Conditional Deformable Image Registration with Convolutional Neural Network

This is the official Pytorch implementation of "Conditional Deformable Image Registration with Convolutional Neural Network" (MICCAI 2021), written by Tony C. W. Mok and Albert C. S. Chung.

Prerequisites

• Python 3.5.2+
• Pytorch 1.3.0 - 1.7.1
• NumPy
• NiBabel

This code has been tested with Pytorch 1.7.1 and NVIDIA TITAN RTX GPU.

Inference

python Test_cLapIRN.py

The regularization weight can be changed by appending --reg_input {{normalized weight within [0,1]}} argument to the inference script. For example,

python Test_cLapIRN.py --reg_input 0.4

is equivalent to output the solution with regularization weight set to 4.

Train your own model

Step 1: Replace /PATH/TO/YOUR/DATA with the path of your training data. You may also need to implement your own data generator (Dataset_epoch in Functions.py).

Step 2: Change the imgshape variable in Train_cLapIRN.py to match the resolution of your data.

Step 3: python Train_cLapIRN.py to train the model. Remember the data should be normalized within [0,1]. Otherwise, set norm=True in the provided data loader.

(Optional): Implement the custom validation code in line 368 at Train_cLapIRN.py.

Scalability/Out of memory error

1. You may adjust the size of the model by manipulating the argument --start_channel in Train_cLapIRN.py and Test_cLapIRN.py

2. You may modify the number of conditional image registration module in resblock_seq function (at Functions.py).

(Example) Training on the preprocessed OASIS dataset without cropping

If you want to train on the preprocessed OASIS dataset in https://github.com/adalca/medical-datasets/blob/master/neurite-oasis.md. We have an example showing how to train on this dataset.

1. Download the preprocessed OASIS dataset, unzip it and put it in "Data/OASIS".
2. To train a new conditional LapIRN model, python Train_cLapIRN_lite.py will create a conditional LapIRN model trained on all cases in the dataset.
3. To test the model, python Test_cLapIRN_lite.py --modelpath {{pretrained_model_path}} --fixed ../Data/image_A_fullsize.nii.gz --moving ../Data/image_B_fullsize.nii.gz will load the assigned model and register the image "image_A_fullsize.nii.gz" and "image_B_fullsize.nii.gz".

Note that the conditional LapIRN model in Train_cLapIRN_lite.py is a lightweight version, which reduced the number of feature maps in the original model. A pretrained model and its log file are available in "Model/LDR_OASIS_NCC_unit_disp_add_fea4_reg01_10_lite_stagelvl3_54000.pth" and "Log/LDR_OASIS_NCC_unit_disp_add_fea4_reg01_10_lite_.txt", respectively.

(Example) Training on 2D images

We demonstrate the 2D deformable image registration with 2D coronal slices extracted from the preprocessed OASIS dataset (available in https://github.com/adalca/medical-datasets/blob/master/neurite-oasis.md) as follows:

1. Download the preprocessed OASIS dataset, unzip it and put it in "Data/OASIS".
2. To train a new conditional LapIRN model, python Train_cLapIRN_2D.py will create a conditional LapIRN model trained on all cases in the dataset.
3. To test the model, python Test_cLapIRN_2D.py --modelpath {{pretrained_model_path}} --fixed ../Data/image_A_2D.nii.gz --moving ../Data/image_B_2D.nii.gz will load the assigned model and register the image "image_A_fullsize.nii.gz" and "image_B_fullsize.nii.gz".

Note that the 2D images in the dataset are in resolution (160, 192, 1). During training and testing, we drop the last dimension using .squeeze(-1). The pretrained model for 2D conditional LapIRN can be downloaded here.

Publication

If you find this repository useful, please cite:

• Conditional Deformable Image Registration with Convolutional Neural Network
  Tony C. W. Mok, Albert C. S. Chung
  MICCAI 2021. eprint arXiv:2106.12673

• Large Deformation Diffeomorphic Image Registration with Laplacian Pyramid Networks
  Tony C. W. Mok, Albert C. S. Chung
  MICCAI 2020. eprint arXiv:2006.16148

Acknowledgment

Some codes in this repository are modified from IC-Net and VoxelMorph.

Keywords

Keywords: Conditional Image registration, Controllable Regularization, Deformable Image Registration