¶

WSSS4LUAD¶

Update:

*(Feb 10, 2022) The masks of the test set is now available in google drive.

*(November 18) The dataset will be available again after ISICDM 2021.

*(October 24) This challenge will be presented on the 5-th ISICDM 2021 conference (December 17-20, 2021, Guilin, China, http://www.imagecomputing.org/2021/). Top 3 ranking teams will be invited to share their methodologies on the conference.

*(October 12) We are reproducing the model of the top-ranking participants. Final ranking scores will be released soon.

*Top 3 ranking participants will be ask to provide their code via a docker image after the deadline (30 Sept) of this challenge.

Histopathology slide is the gold standard of cancer diagnosis. It delivers massive information on tumor microenvironment (TME), which not only plays a vital role in interpreting tumor initiation and progression, but also influences therapeutic effect and prognosis of cancer patients. The crosstalk between different types of tissues are highly related to tumor progression. Therefore, it is urgent to segment  and differentiate different tissues for further clinical researches.

Lung cancer is the leading cause of cancer death worldwide . In this challenge, we aim to perform tissue semantic segmentation in H&E stained Whole Slide Image (WSI) for lung adenocarcinoma. The current challenge is that obtaining pixel-level annotations of tissue semantic segmentation is extremely difficult and time-consuming. Inspired by Weakly Supervised Semantic Segmentation (WSSS) in computer vision, we decided to provide only image-level annotations to perform tissue semantic segmentation. 

In this challenge, we scanned 67 H&E stained slides from Guangdong Provincial People' Hospital (GDPH) and collected 20 WSIs from The Cancer Genome Atlas (TCGA). Only one WSI was extracted per patient. The goal of this challenge is to use only image-level annotations to achieve pixel-level prediction of three common and meaningful tissue types, tumor epithelial tissue, tumor-associated stroma tissue and normal tissue. Participants are only given image-level annotations (3-digit multi-class labels) for machine learning algorithm training, and pixel-level ground truth for validation and testing.

Statistics¶

Training set: 49 WSIs from GDPH and 14 WSIs from TCGA.
Total 10091 patches were cropped in the training set.
Label Distribution:

• Tumor:  6579
• Stroma:  7076
• Normal:  1832

Validation set: 9 WSIs from GDPH and 3  WSIs from TCGA

Total 40 patches were cropped in the validation set. Including 9 large patches (around 1500~5000*1500~5000 ) and 31 small patches (around 200~500*200~500)

Test set: 9 WSIs from GDPH and 3 WSIs from TCGA.

Total 80 patches were cropped in the test set. Including 14 large patches (around 1500~5000*1500~5000 ) and 66 small patches (around 200~500*200~500)

Evaluation Metrics¶

We use mIOU for model evaluation.

The white background inside the alveoli will be excluded when calculating mIOU. We have provided background mask in the validation and testing data. Participants can directly overlay the white background mask on the prediction results.

Price¶

Top 3 ranking winners will receive 500$, 400$, 300$, respectively. To be eligible for the price, the source code and the model parameters must be submitted, besides the results. 

Citation¶

[1] Han, Chu, et al. "Multi-layer pseudo-supervision for histopathology tissue semantic segmentation using patch-level classification labels." Medical Image Analysis (2022): 102487.

  @article{Han2022Multilayer,
    title = {Multi-Layer Pseudo-Supervision for Histopathology Tissue Semantic Segmentation using Patch-level Classification Labels},
    journal = {Medical Image Analysis},
    pages = {102487},
    year = {2022},
    issn = {1361-8415},
    author = {Chu Han and Jiatai Lin and Jinhai Mai and Yi Wang and Qingling Zhang and Bingchao Zhao and Xin Chen and Xipeng Pan and Zhenwei Shi and Zeyan Xu and Su Yao and Lixu Yan and Huan Lin and Xiaomei Huang and Changhong Liang and Guoqiang Han and Zaiyi Liu}
  }

[2] Han, Chu, et al. (2022).  WSSS4LUAD: Grand Challenge on Weakly-supervised Tissue Semantic Segmentation for Lung Adenocarcinoma. (https://arxiv.org/abs/2204.06455)

  @inproceedings{han2022wsss4luad,
    author = {Han, Chu and Pan, Xipeng and Yan, Lixu and Lin, Huan and Li, Bingbing and Yao, Su and Lv, Shanshan and Shi, Zhenwei and Mai, Jinhai and Lin, Jiatai and Zhao, Bingchao and Xu, Zeyan and Wang, Zhizhen and Wang, Yumeng and Zhang, Yuan and Wang, Huihui and Zhu, Chao and Lin, Chunhui and Mao, Lijian and Wu, Min and Duan, Luwen and Zhu, Jingsong and Hu, Dong and Fang, Zijie and Chen, Yang and Zhang, Yongbing and Li, Yi and Zou, Yiwen and Yu, Yiduo and Li, Xiaomeng and Li, Haiming and Cui, Yanfen and Han, Guoqiang and Xu, Yan and Xu, Jun and Yang, Huihua and Li, Chunming and Liu, Zhenbing and Lu, Cheng and Chen, Xin and Liang, Changhong and Zhang, Qingling and Liu, Zaiyi},
    title = {WSSS4LUAD: Grand Challenge on Weakly-supervised Tissue Semantic Segmentation for Lung Adenocarcinoma},
    publisher = {arXiv},
    year = {2022}
  }