Mamba-Sea

Mamba-Sea: A Mamba-based Framework with Global-to-Local Sequence Augmentation for Generalizable Medical Image Segmentation

Mamba-Sea has been accepted by IEEE TMI. 【Paper】

Tips

The core part of Mamba-Sea's code is in Network1 and Network2.By using these two files as a backbone, you can easily incorporate our code into your tasks.

Prepare Datasets

• './dataset/fundus/' or './dataset/ProstateSlice/' or './dataset/skin/'

  -Domain1

  • train
    • images
      • .png
    • masks
      • .png
  • test
    • images
      • .png
    • masks
      • .png ......

Pre_trained Weights

The weights of the pre-trained VMamba could be downloaded Google Drive. After that, the pre-trained weights should be stored in './pretrained_weights/'.

Dataset

Fundus

Download dataset Fundus (Provided by DoFE).

Prostate

Download dataset Prostate (Originally Provided by SAML and RAM-DSIR).

Skin

Download dataset ISIC2018 and PH2. (Following ESP-MedSAM).

Train and Test

For exmaple,

./train.sh parallel train_dg.py fundus VMUnet 0,0,1,1 1.0 adamw

Contact Information

Email: chengzihan@sjtu.edu.cn or czh@smail.nju.edu.cn, each one is OK.

Acknowledgments

Thanks a lot for the valuable work of VM-UNet and VMamba.