DAMS

Domain Adaptation of MRI Scanners as an alternative to MRI harmonization

The paper has been accepted for presentation at the 5th MICCAI Workshop on Domain Adaptation and Representation Transfer (DART). Paper link

Download the slide of the presentation DART_Rafsanjany_Kushol.pdf

@inproceedings{kushol2023domain,
  title={Domain adaptation of MRI scanners as an alternative to MRI harmonization},
  author={Kushol, Rafsanjany and Frayne, Richard and Graham, Simon J and Wilman, Alan H and Kalra, Sanjay and Yang, Yee-Hong},
  booktitle={MICCAI Workshop on Domain Adaptation and Representation Transfer},
  pages={1--11},
  year={2023},
  organization={Springer}
}

Abstract

Combining large multi-center datasets can enhance statistical power, particularly in the field of neurology, where data can be scarce. However, applying a deep learning model trained on existing neuroimaging data often leads to inconsistent results when tested on new data due to domain shift caused by differences between the training (source domain) and testing (target domain) data. Existing literature offers several solutions based on domain adaptation (DA) techniques, which ignore complex practical scenarios where heterogeneity may exist in the source or target domain. This study proposes a new perspective in solving the domain shift issue for MRI data by identifying and addressing the dominant factor causing heterogeneity in the dataset. We design an unsupervised DA method leveraging the maximum mean discrepancy and correlation alignment loss in order to align domain-invariant features. Instead of regarding the entire dataset as a source or target domain, the dataset is processed based on the dominant factor of data variations, which is the scanner manufacturer. Afterwards, the target domain's feature space is aligned pairwise with respect to each source domain's feature map. Experimental results demonstrate significant performance gain for multiple inter- and intra-study neurodegenerative disease classification tasks.

Requirements

PyTorch
nibabel
scipy
scikit-image

Datasets

ADNI1, ADNI2, and AIBL dataset can be downloaded from ADNI (Alzheimer’s Disease Neuroimaging Initiative)

MIRIAD dataset can be downloaded from MIRIAD (Minimal Interval Resonance Imaging in Alzheimer's Disease)

CALSNIC dataset can be requested from CALSNIC (Canadian ALS Neuroimaging Consortium)

Preprocessing

Skull stripping using Freesurfer v7.3.2

Command mri_synthstrip -i input -o stripped

Details can be found SynthStrip (SynthStrip: Skull-Stripping for Any Brain Image)

Registration to MNI-152 using FSL FLIRT function

Details can be found FSL

One implementation can be found here. After registration, the image dimension will be $182\times218\times182$ and the voxel dimension will be $1\times1\times1$ $mm^3$.

Training

Run python train.py to train the network. It will generate dataset_source1_source2_to_target_max_accuracy.pth in Results folder

Testing

Run python test.py. It will load the pre-trained model dataset_source1_source2_to_target_max_accuracy.pth and generate the classification results based on the given target dataset

Contact

Email at: kushol@ualberta.ca

Acknowledgement

This basic structure of the code relies on the project of Deep Transfer Learning in PyTorch

Aligning Domain-specific Distribution and Classifier for Cross-domain Classification from Multiple Sources

Deep CORAL: Correlation Alignment for Deep Domain Adaptation