Nature Communications
Journal Link | Open Access Read Link | Download TISSUE Model | Download ARTIFACT Models | Cite
Abstract: Histological slides contain numerous artifacts that can significantly deterioratethe performance of image analysis algorithms. Here we develop the GrandQCtool for tissue and multi-class artifact segmentation. GrandQC allows for high-precision tissue segmentation (Dice score 0.957) and segmentation of tissuewithout artifacts (Dice score 0.919–0.938 dependent on magnification). Slidesfrom 19 international pathology departments digitized with the most commonscanning systems and from The Cancer Genome Atlas dataset were used toestablish a QC benchmark, analyzing inter-institutional, intra-institutional,temporal, and inter-scanner slide quality variations. GrandQC improves theperformance of downstream image analysis algorithms. We open-source theGrandQC tool, our large manually annotated test dataset, and all QC masks forthe entire TCGA cohort to address the problem of QC in digital/computationalpathology. GrandQC can be used as a tool to monitor sample preparation andscanning quality in pathology departments and help to track and eliminatemajor artifact sources.
GrandQC is an open-source tissue and multi-class artifact segmentation toolbox based on PyTorch.
The main branch works with PyTorch 2+.
Major features
-
GrandQC Design
GrandQC is composed of two components: tissue segmentation and artifact segmentation. Tissue segmentation serves as the foundation for artifact segmentation, offering a more precise tissue target for detecting artifacts
-
Comprehensive Artifact Segmentation
GrandQC offers advanced detection of all common artifacts on segmented whole slide images (WSI), including tissue folding, pen marks, bubbles, edges, black spots, foreign objects, and out-of-focus areas. This level of coverage surpasses that of any existing open-source quality control tools.
-
High efficiency
For tissue segmentation, GrandQC processes a whole slide image (WSI) in an average of 0.4 seconds. Artifact segmentation takes between 27 and 45 seconds per WSI, depending on the model and magnification level.
-
Models with Different Magnifications
To balance efficiency and segmentation quality, GrandQC offers three models trained at different magnifications: 5x, 7x, and 10x. Higher magnifications result in more accurate artifact segmentation but require longer processing times.
-
State of the art
We performed an extensive comparison with existing open-source tools, demonstrating that GrandQC significantly outperforms them across multiple metrics.
-
Building a Benchmark
We analyzed whole slide images (WSI) from various institutions and hospitals across different countries to develop a comprehensive quality benchmark. This benchmark enables each hospital and institution to evaluate their tissue slide preparation processes and make targeted improvements.
-
Scanner Selection Assistant
GrandQC allows for direct comparison of different scanners by analyzing the same batch of tissue slides. This helps in identifying the strengths and weaknesses of each scanner, particularly in relation to out-of-focus issues.
To balance accuracy and inference efficiency, we have trained three different models based on three magnifications (5x, 7x, and 10x). As the magnification decreases, the inference speed increases, but accuracy may be slightly affected.
We consider 7x model to be an optimal choice for most applications.
Please refer to publication for results of formal validation (pixel-wise segmentation accuracy) for 5x, 7x, and 10x GrandQC versions.
Here is an example of one ROI(Region of Interest) image from one Whole Slide Image:
GrandQC TISSUE models segmentt tissue (1) and background (0).
GrandQC ARTIFACT models segment tissue (1), background (7), and five different types of artifacts, including: Tissue folds (2), Dark spots & Foreign objects (3), Pen markings (4), Air Bubble & Slide Edge artifacts (5), and Out-of-focus artifacts (6).
In the final model output, white represents tissue, black represents background, and the five types of artifacts are displayed in distinct colors, as shown in the example below.
conda create -n grandqc python==3.10 -y && conda activate grandqc
git clone https://github.com/cpath-ukk/grandqc.git && cd grandqc
pip install -r requirements.txt
conda install -c conda-forge openslide openslide-python
Links to download model checkpoints:
NOTE: the intended use of the GrandQC is firstly running a tissue segmentation model and then running artifact segmentation model. Although the artifact segmentation models also can detect background, the tissue segmentation model does it much quicker (at 1x magnification), within seconds.
Tissue Segmentation Model: Model
Artifact Segmentation Models: Models
Download the model to the specified folder. The final folder structure should be as follows:
grandqc
├── 01_WSI_inference_OPENSLIDE_QC
│ ├── models
│ │ ├── qc
│ │ │ ├── GrandQC_MPP1.pth
│ │ │ ├── GrandQC_MPP15.pth
│ │ │ └── GrandQC_MPP2.pth
│ │ └── td
│ │ └── Tissue_Detection_MPP10.pth
│ ├── main.py
│ ├── run_art.sh
│ └── ...
├── 02_WSI_inference_OME_TIFF_QC
│ ├── models
│ │ ├── qc
│ │ │ ├── GrandQC_MPP1.pth
│ │ │ ├── GrandQC_MPP15.pth
│ │ │ └── GrandQC_MPP2.pth
│ │ └── td
│ │ └── Tissue_Detection_MPP10.pth
│ ├── main.py
│ ├── run_art.sh
│ └── ...
├── requirements.txt
└── README.md
The default version is 7x (Checkpoint: GrandQC_MPP15.pth)
To use 5x and 10x checkpoints, the main.py script should be modified.
For 5x, use:
MODEL_QC_NAME = 'GrandQC_MPP2.pth'
MPP_MODEL = 2
For 10x, use:
MODEL_QC_NAME = 'GrandQC_MPP1.pth'
MPP_MODEL = 1
Fot this case, you need to use the scripts in 01_WSI_inference_OPENSLIDE_QC:
cd 01_WSI_inference_OPENSLIDE_QC
- Tissue Segmentation
Before running the Tissue-Segmentation script, you need to define the slides path SLIDE_FOLDER and OUTPUT_DIR in run_tis.sh first and then run:
sh run_tis.sh
- Artifacts Segmentation
Similar to the Tissue Segmentation, before running the Artifacts-Segmentation script, you need to define the slides path SLIDE_FOLDER and OUTPUT_DIR in run_art.sh first and then run:
sh run_art.sh
Fot this case, you need to use the scripts in 02_WSI_inference_OME_TIFF_QC:
cd 02_WSI_inference_OME_TIFF_QC
abd then same as the usage above.
If you use GrandQC or benchmark in your research, please consider citing our paper
Weng Z. et al. "GrandQC: a comprehensive solution to quality control problem in digital pathology"
Nature Communications (2024). https://doi.org/10.1038/s41467-024-54769-y
@article{Weng2024,
author = {Weng, Zhilong and Seper, Alexander and Pryalukhin, Alexey and Mairinger, Fabian and Wickenhauser, Claudia and Bauer, Marcus and Glamann, Lennert and Bläker, Hendrik and Lingscheidt, Thomas and Hulla, Wolfgang and Jonigk, Danny and Schallenberg, Simon and Bychkov, Andrey and Fukuoka, Junya and Braun, Martin and Schömig-Markiefka, Birgid and Klein, Sebastian and Thiel, Andreas and Bozek, Katarzyna and Netto, George J. and Quaas, Alexander and Büttner, Reinhard and Tolkach, Yuri},
title = {GrandQC: A comprehensive solution to quality control problem in digital pathology},
journal = {Nature Communications},
year = {2024},
pages = {10685},
doi = {10.1038/s41467-024-54769-y},
url = {https://doi.org/10.1038/s41467-024-54769-y},
issn = {2041-1723}
}
This project is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.
