MIL-Based Multi-Instance Learning Code for Ovarian Cancer Histopathology Image Classification from kaggle_UBC_ocean
choose Language:
Competition Link: kaggle_UBC_ocean
This project provides a method for classifying high-probability cancerous regions in high-resolution WSI images (up to 60000px):
The 236 mask images provided by the competition organizers are scaled down to thumbnail size, and then the U_net segmentation network is trained using the WSI thumbnails in train_thumbnails. The segmentation model results are as follows:
- For the image
train_images/22489.png:
After segmentation using U_net, the mask is as follows:
From the above mask, the smallest squares that can enclose the segmented regions are selected, and then coordinate scaling and transformation are performed to crop the corresponding positions in the original WSI image, as shown below:
The blue boxes represent the smallest squares that enclose the cancerous regions, and within each blue box, 1000*1000px random crops are performed as shown in the green squares. The cropped images are then sent to CNN for multi-instance classification of high-probability cancerous regions.
Using the technique from the paper TransMIL: First, extract features using cnn, then convert the feature map of each image into a single token, and send it to the encoder of the transformer for multi-instance learning.
The workflow is as follows:
- This diagram is from the transmil paper, which is a link to the transmil paper: arxiv.org/abs/2106.00908
This technique includes an image segmentation part compared to TransMIL, which reduces computational resources and improves classification accuracy.
The workflow is as follows:
After obtaining high-probability disease regions, CNN is used to extract features, which are then flattened and sent as sequences to the transformer for multi-instance learning. The code environment is the kaggle competition environment, and the code is in .ipynb format.
The environment can be configured as follows:
pip install -r requirements.txt