Pytorch implementation of the paper Bayesian Generative Models for Knowledge Transfer in MRI Semantic Segmentation Problems

We use Deep Weight Prior (DWP) [1] to perform transfer learning from large source to the smaller target dataset with medical images.

We use a common benchmark - BRATS18 [2] as a target and MS [3] dataset as source dataset.

The method perform better that randomly initialized and fine-tuned models.

[1] Atanov, A., Ashukha, A., Struminsky, K., Vetrov, D., and Welling, M. (2018). The deep weight prior.

[2] Menze, B. H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, J., et al. (2015). The multimodal brain tumor image segmentation benchmark (brats).

[3] CoBrain analytics, (2018). Multiple Sclerosis Human Brain MR Imaging Dataset

Experiments

Environment setup

The code is base on PyTorch 1.5.0 and PyTorch-Lightning 0.8.0. The exact specification of our environment is provided in the file environment.yml and can be created via

conda env create -f environment.yml

Example on toy dataset

• Load Source and target dataset into data/dataset_name folders. Below are examples for MNIST abd notMNIST datasets (they will be loaded automatically)

• Train N models on the source dataset. Here full notMNIST dataset is used

  python3 train.py --dataset_name notMNIST --trian_size -1 
  

• Train VAE on the kernels

  python3 train_vae.py --kernel_size 7
  

• Train model on the target dataset with VAE from the previous step as a prior

  python3 train.py --dataset_name MNIST --prior notMNIST --trian_size 100 
  

Citation

If you find our paper or code useful, feel free to cite the paper:

@article{kuzina2019bayesian,
  title={Bayesian generative models for knowledge transfer in mri semantic segmentation problems},
  author={Kuzina, Anna and Egorov, Evgenii and Burnaev, Evgeny},
  journal={Frontiers in neuroscience},
  volume={13},
  pages={844},
  year={2019},
  publisher={Frontiers}
}