PathoEye

PathoEye: a deep learning framework for the whole-slide image analysis of skin tissue

Table of Contents

• Background
• Installation
• Tutorial
• Maintainers
• Contributing
• License

Background

The whole-slide images (WSI) examination of skin biopsy is the golden standard for pathological diagnosis of most skin diseases. While most studies focus on the classification tasks, an interpretable computational framework is lacking for WSI analysis. To this end, we developed PathoEye for WSI analysis in dermatology, which integrates epidermis-guided sampling, deep learning and radiomics. PathoEye performed better than the existing state-of-the-art methods in discriminating the young and aged skin tissues. Moreover, PathoEye performs comparably with the existing methods in the binary classification of healthy and diseased skin while performing better in multi-classification tasks.

Installation

Clone PathoEye source code to your server.

cd PATHOEYE_INSTALL_DIR
git clone https://github.com/lysovosyl/PathoEye.git

Make sure you have installed all the packages listed in requirements.txt before running PathoEye.

conda create -n PathoEye python==3.8
pip install -r requirements.txt
conda activate PathoEye

The detailed dependencies are listed as follows:

SimpleITK==2.2.1
torch==2.0.0
torchmetrics==0.9.0
torchvision==0.15.1
opencv-python==4.7.0.72
opencv-python-headless==4.7.0.72
scikit-image==0.20.0
scikit-learn==1.2.2
openslide-python==1.2.0
matplotlib==3.7.0
matplotlib-inline==0.1.6
pyradiomics==3.0.1
six==1.16.0
numpy==1.23.5
pandas==2.0.1
pypickle==1.1.0

Tutorial

Testing dataset

The example dataset and the pretrained weights for testing PathoEye can be downloaded from figShare (https://doi.org/10.6084/m9.figshare.30566108.v1). The full dataset for the young and old skin analysis are free available at GTEx project.

Module1: epidermis extraction

Epidermis extraction

The following example assumes that the whole slide images (WSIs) data is organized in well known standard formats (such as .svs, .ndpi, .tiff etc.) and stored at a folder named DATA_DIRECTORY.

    DATA_DIRECTORY/
        ├──slide_1.svs
        ├──slide_1.svs
        ├──slide_2.svs
        ├──slide_3.svs
        └── ...

You can run epidermis_extract.py to extract epidermis of each WSI in the DATA_DIRECTORY. In this stage, you can obtain the level 1 images.

python epidermis_extract.py -data_dir /DATA_DIRECTORY -save_path ./EPIDERMIS_SAVEPATH

Epidermis thickness and variance of rete ridge length calculation

You can obtain the thickness and the variance of the rete ridge for each image in the EPIDERMIS_SAVEPATH using epidermis_extract.py and the outputs will be stored in the save_dir directory.

python thickness.py -data_dir /EPIDERMIS_SAVEPATH -save_path ./RESULT_SAVEPATH

Module2: epidermis-guided patch sampling

In this part, you can generate the level 2 and level 3 images by create_patches.py with WSIs as input.
The organized WSIs must be stored under a folder named TRAIN_DIRECTORY(train dataset) , VAL_DIRECTORY (validation dataset) and TSET_DIRECTORY (test dataset).

    TRAIN_DIRECTORY/
        ├── class_1
            ├──slide_1.svs
            ├──slide_2.svs
            ├──slide_3.svs
            └── ...
        ├── class_2
            ├──slide_1.svs
            ├──slide_2.svs
            ├──slide_3.svs
            └── ...
        └── ...

    VAL_DIRECTORY/
        ├── class_1
            ├──slide_1.svs
            ├──slide_2.svs
            ├──slide_3.svs
            └── ...
        ├── class_2
            ├──slide_1.svs
            ├──slide_2.svs
            ├──slide_3.svs
            └── ...
        └── ...
        
    TEST_DIRECTORY/
        ├── class_1
            ├──slide_1.svs
            ├──slide_2.svs
            ├──slide_3.svs
            └── ...
        ├── class_2
            ├──slide_1.svs
            ├──slide_2.svs
            ├──slide_3.svs
            └── ...
        └── ...

Then, you can generate patches for the training dataset, validation dataset and testing dataset.

python create_patches.py -input_path /TRAIN_DIRECTORY -save_path /TRAIN_DATASET -device cuda:0
python create_patches.py -input_path /VAL_DIRECTORY -save_path /VAL_DATASET -device cuda:0
python create_patches.py -input_path /TEST_DIRECTORY -save_path /TEST_DATASET -device cuda:0

Module3: DCNN classification

For the binary classification and multiple classes classfication, we applied a Deep Convolutional Neural Network (DCNN) model to classify patch-level images that were generated in the previous steps. It takes the extracted patches as input and learns discriminative features among different skin conditions.

python train.py -train_path /TRAIN_DATASET -val_path /VAL_DATASET -save_path /MODEL_SAVEPATH

This program evaluates the trained DCNN model on the test dataset. The model predicts the class of each input image patch, and the results (including predicted labels and confidence scores) are saved for further performance analysis.

python test.py -test_path /TEST_DIRECTORY -model_path /MODEL_SAVEPATH -save_path /RESULT_SAVEPATH

Module4: explanation and discovery

Disease inference and feature tracing

python inference.py -input_path /SLIDER.SVS -model_path /MODEL_SAVEPATH -save_path /RESULT_SAVEPATH

Radiomic analysis of skin features

python radiomic_feature.py -input_path /TEST_DATASET -save_path /RESULT_SAVEPATH

This module aims to interpret the decision process of the trained classification model and discover meaningful histological characteristics associated with different classes. It contains two main functionalities:

1. Model Inference with Visual Explanation: The inference.py script performs inference on a single whole-slide image (WSI). It generates both class prediction results and Grad-CAM heatmaps that highlight the most discriminative tissue regions used by the model.
2. Radiomic analysis: The radiomic_feature.py script extracts radiomic features from each image to quantify texture, shape, and intensity patterns. These features help reveal interpretable and human-understandable morphological characteristics linked to the prediction process of the model.

Please cite

Lin Y., Lin F., Zhang Y. et al. PathoEye: a deep learning framework for histopathological image analysis of skin tissue. Submitted.

Maintainer

Any questions, please contact @Yusen Lin

Contributors

Mr. Yusen Lin, Dr. Jiajian Zhou, Dr. Feiyan Lin and Miss Jiayu Wen.

License

MIT © Yusen Lin