Pipeline-Processing-TCGA-Slides-for-MIL

[Overview] | [Walkthrough] | [Acknowledgement] | [Citation]

📚 Recent updates:

• 24/06/01: add new features: allowing users to specify a unified magnification for patching and feature extraction
• 24/05/08: update codes & readme

Overview

This repo provides a complete and detailed tutorial for users who intend to process TCGA Whole-Slide Images (WSIs) for downstream computational tasks, such as WSI classification and survival analysis (basically with multiple-instance learning (MIL) as the learning paradigm).

🔥 Moreover, this repo also provides an improved version of original CLAM. This improvoed version has more practical features:

• allowing users to specify a unified magnification for patching and feature extractiong,
• patch-level image transformations for model rubostness test, e.g. Gaussian bluring and stain variation.
• diverse architectures / networks for patch feature extracting

💡 A quick overview of what convenience this repo could provide for you:

• Step-by-step instructions to download complete TCGA data, including WSIs and corresponding label of interest. This is not presented in CLAM.
• Detailed steps to process WSIs, including WSI segmentation & patching and patch feature extraction. Two notebooks are given to show the details that need to be noticed when using CLAM for WSI preprocessing.
• More alternative pretrained models for patch feature extraction, currently with support of truncated ResNet50 (CLAM), ResNet18 w/ SimCL (DSMIL), CTransPath, CLIP, PLIP, and CONCH.

📝 This repo is developed from PseMix, previously. Now it has been moved out as an individual project for maintaining convenience.

On updating. Stay tuned!

Feel free to post your issue in this repo if you encounter any problems.

Walkthrough

👩‍💻 Here show you how to use this repo:

• S01: Downloading slides from TCGA websites. It shows you, step by step, how to obtain the data you want from TCGA.
• S02: Reorganizing slides at patient-level. It provides the code for organizing slides at patient-level and shows you how to get the useful label data (slide-level) that you would possibly utilize in downstream computational tasks.
• S03: Segmenting and patching slides. It shows you the complete procedures and results of tissue segmentation and patching. Also, the basic knowledge of WSI structure is given.
• S04: Extracting patch features. It shows you the complete procedures and results of patch feature extraction. 6+ pertained models are supported and can be easily used with this step.

📝 Some notes:

• It is strongly recommended to use CONCH in patch feature extraction (S04)
• If you cannot use CONCH due to limited access rights, CTransPath and PLIP are good alternatives (both are free for use). Their usage is provided in tools/scripts.
• The truncated ResNet50 and ResNet18 w/ SimCL are NOT recommended in patch feature extraction.

Acknowledgement

We thank CLAM team for contributing such an efficient and easy-to-use tool for WSI preprocessing, and TCGA team for making WSI data publicly-available to facilitate research.

Citation

If this project helps you more or less, please cite it via

@article{liu10385148,
  author={Liu, Pei and Ji, Luping and Zhang, Xinyu and Ye, Feng},
  journal={IEEE Transactions on Medical Imaging}, 
  title={Pseudo-Bag Mixup Augmentation for Multiple Instance Learning-Based Whole Slide Image Classification}, 
  year={2024},
  volume={43},
  number={5},
  pages={1841-1852},
  doi={10.1109/TMI.2024.3351213}
}

In addition, if you paper use TCGA data and CLAM tools, please also cite them via

@article{lu2021data,
  title={Data-efficient and weakly supervised computational pathology on whole-slide images},
  author={Lu, Ming Y and Williamson, Drew FK and Chen, Tiffany Y and Chen, Richard J and Barbieri, Matteo and Mahmood, Faisal},
  journal={Nature biomedical engineering},
  volume={5},
  number={6},
  pages={555--570},
  year={2021},
  publisher={Nature Publishing Group UK London}
}

@article{kandoth2013mutational,
  title={Mutational landscape and significance across 12 major cancer types},
  author={Kandoth, Cyriac and McLellan, Michael D and Vandin, Fabio and Ye, Kai and Niu, Beifang and Lu, Charles and Xie, Mingchao and Zhang, Qunyuan and McMichael, Joshua F and Wyczalkowski, Matthew A and others},
  journal={Nature},
  volume={502},
  number={7471},
  pages={333--339},
  year={2013},
  publisher={Nature Publishing Group UK London}
}