---

I2I-Mamba
_{Multi-modal medical image synthesis via selective state space modeling}

Omer F. Atli^1,2   ·   Bilal Kabas^1,2   ·   Fuat Arslan^1,2   ·   Arda C. Demirtas^1,2   ·   Mahmut Yurt³   ·   Onat Dalmaz³   ·   Tolga Çukur^1,2  

¹Bilkent University   ²UMRAM   ³Stanford University

---

[arXiv]

Official PyTorch implementation of I2I-Mamba, a novel adversarial model for multi-modal medical image synthesis that leverages selective state space modeling (SSM) to efficiently capture long-range context while maintaining local precision. I2I-Mamba injects channel-mixed Mamba (cmMamba) blocks in the bottleneck of a convolutional backbone. In cmMamba blocks, SSM layers are used to learn context across the spatial dimension and channel-mixing layers are used to learn context across the channel dimension of feature maps.

Dependencies

python>=3.8.13
cuda=>11.6

torch>=2.2
torchvision>=0.17
visdom
dominate
scikit-image
h5py
scipy
ml_collections
mamba-ssm==1.1.3

You are welcome to ask any library issues.

Dataset

To reproduce the results reported in the paper, we recommend the following dataset processing steps:

Sequentially select subjects from the dataset. Apply skull-stripping to 3D volumes. Select 2D cross-sections from each subject. Normalize the selected 2D cross-sections before training and before metric calculation.

You should structure your aligned dataset in the following way:

/datasets/IXI/
  ├── T1_T2
  │   ├── train
  |    - 1.png
  |    - 2.png
  |    - ...
  │   ├── val
  │   └── test
  ├── T2_PD
  │   ├── train
  │   └── ...
  ├── T1_PD__T2
  │   ├── train
  │   └── ...
  

x.png is a single image containing both source and target image side-by-side. Thus the resulting image for one-to-one task will be 512 pixels horizontally and 256 pixels vertically and for many-to-one task it will be 768 pixels horizontally and 256 pixels vertically.

Note that for many-to-one tasks, source modalities should be in the Red and Green channels.

Training

Step 0

Before training your data you are required to run path_generate.py to generate 'spiral_eye.npy', 'spiral_r_eye_.npy', 'despiral_r_eye_.npy' which are required for training and testing. Note that this file will be run only for the first time.

python3 models/path_generate.py

Commands

1. One-to-one training, e.g. T2->PD

python3 train.py --dataroot datasets/IXI/T2_PD/ --name ixi_t2__pd --gpu_ids 0 --model i2i_mamba_one --which_model_netG i2i_mamba --which_direction AtoB --lambda_A 100 --dataset_mode aligned --norm batch --pool_size 0 --output_nc 1 --input_nc 1 --loadSize 256 --fineSize 256 --niter 30 --niter_decay 30 --save_epoch_freq 5 --checkpoints_dir checkpoints/ --display_id 0 --lr 0.0001

2. Many-to-one training, e.g. T1,T2->PD

python3 train.py --dataroot datasets/IXI/T1_T2__PD/ --name ixi_t1_t2__pd --gpu_ids 0 --model i2i_mamba_many --which_model_netG i2i_mamba --which_direction AtoB --lambda_A 100 --dataset_mode aligned --norm batch --pool_size 0 --output_nc 1 --input_nc 3 --loadSize 256 --fineSize 256 --niter 30 --niter_decay 30 --save_epoch_freq 5 --checkpoints_dir checkpoints/ --display_id 0 --lr 0.0001

Argument descriptions

Argument	Description
--dataroot	Root directory path of the dataset.
--name	Name of the experiment, used for storing model checkpoints and results.
--gpu_ids	List of GPU IDs to use for training (e.g., 0 for the first GPU).
--model	Model type to use for training.
--which_model_netG	Specifies the generator architecture to use.
--which_direction	Direction of the image translation, e.g., AtoB or BtoA.
--lambda_A	Weight for cycle consistency loss from domain A to B and back to A.
--dataset_mode	Dataset loading mode, specifying how the dataset is loaded (e.g., aligned).
--norm	Normalization type used in the model, e.g., batch normalization.
--pool_size	Size of the image buffer that stores previously generated images.
--output_nc	Number of output image channels.
--input_nc	Number of input image channels.
--loadSize	Scale images to this size before cropping.
--fineSize	Crop images to this size for training.
--niter	Number of iterations with the initial learning rate.
--niter_decay	Number of iterations to linearly decay the learning rate to zero.
--save_epoch_freq	Frequency of saving checkpoints at the end of epochs.
--checkpoints_dir	Directory where model checkpoints are saved.
--display_id	ID of the display window for visualizing training results (0 to disable).
--lr	Initial learning rate for the optimizer.

Testing

Commands

1. One-to-one testing, e.g. T2->PD

python3 test.py --dataroot datasets/IXI/T2_PD/ --name ixi_t2__pd --gpu_ids 0 --model i2i_mamba_one --which_model_netG i2i_mamba --dataset_mode aligned --norm batch --phase test --output_nc 1 --input_nc 1 --how_many 10000 --serial_batches --fineSize 256 --loadSize 256 --results_dir results/ --checkpoints_dir checkpoints/ --which_epoch 60

2. Many-to-one testing, e.g. T1,T2->PD

python3 test.py --dataroot datasets/IXI/T1_T2__PD/ --name T2_PD_T1_mambaoffTekrar_nonorm --gpu_ids 0 --model i2i_mamba_many --which_model_netG i2i_mamba --dataset_mode aligned --norm batch --phase test --output_nc 1 --input_nc 3 --how_many 10000 --serial_batches --fineSize 256 --loadSize 256 --results_dir results/ --checkpoints_dir checkpoints/ --which_epoch 60

Argument descriptions

Argument	Description
--results_dir	Directory where test results are saved.
--serial_batches	If set, takes images in order to make batches, otherwise takes them randomly.
--how_many	Number of test images to run.
--which_epoch	Which epoch to load? Set to 'latest' to use the latest cached model.

Citation

You are encouraged to modify/distribute this code. However, please acknowledge this code and cite the paper appropriately.

@article{atli2024i2imamba,
  title={I2I-Mamba: Multi-modal medical image synthesis via selective state space modeling}, 
  author={Omer F. Atli and Bilal Kabas and Fuat Arslan and Mahmut Yurt and Onat Dalmaz and Tolga Çukur},
  year={2024},
  journal={arXiv:2405.14022}
}

For any questions, comments and contributions, please feel free to contact Omer Faruk Atli (faruk.atli[at]bilkent.edu.tr)

Acknowledgments

This code uses libraries from ResViT and mamba repository.

---

Copyright © 2024, ICON Lab.