Create README.md
Browse files# Dataset for Singpath-VL
- **Paper:** [Singpath-VL](https://arxiv.org/pdf/2602.09523)
## Singpath-CytoText
### Dataset Description
Singpath-CytoText is a cervical cytology image dataset with structured pathological descriptions and annotations. Each image contains a bounding box highlighting anchor cell and provides a **structured description** (e.g., nuclear size, chromatin pattern, nuclear membrane status, etc.), a natural language **caption**, and a final cytological **label** (e.g., ASC-H). This dataset can be used for multimodal learning, vision-language pre-training, and medical image report generation.
### Supported Tasks and Leaderboards
- **Image Classification**: Predict the Bethesda category (e.g., ASC-H, NILM) from a cell image.
- **Vision-Language Modeling**: Generate structured pathological descriptions or reports from images.
- **Cross-Modal Retrieval**: Retrieve similar cell images using natural language queries.
### Languages
All text fields (structured description, caption, label) are in Chinese.
## Dataset Structure
### Data Instances
A typical data point from `Singpath-CytoText.json`:
```json
{
"image_name": "dEPhZAhdwBFttBzk_2363_130_280_244.jpg",
"structured description": {
"核大小": "增大",
"核染色质": "不均匀，颗粒粗，聚集明显",
"核数量": "单核",
"核质比": "增高",
"核膜": "不规则",
"核仁": "未见",
"胞质状态": "胞浆量少，着色浅，无特殊结构",
"异常病理指征": "细胞核异型性：存在；角化异常：未见，不存在挖空细胞",
"细胞空间构型": "未见聚集成团"
},
"caption": "框中细胞核增大，核染色质不均匀、深染，颗粒较粗且聚集明显；核膜轮廓不规则，核质比增高，为单核结构，未见明显核仁。胞浆量较少，着色较浅，无空泡化或其他特殊结构。细胞呈分散分布，未见聚集成团。。不存在挖空细胞。",
"label": "ASC-H"
}
```
## CytoCell-Bench
### Dataset Description
CytoCell-Bench is a benchmark dataset for cervical cytology image classification. Each image corresponds to a cell region and provides a classification label according to the Bethesda system (e.g., NILM, ASC-US, LSIL, HSIL). This dataset is intended to serve as a standardized testbed for cytology image classification algorithms.
### Supported Tasks and Leaderboards
- **Image Classification**: Predict the Bethesda category of a cell image. This task can serve as a benchmark in medical image classification.
### Languages
Labels are in English abbreviations (e.g., NILM) following the international Bethesda reporting system.
## Dataset Structure
### Data Instances
A typical data point from `CytoCell-Bench.json`:
```json
{
"image_name": "AIMS-419_19096.0_44486.0_crop_000.jpg",
"label": "NILM"
}
```
## 📖 Citation
If you find the dataset useful in your research, please consider citing:
```
@article {qiu2026singpath,
title={Singpath-VL Technical Report},
author={Qiu, Zhen and Xiao, Kaiwen and Lu, Zhengwei and Liu, Xiangyu and Zhao, Lei and Zhang, Hao},
journal={arXiv preprint arXiv:2602.09523},
year={2026}
}
```