Informing QSM Background Field Correction with an External Reference Scan
Poster No:
1513
Submission Type:
Abstract Submission
Authors:
Xincheng Ye1, Ashley Stewart1, Frederik Testud2, Kieran O’Brien3, Thomas Andersen4, Shekhar Chandra1, Steffen Bollmann1
Institutions:
1The University of Queensland, Brisbane, Australia, 2Siemens Healthcare AB, Lund, Sweden, 3Siemens Healthcare Pty Ltd, Brisbane, Australia, 4Copenhagen University Hospital Rigshospitalet, Copenhagen, Denmark
First Author:
Co-Author(s):
Introduction:
Quantitative Susceptibility Mapping (QSM) is an advanced MRI technique that maps tissue magnetic susceptibility, offering insights into iron deposits and calcification. It enhances diagnostic accuracy for diseases like Parkinson's and Alzheimer's by revealing subtle tissue changes (Jung et al., 2022).
QSM requires dedicated processing steps, including phase unwrapping, background field removal, and solving an ill-posed dipole inversion problem where error accumulations across steps can lead to large susceptibility errors (Jung et al., 2022).
Background field removal is a critical step in QSM to isolate local tissue signals from external field effects. Traditional methods like PDF (Liu et al., 2011), LBV (Zhou et al., 2014), and RESHARP (Sun & Wilman, 2014) address this challenge with distinct mathematical frameworks but face limitations such as noise sensitivity and computational demands. Modern deep learning approaches like BFRnet (Zhu et al., 2023) and SHARQnet (Bollmann et al., 2019) offer faster and more accurate removal but depend on high-quality training data and may struggle with unseen scenarios. Moreover, during training, they simulate background fields by simple susceptibility geometrical shapes, resulting in less accurate estimates of realistic background fields.
Here, we present a new method that uses a novel deep-learning-based background field removal algorithm, which leverages the MRI-derived attenuation correction µ-maps (Ladefoged et al., 2020) to serve as a prior for structures causing the background field.
QSM requires dedicated processing steps, including phase unwrapping, background field removal, and solving an ill-posed dipole inversion problem where error accumulations across steps can lead to large susceptibility errors (Jung et al., 2022).
Background field removal is a critical step in QSM to isolate local tissue signals from external field effects. Traditional methods like PDF (Liu et al., 2011), LBV (Zhou et al., 2014), and RESHARP (Sun & Wilman, 2014) address this challenge with distinct mathematical frameworks but face limitations such as noise sensitivity and computational demands. Modern deep learning approaches like BFRnet (Zhu et al., 2023) and SHARQnet (Bollmann et al., 2019) offer faster and more accurate removal but depend on high-quality training data and may struggle with unseen scenarios. Moreover, during training, they simulate background fields by simple susceptibility geometrical shapes, resulting in less accurate estimates of realistic background fields.
Here, we present a new method that uses a novel deep-learning-based background field removal algorithm, which leverages the MRI-derived attenuation correction µ-maps (Ladefoged et al., 2020) to serve as a prior for structures causing the background field.
Methods:
The MRI-derived attenuation correction µ-maps provide a skull estimation of brain MR by overcoming the lack of bone signal. To estimate a realistic background field from a µ-map, we discretize a µmap into three categories: cerebrospinal fluid, air, and bone. Referring to the QSM challenge (Marques et al., 2021), we assign corresponding susceptibility values (ppm) to these three categories (µ-map chi) and then apply the QSM forward operation, obtaining a µ-map-based background field prior. To simulate more background field configurations for training, we utilize SynthSeg (Billot et al., 2023), which can generate synthetic MR brain and segmentation pairs of any contrast and resolution, given one segmentation label. Here, µ-map chi is used as a segmentation label.
To simulate total fields and local fields for training, we use SynthSeg on the QSM challenge data and apply a brain mask to the generated brain and QSM forward operation. The total fields are simulated by applying the QSM forward operation on the sum of the µ-map chi and the masked brain.
The above workflow is summarized in Figure 1, and was used to simulate 60 training data pairs, and trained a UNet with Mean Absolute Error (MAE) as the loss function. To suppress noise and artifacts with high-intensity values, we used Total Variation as a regularization term with a scale factor of 0.1. The learning rate was 0.0001, and the batch size was 2. The model was trained for 150 epochs. The experiments were conducted on an NVIDIA L40 with 48GB memory. Due to the memory limitations, all images were resized to (192, 192, 192).
To simulate total fields and local fields for training, we use SynthSeg on the QSM challenge data and apply a brain mask to the generated brain and QSM forward operation. The total fields are simulated by applying the QSM forward operation on the sum of the µ-map chi and the masked brain.
The above workflow is summarized in Figure 1, and was used to simulate 60 training data pairs, and trained a UNet with Mean Absolute Error (MAE) as the loss function. To suppress noise and artifacts with high-intensity values, we used Total Variation as a regularization term with a scale factor of 0.1. The learning rate was 0.0001, and the batch size was 2. The model was trained for 150 epochs. The experiments were conducted on an NVIDIA L40 with 48GB memory. Due to the memory limitations, all images were resized to (192, 192, 192).
Results:
We evaluated the model on a simulated test image, as shown in Figure 2. The left image is the ground truth, and the middle image is the prediction. The right image is the difference map between the ground truth and the prediction. The prediction accuracy was estimated using XSIM (Milovic et al., n.d.), which is a structure similarity optimized for QSM. The higher the XSIM is, the better the QSM is. The test image resulted in 0.9155.
Conclusions:
In summary, our approach has two main contributions: 1. By leveraging the MR-derived µ-map, we can have a better and more realistic estimate of the background field for training. 2. During prediction, this estimated background field serves as a prior, allowing the model to better predict the subtle differences covered by strong intensity from background sources.
Modeling and Analysis Methods:
Methods Development 1
Novel Imaging Acquisition Methods:
Imaging Methods Other 2
Keywords:
Computational Neuroscience
Machine Learning
MRI
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
Bollmann, S. (2019). SHARQnet – sophisticated harmonic artifact reduction in quantitative susceptibility mapping using a deep convolutional neural network. Zeitschrift Für Medizinische Physik, 29(2), 139–149.
Jung, W. (2022). Overview of quantitative susceptibility mapping using deep learning: Current status, challenges and opportunities. NMR in Biomedicine, 35(4), e4292.
Ladefoged, C. N. (2020). AI-driven attenuation correction for brain PET/MRI: Clinical evaluation of a dementia cohort and importance of the training group size. NeuroImage, 222, 117221.
Liu, T. (2011). A novel background field removal method for MRI using projection onto dipole fields (PDF). NMR in Biomedicine, 24(9), 1129–1136.
Marques, J. P. (2021). QSM reconstruction challenge 2.0: A realistic in silico head phantom for MRI data simulation and evaluation of susceptibility mapping procedures. Magnetic Resonance in Medicine, 86(1), 526–542.
Milovic, C. (n.d.). XSIM: A structural similarity index measure optimized for MRI QSM. Magnetic Resonance in Medicine, n/a(n/a).
Sun, H., & Wilman, A. H. (2014). Background field removal using spherical mean value filtering and Tikhonov regularization. Magnetic Resonance in Medicine, 71(3), 1151–1157.
Zhou, D. (2014). Background field removal by solving the Laplacian boundary value problem: BACKGROUND FIELD REMOVAL BY SOLVING LAPLACIAN BOUNDARY VALUE PROBLEM. NMR in Biomedicine, 27(3), 312–319.
Zhu, X. (2023). BFRnet: A deep learning-based MR background field removal method for QSM of the brain containing significant pathological susceptibility sources. Zeitschrift Für Medizinische Physik, 33(4), 578–590.