Source-linked AI summary
Quantitative Susceptibility Mapping using Deep Neural Network: QSMnet
Jaeyeon Yoon, Enhao Gong, Itthi Chatnuntawech, Berkin Bilgic, Jingu Lee, Woojin Jung, Jingyu Ko, Hosan Jung, Kawin Setsompop, Greg Zaharchuk, Eung Yeop Kim, John Pauly, Jongho Lee
TL;DR
QSM requires ill-conditioned dipole deconvolution, while existing approaches require multiple orientations or introduce artifacts. This paper presents QSMnet, a deep neural network trained with COSMOS-derived labels for single-orientation reconstruction; it produced high-quality, orientation-consistent maps, though clinical reliability and physical correctness remain limited concerns.
Problem
QSM dipole deconvolution is ill-conditioned, and existing methods either require multiple-orientation scans or suffer from artifacts.
Method
QSMnet uses a modified 3D U-net trained with orientation-matched COSMOS QSM maps and losses for model consistency, voxel agreement, and edge preservation.
Results
QSMnet produced high-fidelity maps with superior image quality and better multi-orientation consistency than TKD or MEDI, approaching COSMOS quality.
Takeaways & Limitations
The results support QSMnet as a promising single-orientation reconstruction approach and suggest potential applicability to abnormalities not represented in training.
Takeaways & Limitations
Clinical reliability is difficult to guarantee, and white-matter consistency may not be physically correct because training used isotropic susceptibility data.
Abstract
from arXiv · showhide
Deep neural networks have demonstrated promising potential for the field of medical image reconstruction. In this work, an MRI reconstruction algorithm, which is referred to as quantitative susceptibility mapping (QSM), has been developed using a deep neural network in order to perform dipole deconvolution, which restores magnetic susceptibility source from an MRI field map. Previous approaches of QSM require multiple orientation data (e.g. Calculation of Susceptibility through Multiple Orientation Sampling or COSMOS) or regularization terms (e.g. Truncated K-space Division or TKD; Morphology Enabled Dipole Inversion or MEDI) to solve the ill-conditioned deconvolution problem. Unfortunately, they either require long multiple orientation scans or suffer from artifacts. To overcome these shortcomings, a deep neural network, QSMnet, is constructed to generate a high quality susceptibility map from single orientation data. The network has a modified U-net structure and is trained using gold-standard COSMOS QSM maps. 25 datasets from 5 subjects (5 orientation each) were applied for patch-wise training after doubling the data using augmentation. Two additional datasets of 5 orientation data were used for validation and test (one dataset each). The QSMnet maps of the test dataset were compared with those from TKD and MEDI for image quality and consistency in multiple head orientations. Quantitative and qualitative image quality comparisons demonstrate that the QSMnet results have superior image quality to those of TKD or MEDI and have comparable image quality to those of COSMOS. Additionally, QSMnet maps reveal substantially better consistency across the multiple orientations than those from TKD or MEDI. As a preliminary application, the network was tested for two patients. The QSMnet maps showed similar lesion contrasts with those from MEDI, demonstrating potential for future applications.
INTRODUCTION
QSM reconstructs magnetic susceptibility from GRE-derived field maps by inverting a dipole convolution, but the Fourier-domain inversion is ill-conditioned. Existing solutions trade scan time against artifacts, motivating deep-learning reconstruction.
- Magnetic susceptibility is a tissue property relevant to clinical diagnosis and quantifying susceptibility sources.
- GRE phase images provide field maps that can be deconvolved to regenerate the underlying susceptibility distribution.
- QSM inversion is ill-conditioned because the Fourier-transformed dipole pattern contains zeros, creating division-by-zero regions.
- TKD avoids division singularities by truncating k-space inversion but suffers from streaking artifacts.
- MEDI reduces streaking by constraining susceptibility-map edges to resemble the T2*-weighted magnitude image, although errors remain.
- Deep neural networks offer nonlinear input-output mapping and have shown promise for medical image reconstruction.
MRI data acquisition and processing
The study acquired multi-orientation GRE data from healthy volunteers and single-orientation data from two patients, then processed the images into local field maps and reference reconstructions.
- 35 scans from seven healthy volunteers were acquired at five head orientations, while two patients were scanned at one orientation.
- Healthy-volunteer GRE scans used 1 mm isotropic voxels, TR = 33 ms, TE = 25 ms, and approximately 5-minute acquisition times.
- Magnitude and phase images were reconstructed offline, brain masks were generated, and phase images were spatially unwrapped to produce local field maps.
- Five-orientation local field maps were registered and processed with COSMOS to generate gold-standard susceptibility maps.
- Single-orientation maps were also generated with TKD and MEDI, with five orientation-specific maps used to assess contrast consistency.
Deep neural network for QSM: QSMnet
QSMnet is a 3D U-net trained to learn dipole deconvolution from local field images using rotated COSMOS maps as labels. Its physics-aware and image-based losses support susceptibility reconstruction and edge preservation.
- 25 scans from five subjects trained QSMnet, while five scans each from separate subjects formed validation and test sets.
- QSMnet receives unregistered local field images and uses orientation-matched rotated COSMOS QSM maps as training labels.
- Data augmentation doubled the training data by rotating COSMOS maps and regenerating local field maps through dipole convolution.
- The network is a modified 3D U-net with convolutional, pooling, deconvolution, normalization, nonlinear, and feature-concatenation layers.
- Three losses enforce dipole-model consistency, voxel-wise agreement, and edge preservation in the reconstructed map.
- QSMnet maps showed high fidelity to COSMOS, whereas TKD and MEDI maps exhibited streaking artifacts in test-set reconstructions.
Evaluation of QSM algorithms
The evaluation compared QSMnet with TKD and MEDI using image-quality metrics, multi-orientation ROI consistency, and preliminary patient reconstructions.
- Test-set maps were compared using pSNR, RMSE, HFEN, SSIM, residual-error maps, and reconstruction time.
- Five ROIs were analyzed by calculating mean susceptibility and standard deviation across the five head orientations.
- QSMnet was preliminarily evaluated in two untrained patients, one with a microbleed and one with multiple sclerosis lesions, against MEDI.
RESULTS
QSMnet produced higher-quality QSM maps than TKD and MEDI, with fewer artifacts, better quantitative metrics, and stronger consistency across head orientations. Its maps closely matched COSMOS and preserved structural detail, while reconstruction was substantially faster than MEDI.
- QSMnet had the lowest RMSE against COSMOS: 0.016, versus 0.034 for TKD and 0.029 for MEDI.
- QSMnet achieved the best pSNR, RMSE, HFEN, and SSIM among the three reconstruction methods.
- Across five head orientations, QSMnet closely matched COSMOS and avoided the streaking artifacts and contrast variations seen in TKD and MEDI.
- QSMnet preserved cortical-ribbon detail across orientations, whereas TKD and MEDI lost structural information.
- ROI susceptibility values from QSMnet most closely matched COSMOS and showed the smallest error across head orientations.
- QSMnet reconstructed maps in 6.3 ± 0.0 seconds, compared with 255.8 ± 18.2 seconds for MEDI.
- In two preliminary patient applications, QSMnet produced lesion contrasts comparable to MEDI.
DISCUSSION AND CONCLUSION
QSMnet produced high-quality, orientation-consistent susceptibility maps and showed comparable lesion contrasts to MEDI, but its clinical applicability and physical interpretation remain limited by scope and interpretability concerns.
- DISCUSSION AND CONCLUSION: QSMnet produced high-quality QSM maps close to gold-standard COSMOS maps and consistent susceptibility results across multiple head orientations.The consistency supports reproducibility for longitudinal studies requiring repeated scans.
- DISCUSSION AND CONCLUSION: QSMnet produced consistent white-matter contrasts across head orientations, but these results may not be physically correct because white-matter susceptibility is anisotropic.The network was trained only on isotropic COSMOS susceptibility, which may suppress anisotropy.
- DISCUSSION AND CONCLUSION: QSMnet showed comparable lesion contrasts to MEDI in preliminary microbleed and multiple sclerosis patient maps.The patient evaluation was limited to two patients.
- DISCUSSION AND CONCLUSION: MEDI results retained substantial variability across multiple head orientations despite testing different regularization factors.This variability reflects the sensitivity of image quality to the regularization factor.
- DISCUSSION AND CONCLUSION: The network has fixed input-resolution and B0-axis assumptions that constrain processing of higher-resolution or differently oriented data.Higher-resolution data may require a new network and differently oriented inputs require matrix rotation.
- DISCUSSION AND CONCLUSION: Extensive validation in healthy volunteers and patients is still needed to confirm QSMnet validity and applicability.The authors also identify continued work on understanding the network’s characteristics.
ACKNOWELDGEMENTS
The research received support from the National Research Foundation of Korea and Seoul National University’s Creative-Pioneering Researchers Program.
- ACKNOWELDGEMENTS: The research was supported by the National Research Foundation of Korea through grant NRF-2017M3C7A1047864.
- ACKNOWELDGEMENTS: Additional support came from the Creative-Pioneering Researchers Program of Seoul National University.
- ACKNOWELDGEMENTS: The supplied acknowledgements passage identifies funding sources but does not describe additional institutional or personnel contributions.
SUPPLEMENTARY INFORMAITON
The supplementary information documents the QSMnet processing pipeline, regions of interest, residual errors, orientation comparisons, and MEDI regularization effects.
- Data processing pipeline: The data-processing pipeline registers local field maps, computes COSMOS QSM maps, and rotates them back to original head orientations for QSMnet training pairs.
- Regions of interest: The supplementary material identifies five ROIs: red nucleus, substantia nigra, globus pallidus, putamen, and caudate.
- Residual error maps: Residual error maps referenced to COSMOS show large errors and more pronounced streaking artifacts in TKD and MEDI results.
- Multiple head orientations: Axial and sagittal comparisons across head orientations show streaking artifacts and orientation variability for TKD and MEDI, whereas QSMnet shows no visually detected streaking artifacts.
- Effects of the regularization factor in MEDI: MEDI regularization of λ = 1000 caused substantial smoothing, while λ = 3000 improved structure delineation at the cost of increased streaking artifacts.
- Effects of the regularization factor in MEDI: Changing MEDI regularization factors did not eliminate variability across multiple head orientations in ROI analyses.