citation

@article{GazeSearch2023,
    title={GazeSearch: Radiology Findings Search Benchmark},
    author={Trong Thang Pham and Tien-Phat Nguyen and Yuki Ikebe and Akash Awasthi and Zhigang Deng and Carol C. Wu and Hien Nguyen and Ngan Le},
    journal={IEEE Winter Conference on Applications of Computer Vision (WACV)},
    year={2025}
}

License

MIT License — Copyright (c) 2024 AICV@University of Arkansas. Note that the underlying MIMIC-CXR images are subject to the PhysioNet Credentialed Health Data License; only users credentialed for MIMIC-CXR should use this dataset.

Downloads last month: 12

Total file size:

6.82 GB

Access to GazeSearch requires MIMIC-CXR credentials