Benchmarking Experiments for "ViM-UNet: Vision Mamba for Biomedical Segmentation"

Experiments performed using reference methods to benchmark for ViM-UNet described in our preprint (accepted to MIDL 2024 - Short Paper):

• nnU-Net
• U-Mamba

Installation

For nnU-Net:

Here is the detailed instruction on how to install nnU-Net.

TLDR:

1. Install PyTorch
2. Install nnU-Net from source:

$ git clone https://github.com/MIC-DKFZ/nnUNet.git
$ cd nnUNet
$ pip install -e .

For U-Mamba:

Here is the detailed instruction on how to install U-Mamba.

Below is my piece of installation (dropping it here as some parts needed some attention)

1. Create a new mamba environment:

$ mamba env create -n umamba python=3.10 -y
$ mamba activate umamba

2. Install PyTorch:

mamba install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

3. Install packaging: pip install packaging

CUDA_HOME needs to match the installed cuda version, and the path should be visible. For HLRN users, here's an example: export CUDA_HOME=/usr/local/cuda-11.8/.

4. Install causal-conv1d: pip install causal-conv1d==1.1.1
5. Install Mamba: pip install mamba-ssm
6. Clone the repository from scratch and install U-Mamba (we store the data at U-Mamba/data for performing the experiments)

$ git clone https://github.com/bowang-lab/U-Mamba.git
$ cd U-Mamba/umamba
$ pip install -e .

To cite our paper:

@inproceedings{
  archit2024vimunet,
  title={ViM-{UN}et: Vision Mamba for Biomedical Segmentation},
  author={Anwai Archit and Constantin Pape},
  booktitle={Medical Imaging with Deep Learning},
  year={2024},
  url={https://openreview.net/forum?id=PYNwysgFeP}
}