isyuchenyuan / ssci

The PyTorch implementation of Semi-supervised Class Imbalanced Deep Learning for Cardiac MRI Segmentation (MICCAI 2023)

Semi-supervised Class Imbalanced Deep Learning for Cardiac MRI Segmentation (MICCAI 2023)

This repository contains the source code of the paper Semi-supervised Class Imbalanced Deep Learning for Cardiac MRI Segmentation.

Prepare

Datasets

Our method is evaluated with two datasets: ACDC and MMWHS.

• For the ACDC dataset, please download the original training images from the official site. Place the folder training as follows:

├── data
│   ├── ACDC
│   │   ├──training
│   │   ├──train.txt

• For the MMWHS dataset, please download the original training images from the official site. Only the MRI images are used in this work. Place the folder mr_train as follows:

├── data
│   ├── MMWHS
│   │   ├──mr_train
│   │   ├──train_c.txt

You need to change the name of data paths according to your actual experiments.

Environment

In order to install the correct environment, please run the following script:

conda create -n ssci python=3.7.13
conda activate ssci
pip install -r requirements.txt

Building from source

Before the training, you need to first compile the C++ extension of the tree filters implementation.

• cd code/utils/lib_tree_filter
• sudo python3 setup.py build develop

Run

MMWHS preprocessing

First, do the preprocessing for the MMWHS dataset by runing:

python cardiac_processing.py

Model training

Then, train the model by using the MMWHS dataset

python train_MMWHS.py

Train the model by using the ACDC dataset

python train_ACDC.py

Model inference

Test te trained model on the evaluation set, for the MMWHS dataset

python test_MMWHS.py

For the ACDC dataset

python test_ACDC.py

Acknowledgement

The implementations of tree filters and prototype learning are borrowed from TreeFilter-Torch and ProtoSeg

• TreeFilter-Torch: https://github.com/Megvii-BaseDetection/TreeFilter-Torch
• ProtoSeg: https://github.com/tfzhou/ProtoSeg

Thanks a lot for their splendid work and code sharing!

Results

The qualitative analyses of ablation studies.

(a) Original slice; (b) Label; Prediction results from (c) 3D UNet; (d) 3D UNet + PC + CRS ; (e) H-UNet + LC; (f) H-UNet + LC + CRS; (g) H-UNet + PC; (h) H-UNet + LC + PC; (e) H-UNet + LC + PC + CRS (ours)

Contact

If you have any questions or comments, please feel free to cantact me via [email protected]