zeyugao/GastricSemanticSegmentation

Gastric Cancer Tissue Segmentation Dataset

License: Apache-2.0

---

Overview

This dataset is designed for tissue segmentation in gastric cancer cases. It consists of 100 Regions of Interest (ROIs) extracted from Whole Slide Images (WSIs) of 100 gastric cancer cases.

Tissue Types

Six tissue types are annotated:

1. Tumor
2. Lymphoid stroma
3. Desmoplastic stroma
4. Smooth muscle
5. Necrosis
6. Others

---

Data Source

The original WSIs are sourced from the TCGA (The Cancer Genome Atlas) database.

• Mean size of ROIs: 4655 × 5276 pixels

---

Annotation Process

• The annotated ROIs achieved a 78% one-time acceptance rate.
• The remaining annotations were accepted after one revision.
• Pathologists performed minor corrections on 8.4% of all pixels in total.

---

Data Organization

The dataset includes:

1. ROIs (image patches):
   • Saved as .png files under the corresponding folders.
2. Annotations:
   • Each ROI's annotation is saved as a .txt file under the corresponding folders.
   • The annotation is a pixel-wise matrix with the following values:
     • 1: Tumor
     • 2: Lymphoid stroma
     • 3: Desmoplastic stroma
     • 4: Smooth muscle
     • 5: Necrosis
     • 6: Others
     • -1: Equal to 6 (others)

---

Usage and Restrictions

• This dataset is for research purposes only.
• Commercial use is strictly prohibited.

If you use this dataset in your research, you must cite the following publication:

@article{gao2022unsupervised,
  title={Unsupervised representation learning for tissue segmentation in histopathological images: From global to local contrast},
  author={Gao, Zeyu and Jia, Chang and Li, Yang and Zhang, Xianli and Hong, Bangyang and Wu, Jialun and Gong, Tieliang and Wang, Chunbao and Meng, Deyu and Zheng, Yefeng and others},
  journal={IEEE Transactions on Medical Imaging},
  volume={41},
  number={12},
  pages={3611--3623},
  year={2022},
  publisher={IEEE}
}