In quantitative susceptibility mapping (QSM), the relationship between measured off-resonance shifts and underlying tissue susceptibility possesses mathematical singularities. This complicates computational approaches to the QSM problem¹. Currently, the gold standard for QSM utilizes multiple-orientation acquisitions² (where the patient moves to different angles with respect to the static magnetic field direction). This approach is robust to streaking artifacts, but is clinically impractical. Single-orientation datasets are the only clinically viable QSM alternative, but remain vulnerable to streaking artifacts, thus limiting the viability of QSM as a robust clinical imaging alternative. Here, we present a technique, volume-parcellated quantitative susceptibility mapping (VP-QSM), which performs QSM inversion on multiple reduced field-of-view (FOV) off-resonance maps, henceforth referred to as parcels, to create a composite susceptibility map with reduced streaking.

In VP-QSM, distal field information is neglected in the inversion of a target parcel. An assumption of VP-QSM is that sufficient parcel padding can be used to include relevant distal field information to within the numerical accuracy of the applied inversion. To demonstrate the validity of this principle, we can consider a spherically-approximated voxel of susceptibility, which has a well-known analytic dipole dependence. The distance at which this susceptibility source’s induced field perturbation contributes negligibly to the field-offset measurement ultimately determines the amount of necessary additional spatial information outside a target parcel that is needed to account for all local susceptibility sources within the target parcel. This threshold is dependent on field map noise, anticipated maximum tissue susceptibility offsets, and optimization voxel sizes.

In addition to this analytic derivation, a numerical phantom was developed to test the effect of removing distal information with varying amounts of overlap and zero-padding (Fig 1). Optimal parameters derived from these results are applied to multiple parcels covering the full FOV to create multiple local susceptibility maps, which are combined to form a composite susceptibility map. A small overlapping boundary between adjacent parcels was utilized to reduce parcel-combination artifacts (Fig 2).

In-vivo results are shown from a healthy control dataset acquired on a 7T GE Healthcare Discovery 950 scanner. In addition, a cohort analysis of ten brain cancer patients, ten recently-concussed patients, and ten controls scanned on a GE 3T Discovery 750 scanner was performed to compare streaking artifact reduction with VP-QSM. Brains were masked with brain extraction tool³, background field was removed with projection onto dipole fields⁴, and susceptibility maps were generated with MEDI⁵. Streaking artifacts were estimated by subtracting the absolute values of both the volume-parcellated and full FOV susceptibility map. Computational performances for different parcel amounts were performed in MATLAB with up to 12 cores.

Analytic derivations and numerical simulations both showed that roughly 1.0 cm of additional spatial information (red X, Fig 1e) is sufficient to retain quantitative accuracy with reduced FOV regularization. This estimate is provided for the presented 3T imaging data, using a maximum expected susceptibility of 0.70 ppm, 2mm³ voxels, and field noise of 0.1 Hz. This value can change depending on static field strength, field map SNR, and voxel resolution. Computational performance (Table 1) determined 512 equally-spaced parcels provided sufficient tradeoff for computational expense and parcel size, which was used for VP-QSM with the numerical phantom and in-vivo datasets. For the numerical phantom, volume-parcellation provided comparable RMSE to standard full FOV processing (6.2E-6 to 5.9E-6, respectively). An example of streaking reduction feasible with VP-QSM is shown in the 7T in-vivo dataset (red arrows, Fig 3). For the cohort analysis, a noticeable reduction in streaking artifacts was observed for all datasets (Figure 4).Volume-parcellation constrains streaking artifacts to individual parcels, limiting their propagation throughout the volume and improving the quality of the composite susceptibility map. In addition, improved spatial field information within a parcel will improve convergence in the regularization in the target parcel. With sufficient additional spatial information and proper background correction methods, this provides a means to create quantitatively accurate susceptibility maps with limited streaking artifacts.VP-QSM, which is compatible with all existing susceptibility mapping algorithms, allows for improved stabilization of QSM near regions of poor field estimates. In addition, VP-QSM is well-suited for use with distributed computing algorithms for fast computation. Streaking artifacts are often encountered in ultra-high field QSM, as well as in large-cohort studies of clinical or untrained volunteer populations. In addition, automated pipelines are often more vulnerable to streaking due to intermittent errors in phase-unwrapping, background field removal, or brain masking algorithms. VP-QSM offers a mechanism to add robustness to existing QSM studies that are often compromised by these systematic limitations.