GAA-DETR

MICCAI-2025 Early Accept: Query-Level Alignment for End-to-end Lesion Detection with Human Gaze

Our work introduces a novel detection framework, GAA-DETR (Gaze-Aligned Attention Detection Transformer), which integrates clinical gaze data to enhance lesion detection accuracy. Inspired by how clinicians search for lesions during diagnosis, our method aligns model attention with gaze patterns, enabling detection models to "see" like doctors.

We also contribute the first large-scale Medical Lesion Detection Gaze Dataset, which includes 1,669 open-sourced gaze annotations.

---

Dataset

Gaze Data

Our gaze data were collected using MicEye-v2.0, a tool that records radiologists' eye movements during bounding box annotations. The processed dataset, including gaze heatmaps, is available for download:

Download Gaze Data (Password: gaze)

Image Data

The image data used in our work is sourced from the following public datasets:

• Breast Dataset: Mammograms with annotations for malignant and benign lesions. The datasets used are:

  INbreast: A full-field digital mammographic database. It can be obtained by sending an email request to medicalresearch@inescporto.pt. Please refer to the following publication for more details:

  @article{moreira2012inbreast,
    title={INbreast: toward a full-field digital mammographic database},
    author={Moreira, Igor C and Amaral, Inês and Domingues, Inês and Cardoso, Ana and Cardoso, Maria J and Cardoso, Jaime S},
    journal={Academic radiology},
    volume={19},
    number={2},
    pages={236--248},
    year={2012}
  }

  CBIS-DDSM: A curated breast imaging subset of the Digital Database for Screening Mammography. This dataset is available for download from The Cancer Imaging Archive (TCIA). You can use the code/down_load_DDSM.py script to automate the download process. For more information, please refer to:

  @article{lee2017curated,
    title={A curated mammography data set for use in computer-aided detection and diagnosis research},
    author={Lee, Ryan S and Gimenez, Fernando and Hoogi, Assaf and Miyake, Kristin K and Gorovoy, Marc and Rubin, Daniel L},
    journal={Scientific data},
    volume={4},
    number={1},
    pages={1--9},
    year={2017}
  }

• ComparisonDetector Dataset: Cervical images annotated for lesion detection. This dataset can be downloaded from GitHub - ComparisonDetector. For more details, please refer to:

  @article{liang2021comparison,
    title={Comparison detector for cervical cell/clumps detection in the limited data scenario},
    author={Liang, Y and Tang, Z and Yan, M and Chen, J and Liu, Q and Xiang, Y},
    journal={Neurocomputing},
    volume={437},
    pages={195--205},
    year={2021}
  }

Please ensure to cite the original datasets when using our dataset in your research.

---

Code

Usage

Install

# Clone this repo
git clone https://github.com/YanKong0408/GAA-DETR.git
cd code/GAA-DETR

# Install Pytorch and torchvision
# Tested with 'python=3.7.3,pytorch=1.9.0,cuda=11.1'. Other versions might work.
conda install -c pytorch pytorch torchvision

# Install other required packages
pip install -r requirements.txt

Model

Our code is based on DETR.

Data can be obtained from the json file split resulting from the above step and should be organized into the coco dataset format.Gaze heatmaps should be placed in the same folder as the medical images and named as XXX_heatmap.jpg/png.

Your_Fold/
  └── annotations/
  	├── instances_train2017.json
  	└── instances_val2017.json  

Image_Fold/ 
  └── XXX1.jpg
  └── XXX1_heatmap.jpg
  └── XXX2.jpg
  └── XXX2_heatmap.jpg

Train

python main.py 
    --output_dir  /path/to/your/output/dir 
    --coco_path /path/to/your/data/dir
    --resume path/to/your/pre-train/model.pth

Inference

python mian.py \
    --output_dir  /path/to/your/output/dir \
    -c config/Gaze-DINO/Gaze_DINO_swin.py \
    --options batch_size=1 \
    --coco_path /path/to/your/data/dir
    --resume path/to/your/pre-train/model.pth
    --eval