Abstract
Quantitative susceptibility mapping (QSM) is frequently hindered by streaking artifacts arising from zeros on the conical surface of the dipole kernel in k-space. This study presents ALOHA-QSM, a novel reconstruction method based on a direct k-space interpolation approach. By formulating the QSM reconstruction as a deconvolution problem under a low-rank Hankel matrix constraint, the method avoids over-smoothing and artifacts typical of conventional approaches. To manage computational complexity, the problem is solved through successive 2D plane reconstructions along three independent axes. Experimental results demonstrate that ALOHA-QSM effectively reduces artifacts and provides superior quantitative accuracy in deep gray matter structures without requiring additional anatomical priors.
Introduction
Quantitative Susceptibility Mapping (QSM) is a non-invasive MRI technique that quantifies the magnetic susceptibility of biological tissues. This is particularly valuable for monitoring iron deposition, calcification, and oxygenation levels in the brain, which serve as biomarkers for neurodegenerative conditions such as Parkinson’s and Alzheimer’s disease.
The fundamental challenge in QSM lies in the "ill-posed" nature of the dipole inversion. The relationship between the measured magnetic field (phase) and the underlying susceptibility is described by a dipole kernel in the frequency domain (k-space). This kernel contains zeros on a conical surface at the "magic angle" (approximately 54.7° relative to the main magnetic field). Attempting to invert this relationship leads to signal voids and significant streaking artifacts, which can obscure clinical details and degrade quantitative accuracy.
Methodology: The ALOHA-QSM Framework
To address the dipole inversion problem, this research introduces ALOHA-QSM (Annihilating Filter-based Low-Rank Hankel Matrix approach). This method shifts the reconstruction from the image domain to k-space, treating the artifact removal as a data interpolation problem.
Low-Rank Hankel Matrix Completion
The core of the ALOHA approach is the mathematical duality between transform-domain sparsity and the rank of a Hankel matrix in the signal domain. If an image is sparse or can be represented sparsely (e.g., via wavelets), its k-space data, when structured into a Hankel matrix, will be low-rank. ALOHA-QSM leverages this property to "predict" the missing or corrupted data on the dipole kernel's conical surface. By enforcing a low-rank constraint, the algorithm recovers the missing k-space information based on the surrounding valid data.
Successive 2D Processing
A significant hurdle for k-space methods is the memory requirement; a full 3D Hankel matrix for a standard brain volume is computationally prohibitive. To overcome this, the authors implemented a successive 2D reconstruction strategy. The 3D volume is processed as a series of 2D planes along the x, y, and z axes iteratively. This relaxation approach reduces the memory footprint and computational load while maintaining the global consistency required for high-quality 3D reconstruction.
Quantitative Correction
K-space interpolation methods can sometimes lead to an underestimation of susceptibility values. To ensure physical accuracy, the researchers introduced a global correction factor ($s_m$). This factor is derived from a linear regression between the original phase and the reformulated phase, ensuring the final reconstructed values align with the physical reality of the tissue magnetism.
Experimental Validation
The performance of ALOHA-QSM was evaluated using three distinct datasets:
- Numerical Phantom: A simulated brain model used to test the algorithm against known ground-truth values under noisy conditions (SNR=10).
- QSM2016 Reconstruction Challenge Data: In vivo human data compared against the "COSMOS" reference, which is considered the gold standard as it is calculated from 12 different head orientations.
- In Vivo Human Scans: Data from five healthy volunteers (ages 21–25) acquired on a 3T Siemens scanner using a multi-echo GRE sequence.
Results and Discussion
ALOHA-QSM demonstrated significant improvements over existing single-orientation methods such as TKD (Truncated K-space Division) and iLSQR (Iterative Least-Squares).
Artifact Suppression and Detail Preservation
Qualitative analysis showed that ALOHA-QSM effectively eliminated streaking artifacts without the over-smoothing often seen in methods that rely on anatomical priors (like MEDI). It was particularly successful in resolving fine vascular structures and the complex geometry of the cerebral cortex.
Quantitative Accuracy
In comparisons using the QSM2016 challenge data, ALOHA-QSM achieved the highest Pearson correlation ($R^2 = 0.43$) and the lowest Root Mean Squared Error (RMSE = 2.12%) among the tested single-orientation algorithms. This indicates that the method not only produces visually cleaner images but also more reliable physical measurements of tissue susceptibility.
Conclusion
ALOHA-QSM provides a robust solution to the dipole inversion problem in QSM. By utilizing the internal mathematical structure of k-space data through low-rank Hankel matrix completion, it avoids the need for multiple scans or external anatomical information. While the computational time (approximately 156–187 seconds per brain) is higher than simple truncation methods, the gain in image quality and quantitative precision makes it a promising tool for clinical and research neuroimaging.
Summary updated