yankon0408/GAA-DETR-breast

# Breast Lesion Detection Gaze Dataset (MICCAI 2025 Spotlight)

This repository contains the Breast subset of the **first large-scale Medical Lesion Detection Gaze Dataset**, as presented in our MICCAI 2025 Spotlight paper: *"Query-Level Alignment for End-to-End Lesion Detection with Human Gaze"*.

## 1. Dataset Overview

This dataset provides a unique combination of high-resolution mammography images, expert radiologist eye-tracking (gaze) data, and gold-standard lesion annotations. It is designed to facilitate research in gaze-assisted medical AI and human-computer interaction in clinical diagnostics.

### Directory Structure

* `image/`: Pre-processed mammogram images (cropped to remove uninformative black backgrounds).
* `gaze_data/`: Raw gaze coordinates and scaling factors.
* `annotations/`: Lesion bounding boxes in COCO format.
* `code/`: Demonstration scripts (e.g., `demo.ipynb`) for data processing.

---

## 2. Data Specifications

### Images (`/image`)
The mammograms have been pre-processed to remove uninformative black backgrounds via cropping.

### Gaze Data (`/gaze_data`)
The original eye-tracking data is provided in the following format:
`[[x_coordinate, y_coordinate], scaling_factor]`
* **x, y**: Coordinates relative to the scaled/resized image.
* **scaling_factor**: The ratio used to resize the original image to the display size during the gaze collection process.

### Annotations (`/annotations`)
Annotations follow the standard **MS COCO** detection dataset format.
* **Bounding Box Format**: `[x1, y1, width, height]`
* **Important Note on Labels**: All malignant and benign lesion classifications have been pathologically verified. Regions with pathology confirmation are strictly annotated. However, please note that for other regions, annotations are based on morphological appearance on the mammogram; while most visible lesions are marked, some minor findings may be unannotated.

---

## 3. Getting Started

A demonstration notebook is provided in `code/demo.ipynb` to show how to process the data. It includes steps to:
1. Load the images and corresponding gaze data.
2. Convert raw `gaze_data` points into attention heatmaps.
3. Overlay gaze heatmaps and bounding box annotations onto the mammograms.

---

## 4. Citation and Acknowledgments

### Primary Reference
If you use this dataset or the associated code in your research, please cite our MICCAI 2025 paper:

```bibtex
@inproceedings{kong2025query,
title={Query-Level Alignment for End-to-End Lesion Detection with Human Gaze},
author={Kong, Yan and Peng, Zhixiang and Yin, Yuan and Li, Yonghao and Cai, Jiangdong and Wang, Sheng and Wang, Qian and Fang, Yuqi and Shan, Caifeng},
booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
pages={496--506},
year={2025},
organization={Springer}
}
```

### Data Sources
The image data in this subset is sourced from the INbreast and CBIS-DDSM public databases. Please ensure you also cite the following original works when using this dataset:

#### INbreast:
```bibtex
@article{moreira2012inbreast,
title={INbreast: toward a full-field digital mammographic database},
author={Moreira, Igor C and Amaral, In{\^e}s and Domingues, In{\^e}s 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:
```bibtex
@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}
}
```

**Contact:** *email to* [[e-mail]]() or *issue in* [https://github.com/YanKong0408/GAA-DETR](https://github.com/YanKong0408/GAA-DETR).

# 乳腺病变检测注视数据集（MICCAI 2025 Spotlight论文）

本仓库包含首个大规模医学病变检测注视数据集的乳腺子数据集，相关成果发表于我们的MICCAI 2025亮点论文《结合人类注视的端到端病变检测查询级对齐》。

## 1. 数据集概览

本数据集创新性地整合了高分辨率乳腺钼靶图像、放射科专家眼动追踪（注视）数据以及金标准病变标注，旨在推动注视辅助医疗人工智能以及临床诊断中人机交互领域的研究。

### 目录结构
* `image/`：经过预处理的乳腺钼靶图像（已裁剪去除无意义的黑色背景）。
* `gaze_data/`：原始注视坐标与缩放因子。
* `annotations/`：采用COCO格式的病变边界框标注。
* `code/`：用于数据处理的演示脚本（例如`demo.ipynb`）。

---

## 2. 数据说明

### 图像数据（`/image`目录）
乳腺钼靶图像已通过裁剪预处理去除无意义的黑色背景。

### 注视数据（`/gaze_data`目录）
原始眼动追踪数据采用如下格式存储：`[[x_coordinate, y_coordinate], scaling_factor]`
* **x, y**：对应缩放/调整尺寸后的图像坐标。
* **scaling_factor**：眼动数据采集过程中，将原始图像调整至显示尺寸所用的比例系数。

### 标注数据（`/annotations`目录）
标注遵循标准微软COCO（Microsoft Common Objects in Context, MS COCO）目标检测数据集格式。
* **边界框格式**：`[x1, y1, width, height]`
* **标注标签重要说明**：所有恶性与良性病变分类均经过病理学验证，仅对经病理学确认的区域进行严格标注。需注意的是，其余区域的标注仅基于乳腺钼靶图像的形态学外观；尽管绝大多数可见病变均已标记，但部分微小异常可能未被标注。

---

## 3. 快速入门

仓库中`code/demo.ipynb`提供了演示脚本，展示如何处理本数据集，具体包含以下步骤：
1. 加载图像与对应注视数据。
2. 将原始`gaze_data`点转换为注意力热力图。
3. 将注视热力图与边界框标注叠加至乳腺钼靶图像上。

---

## 4. 引用与致谢

### 主要引用
若您在研究中使用本数据集或相关代码，请引用我们的MICCAI 2025论文：

bibtex
@inproceedings{kong2025query,
title={Query-Level Alignment for End-to-End Lesion Detection with Human Gaze},
author={Kong, Yan and Peng, Zhixiang and Yin, Yuan and Li, Yonghao and Cai, Jiangdong and Wang, Sheng and Wang, Qian and Fang, Yuqi and Shan, Caifeng},
booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
pages={496--506},
year={2025},
organization={Springer}
}

### 数据来源
本子集的图像数据源自INbreast与CBIS-DDSM公开数据库。使用本数据集时，请同时引用下列原始文献：

#### INbreast数据集：
bibtex
@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数据集：
bibtex
@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}
}

**联系方式**：可发送邮件至[[e-mail]]()或在[https://github.com/YanKong0408/GAA-DETR](https://github.com/YanKong0408/GAA-DETR)提交Issue。