CALAMITI

<meta name="title" content="CALAMITI"/>

Contrast Anatomy Learning and Analysis for MR Intensity Translation and Integration (CALAMITI)

Contrast Anatomy Learning and Analysis for MR Intensity Translation and Integration (CALAMITI) is our current MR harmonization method. It was designed to achieve unsupervised multi-site MR harmonization. The associated publication is:

• L. Zuo, B.E. Dewey, A. Carass, Y. Liu, Y. He, P.A. Calabresi, and J.L. Prince, "Information-based Disentangled Representation Learning for Unsupervised MR Harmonization", 27^th Conference on Information Processing in Medical Imaging (IPMI 2021), Virtually in Bornholm, Denmark, June 28 - July 2, 2021.

Software

CALAMITI (2D)

220k

Instructions

Preprocessing

CALAMITI requires the following preprocessing steps:

• N4 inhomogeneity correction.
• Registration to MNI space (0.8mm isotropic resolution is ideal. Image dimensions after MNI registration is 241*288*241).
• White matter peak normalization (see https://github.com/jcreinhold/intensity-normalization).

Prepare training

CALAMITI requires paired multi-contrast MR images (e.g., T1-w and T2-w) during training. The ideal structure of the data directory and naming convention are as follows:

   ├──absolute_path_to_data
       ├──SiteA
       |    ├──train
       |    |    ├──SiteA_T1w_SUB*_ORIENTATION_SLICE*.nii.gz ("ORIENTATION" should be "AXIAL", "CORONAL", or "SAGITTAL")
       |    |    └──SiteA_T2w_SUB*_ORIENTATION_SLICE*.nii.gz
       |    └──valid
       |         ├──SiteA_T1w_SUB*_ORIENTATION_SLICE*.nii.gz 
       |         └──SiteA_T2w_SUB*_ORIENTATION_SLICE*.nii.gz
       └──SiteB
           ├──train
           |    ├──SiteB_T1w_SUB*_ORIENTATION_SLICE*.nii.gz 
           |    └──SiteB_T2w_SUB*_ORIENTATION_SLICE*.nii.gz
           └──valid
                ├──SiteB_T1w_SUB*_ORIENTATION_SLICE*.nii.gz
                └──SiteB_T2w_SUB*_ORIENTATION_SLICE*.nii.gz

Sample command

Sample command for CALAMITI training and testing (encoding and decoding) is provided under "script" folder.

If you have other questions regarding the method or software, please email Lianrui Zuo at lr_zuo@jhu.edu<\code>