Author: JIANG Maiqi
The whole slide image (WSI) is a kind of medical image with huge size. It is hard to process the whole slide image as general image. The data will be processed by splitting WSI to patches, following the doctor's gaze. The gaze is the doctor's attention to the WSI. So, we can use the gaze to split the WSI to patches which are the input of deep learning model. The model can be used to predict the doctor's attention to the WSI and classify the WSI by a semi-supervised learning method.
This repository is only focus on the model running. The data preprocessing is not included in this repository.
It is based on private wsi data, so we can not provide the data. There is another private data need to be prepared. The data is the
doctor's gaze data. If you want to use your own gaze data, the gaze data loading can reference ./tool/dataProcessor.py and /tool/recReader.py.
These two files (provided by QIAO Siyu, Northeastern University, China) are used to load the gaze data. The gaze data is .rec file and .erc file. These files are binary
files. They are file formats.
As for the installation, you need to check out the environment first. The environment is as follows:
python ~= 3.8.13
CUDA ~= 11.3
CuDNN ~= 8.2.0
(Optional) MiniConda
If you are using Windows, you can install OpenSlide by following the guidance from OpenSlide Website.
For linux, the installation is a little easier. For example, in ubuntu, you can just use the following command:
apt install python-openslide
pip install Openslide-python
OpenSlide is a useful tool for reading WSI.
In order to process the WSI, you need to install OpenCV. In ubuntu, you can install it by following the command:
sudo apt-get install python3-opencv
In Windows, you can install it by following the command:
pip install opencv-python
pip install opencv-contrib-python
Following your CUDA version, you can install Pytorch as suggestion command from Pytorch Official Page. For example, if you are using CUDA 11.3, you can install Pytorch by following the command:
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch # if you are using conda
There are several attempts in this repository. You can open one of the folders to try it, such as following:
cd ./transformer
python train_pure_mil.py
There is also a test file for you to test the model. For example, you can use the following command to test the model:
cd ./transformer
python test_mil_based.py