Χ-separation (chi-separation) toolbox v1.0: updates compared to the beta version and advanced func
Poster No:
1883
Submission Type:
Abstract Submission
Authors:
Sooyeon Ji1, Kyeongseon Min1, Hyeong-Geol Shin2,3, Junhyeok Lee4, Minjun Kim1, Sehong Oh5, Jongho Lee1
Institutions:
1Seoul National University, Seoul, Seoul, 2Johns Hopkins University School of Medicine, Baltimore, MD, 3Kennedy Krieger Institute, Baltimore, MD, 4Seoul National University College of Medicine, Seoul, Korea, Republic of, 5The Cleveland Clinic, Cleveland, OH
First Author:
Co-Author(s):
Hyeong-Geol Shin
Johns Hopkins University School of Medicine|Kennedy Krieger Institute
Baltimore, MD|Baltimore, MD
Johns Hopkins University School of Medicine|Kennedy Krieger Institute
Baltimore, MD|Baltimore, MD
Introduction:
χ-separation,[Shin et al., 2021] a method to separate para and diamagnetic susceptibility sources, demonstrated the ability to offer clinically valuable information.[Kim et al., 2023] For handy application of χ-separation, there is an increasing need for a toolset that generates high-quality χ-separation. Here, we developed the χ-separation toolbox v1.0 that provides a user-friendly GUI-based χ-separation, including three χ-separation algorithms and four advanced processing options to enhance the quality of the χ-separation results.
Methods:
Input/output
Complex 3D multi-echo GRE (mGRE) data with DICOM or NIFTI formats are accepted as input. In case of DICOM input, the relevant parameters (i.e., B0, B0 direction, center frequency, voxel size, echo times (TEs), and delta TE) are automatically read from the DICOM header. For NIFTI input, the parameters are manually filled in. An R2' (=R2*-R2) or R2* map and a brain mask may also be inputted. The outputs of the toolbox are the separated χpara and χdia maps (Figure 1).
Preprocessing of input data
When an R2' or R2* map is not provided, an R2* map is calculated from the mGRE data by fitting an exponential function to the decay curve. The brain mask is calculated using BET (FSL, FMRIB, Oxford, UK) when not provided.
For pipelined phase processing, existing toolboxes must be downloaded. The χ-separation toolbox provides two options for pipelined phase unwrapping according to the consensus [Bilgic et al., 2023]: ROMEO followed by weighted echo averaging, and non-linear complex fit followed by SEGUE. For background field removal, VSHARP[Wu et al., 2012] and PDF[Liu et al., 2011a] are pipelined.
Core functionality
• χ-sepnet: χ-separation using only mGRE data via a deep neural-network (χ-sepnet-R2*)[Kim et al., 2022] is included. The network is trained using COSMOS QSM and R2* maps as input, and χpara and χdia maps reconstructed using multi-orientation data as output.
• χ-separation: When an R2' map is available, two in-house algorithms for optimization-based χ-separation can be used: χ-separation-MEDI and χ-separation-iLSQR. The first option uses the MEDI-regularizer[Liu et al., 2011b], following the original paper.[Shin et al., 2021] The second option calculates χ-separation maps based on iLSQR.[Li et al., 2015]
Advanced processing options (Figure 2)
• Denoising: Two denoising options are provided for mGRE magnitude data used for R2* mapping. The first is MP-PCA denoising,[Veraart et al., 2016] which takes advantage of the intrinsic redundancy of the multi-echo images. Another option is a deep neural-network trained to denoise MR images.
• Background R2* correction: An R2* map is contaminated by background B0 inhomogeneity, due to non-local R2* effects.[Fernández‐Seara and Wehrli, 2000] This effect increases with larger effective voxel size related to k-space windowing. An option to apply k-space windowing along with correction of this effect is included to provide accurate χ-separation maps whilst reducing artifacts by k-space windowing.
• Resolution generalization for χ-sepnet: An option to utilize the "resolution generalization" pipeline for χ-sepnet is provided, which enables the inference of data that has higher or lower resolution than that of the training data (i.e., 1 mm iso).
• Vessel masking: χ-separation maps suffer from vessel artifacts that may hamper accurate calculation of ROI values. The toolbox provides an option to generate vessel masks that can be used to mask out vessels for group analysis.
Complex 3D multi-echo GRE (mGRE) data with DICOM or NIFTI formats are accepted as input. In case of DICOM input, the relevant parameters (i.e., B0, B0 direction, center frequency, voxel size, echo times (TEs), and delta TE) are automatically read from the DICOM header. For NIFTI input, the parameters are manually filled in. An R2' (=R2*-R2) or R2* map and a brain mask may also be inputted. The outputs of the toolbox are the separated χpara and χdia maps (Figure 1).
Preprocessing of input data
When an R2' or R2* map is not provided, an R2* map is calculated from the mGRE data by fitting an exponential function to the decay curve. The brain mask is calculated using BET (FSL, FMRIB, Oxford, UK) when not provided.
For pipelined phase processing, existing toolboxes must be downloaded. The χ-separation toolbox provides two options for pipelined phase unwrapping according to the consensus [Bilgic et al., 2023]: ROMEO followed by weighted echo averaging, and non-linear complex fit followed by SEGUE. For background field removal, VSHARP[Wu et al., 2012] and PDF[Liu et al., 2011a] are pipelined.
Core functionality
• χ-sepnet: χ-separation using only mGRE data via a deep neural-network (χ-sepnet-R2*)[Kim et al., 2022] is included. The network is trained using COSMOS QSM and R2* maps as input, and χpara and χdia maps reconstructed using multi-orientation data as output.
• χ-separation: When an R2' map is available, two in-house algorithms for optimization-based χ-separation can be used: χ-separation-MEDI and χ-separation-iLSQR. The first option uses the MEDI-regularizer[Liu et al., 2011b], following the original paper.[Shin et al., 2021] The second option calculates χ-separation maps based on iLSQR.[Li et al., 2015]
Advanced processing options (Figure 2)
• Denoising: Two denoising options are provided for mGRE magnitude data used for R2* mapping. The first is MP-PCA denoising,[Veraart et al., 2016] which takes advantage of the intrinsic redundancy of the multi-echo images. Another option is a deep neural-network trained to denoise MR images.
• Background R2* correction: An R2* map is contaminated by background B0 inhomogeneity, due to non-local R2* effects.[Fernández‐Seara and Wehrli, 2000] This effect increases with larger effective voxel size related to k-space windowing. An option to apply k-space windowing along with correction of this effect is included to provide accurate χ-separation maps whilst reducing artifacts by k-space windowing.
• Resolution generalization for χ-sepnet: An option to utilize the "resolution generalization" pipeline for χ-sepnet is provided, which enables the inference of data that has higher or lower resolution than that of the training data (i.e., 1 mm iso).
• Vessel masking: χ-separation maps suffer from vessel artifacts that may hamper accurate calculation of ROI values. The toolbox provides an option to generate vessel masks that can be used to mask out vessels for group analysis.
Results:
The advanced options offer high-quality χ-separation with fewer artifacts and less noise, and also a tool for more accurate analysis (Figure 2e).
Conclusions:
With the GUI-based χ-separation toolbox equipped with advanced processing functionalities, reconstruction of high-quality paramagnetic and diamagnetic susceptibility maps becomes readily available, fostering utilization of χ-separation across diverse applications.
Modeling and Analysis Methods:
Methods Development 1
Novel Imaging Acquisition Methods:
Imaging Methods Other 2
Keywords:
MRI
Other - Susceptibility Source Separation, Quantitative Susceptibility Mapping, Toolbox
1|2Indicates the priority used for review
Provide references using author date format
Fernández‐Seara MA (2000): Postprocessing technique to correct for background gradients in image‐based R*2 measurements. Magnetic Resonance in Medicine 44:358–366.
Kim M (2022): Chi-sepnet: Susceptibility source separation using deep neural network. In: . Proceedings of the 30th Annual ISMRM Meeting.
Kim W (2023): χ-Separation Imaging for Diagnosis of Multiple Sclerosis versus Neuromyelitis Optica Spectrum Disorder. Radiology 307:e220941.
Liu T (2011a): A novel background field removal method for MRI using projection onto dipole fields (PDF). NMR in Biomedicine 24:1129–1136.
Liu T (2011b): Morphology enabled dipole inversion (MEDI) from a single‐angle acquisition: Comparison with COSMOS in human brain imaging. Magnetic Resonance in Medicine 66:777–783.
Li W (2015): A method for estimating and removing streaking artifacts in quantitative susceptibility mapping. NeuroImage 108:111–122.
Shin H-G (2021): χ-separation: Magnetic susceptibility source separation toward iron and myelin mapping in the brain. Neuroimage 240:118371.
Veraart J (2016): Denoising of diffusion MRI using random matrix theory. Neuroimage 142:394–406.
Wu B (2012): Whole brain susceptibility mapping using compressed sensing. Magnetic Resonance in Medicine 67:137–147.