Although multislice multishot imaging of the brain offers flexible contrast, it is vulnerable to motion artifacts¹. An integrated framework for retrospectively motion corrected reconstruction of multislice multishot MRI in the presence of 3D rigid motion is developed. The method is devised to fuse previous approaches which treat either within-plane^2,3 or through-plane⁴ motion in isolation. The technique is applied to newborn brain imaging in natural sleep, where the subjects tend to move sporadically even when deeply settled.

This proposal extends our previous contribution for within-plane motion correction³ to the multislice case. To that end, the method aims at jointly solve for the image to be reconstructed and a set of 3D motion states for shots and slices using only partial $$$k$$$-space information aided by the sensitivity encoding (SENSE) information of multiple receiver coils:$$(\mathbf{x}^{\ast},\mathbf{T}^{\ast})=\displaystyle\underset{\mathbf{x},\mathbf{T}}{\mathrm{argmin}}\|\mathbf{A}\boldsymbol{\mathcal{F}}_i\mathbf{H}_j\mathbf{S}\mathbf{T}\mathbf{x}-\mathbf{y}\|_2^2+\lambda\|\mathbf{W}_j\mathbf{x}\|_2^2,$$with $$$\mathbf{x}$$$ the image to be reconstructed, $$$\mathbf{T}=\mathbf{T}_{n_i,r_j}$$$ the motion state corresponding to phase-encode (PE) shot $$$n_i$$$ and slice $$$r_j$$$, modelled using linear phase modulations⁵ so as to avoid blurring, $$$\mathbf{A}=\mathbf{A}_{n_i,r_j}$$$ encoding the shot and slice sampling, $$$\boldsymbol{\mathcal{F}}_i$$$ the discrete Fourier transform in the PE direction $$$i$$$ (encompassing applied $$$k$$$-space oversampling or downsampling), $$$\mathbf{H}_j$$$ the slice profile filter, $$$\mathbf{S}$$$ the coil sensitivity matrix, $$$\mathbf{y}=\mathbf{y}_{n_i,r_j}$$$ the measured data for each shot and slice, and $$$\mathbf{W}_j$$$ a second order smoothness promoting matrix to prevent ill-conditioning in the slice direction $$$j$$$, which is regulated by $$$\lambda$$$. The within-plane correction model is illustrated in Fig. 1. The through-plane correction is based on having data acquired with overlapping slices, modeled by $$$\mathbf{H}_j$$$ (as illustrated in Fig. 2), and superresolution (using $$$\mathbf{W}_j$$$). The within-plane and through-plane processes are connected by the masking matrix $$$\mathbf{A}$$$. Additionally, an outlier rejection strategy is introduced that discards those shots for which the ratio of the error to the measured data with respect to the error of corresponding shots in neighbouring slices exceeds a given threshold. The algorithm alternates between the estimation of motion parameters (using the Newton's method⁶) and the estimation of the underlying volume (using conjugate gradient⁷).

The method is applied to correct Cartesian multislice scans acquired from a cohort of neonates with gestational ages ranging from $$$35+1$$$ to $$$42+2$$$ weeks recruited for the developing Human Connectome Project (dHCP) and scanned using a $$$3\,\mbox{T}$$$ PHILIPS ACHIEVA TX (details in Fig. 3). In our pipeline, motion compensated reconstructions are obtained separately for each view and then views are assembled by slice to volume registration⁴. Certain term rearrangements, coil array compression⁸, and GPU computation are used to efficiently treat the huge dimensionality of the problem. Coil sensitity profiles are estimated from a separate reference scan⁹.

Both sagittal $$$T_2$$$ (Fig. 4) and axial $$$T_1$$$ (Fig. 5) examples have been picked out to illustrate the capability to obtain motion-free reconstructions. First, strong artifacts are observed in the uncorrected sagittal view (Fig. 4a) which have been fully resolved in the motion corrected reconstruction (Fig. 4b). Second, there is significant signal variation from slice to slice in the axial and coronal views (Fig. 4a), whereas the motion corrected version produces consistent volumetric information (Fig. 4b). Third, a corrupted shot seems to be spoiling brain structures imaged in the axial view (Fig. 5a) but after motion correction, an artifact-free image can be recovered (Fig. 5b). Fourth, damaged slices are also affecting a large proportion of the sagittal and coronal information (Fig. 5a), with a much more consistent retrieval when motion correction is applied (Fig. 5b).

Regarding aggregated results on the cohort listed in Fig 3, first, motion correction improves the quality of obtained reconstructions in all the cases studied. Second, in most of them motion is fully or almost fully recovered, with $$$130/131$$$ $$$(99.24\%)$$$ $$$T_2$$$s and $$$104/108$$$ $$$(96.30\%)$$$ $$$T_1$$$s showing consistent brain structures after view assembling. These results imply that no scan interruptions or sedation are required in our neonatal scanning protocol. Finally, regarding computation times, reconstructions take from $$$30\,\mbox{m}$$$ to $$$3\,\mbox{h}$$$ depending on the amount of motion, so they can be performed offline in reasonable times.

A methodology is presented that is able to successfully tackle the problem of rigid motion in multishot multislice acquisitions of the brain anatomy. Results on neonates have proven that the application of the proposed framework usually provides a remarkable image quality improvement compared to uncorrected reconstructions without introducing any major side effect. The approach provides a robust and reliable framework for achieving retrospective correction of 3D rigid motion in multishot multislice data, which is the most commonly acquired imaging method for brain MRI. It thus has wide potential application for both clinical and research examinations.