PatchKGH (pKGH)

♻️Garbage in, garbage out - classic saying for data curation in Computer Vision
The KGH dataset is a private dataset for colorectal polyps classification, consisting of 1037 Whole Slide Images (WSIs) encompassing both healthy colon tissue and colon polyps. The pathological WSIs have been annotated at the region of interest (ROI) level. In order to train deep learning models for colorectal polyp classification applied to this dataset, we need to extrat patches from the WSIs. To achieve a good representation of this dataset, we need to perform strong data curation and extract patches, or tiles, which are cleaned and do not present with any artefacts.

📂 Presentation of the dataset and its annotations

KGH dataset is a private dataset of colorectal polyps. The WSIs come with 4 resolution levels: 20X, 5X, 1.25X and 0.3125X. The figure below shows the different downsampled levels as well as a ROI annotation. The resolution at 20X is of 0.4 mpp (microns per pixels). The tissue thickness in these slides is of 5 microns.

The polyps studied in this dataset are Sessile Serrated Lesions (SSL), Hyperplastic Polyps (HP), Tubulovillous Adenoma (TVA) and Tubular Adenoma (TA). This dataset is also presenting normal, or histological, colon tissues. One WSI can present multiple ROIs.

Studied colorectal polyps

Four different types of colorectal polyps are studied in this dataset:

• Sessile Serrated Lesions (SSL): SSLs, a subtype of serrated polyps, constitute at least 20% of all serrated polyps. They are considered precancerous and are characterized by a flat or slightly elevated shape, predominantly found in the cecum and ascending colon. Distinguishing SSLs from hyperplastic polyps may pose challenges, although certain distortions can be recognized by pathologists;
• Hyperplastic Polyps (HP): These noncancerous growths carry a low risk of malignant transformation and are characterized by an overgrowth of normal cells in the mucosal lining of the colon or distal colon;
• Tubular Adenoma (TA): These are usually small and benign polyps, prevalent in more than 80% of cases. While they are considered precancerous, less than 10% of them have the potential to progress into cancer;
• Tubulovillous Adenomas (TVA): TVAs are a subtype of colonic adenomas exhibiting a combination of tubular and villous features. Considered precancerous, they have the potential to transform into malignant structures. The figure below highlights the four different polyps and a normal patch, under a Field of View (FoV) of 800x800 microns

Regions of interest (ROI) annotations

A pathologist annotated some regions of interest in the slide. It is important to highlight that regions outside ROIs cannot be considered as normal tissue regions nor as pathological regions. They contain normal tissue, distorted tissue and pathological tissue which have not been annotated. The figure below shows 5 WSIs from the 5 different classes with ROI annotations and patches extracted from a ROI when applicable.

📣 Use the dataset in your own framework

The dataset needs to be stored such as:

root/
├── train/
│ ├── Normal/
│ │ ├── NX-id-coords.png
│ │ └── NX-id-coords.jpg
| ├── CP_HP/
│ │ │ ├── ROI
| | │ │ ├── HPX-id-roi_id-coords.png
│ | | │ └── HPX-id-roi_id-coords.jpg
│ │ | ├── nonROI
| | │ │ ├── HPX-id-coords.png
│ | | │ └── HPX-id-coords.jpg
│ ├── ...
├── test/
│ ├── Normal/
│ │ ├── NX-id-coords.png
│ │ └── NX-id-coords.jpg
| ├── CP_HP/
│ │ │ ├── ROI
| | │ │ ├── HPX-id-roi_id-coords.png
│ | | │ └── HPX-id-roi_id-coords.jpg
│ │ | ├── nonROI
| | │ │ ├── HPX-id-coords.png
│ | | │ └── HPX-id-coords.jpg
│ ├── ...

You can then call pkgh class

dataset = pkgh(root_dir, split = 'train', ROI = False, balance = False, transform = transform)

if you want to use all images from ROI and nonROI, from the train set

🤺 Patch extraction challenges

When having a closer look at our dataset (Figure below) , we remark than some slides have been annotated with a marker and contains either contours or handwritten annotations.

The daset presents marker annotations
The application of two state-of-the-art methods revealed limitations in our dataset. The handwritten annotations and inherent slide noise presented challenges for the various segmentation strategies employed. To address these issues, we developed a novel data processing approach specifically designed to mitigate the impact of noisy patches. The custom CNN to detect noisy patches is made of 3 convolutional layers followed by ReLU activation functions. It was trained for 20 epochs on a dataset of 8686 samples made of noisy patches (4475) and tissue patches (4211) extracted at different Fields of view

Custom CNN

🔍 Patch extraction code

You will find the code for patch extraction in this folder.
To run an extraction, you can precise --datafolder where the WSIs are stored, the Field of View (--fov) of extraction, the level of extraction (--level between 0 and 2), the output folder for your new dataset (--output) and the --size under which the patches will be saved:

python3 extraction_huron.py --size 256 --fov 800 --level 1 --method overlap --output ./kgh_800

A basic --threshold 50 for patch/ROI overlap is fixed --method overlap is chosen (default). By default, the background and noisy patches are not saved, you can activate the saving using --save True and these patches will be saved in specific background folders.
A script is given as an example.

💻 Conda environment for patch extraction

To work in the environment supporting TIAToolBox and PyTorch (2.2.1), you can download the conda configuration file and follow the instructions below:

conda env create -f env.yml

Activate the environment:

conda activate tia-env

📬 Contact information

For any questions, you can create an issue or send your question at my courriel adress