MRI is a promising alternative to CT for lung exams due to the lack of radiation exposure and possibility for multi-contrast functional assessment. Challenges associated with lung MRI include respiratory motion and low SNR due to the short T₂* of lung parenchyma. Sequences with ultra-short echo time (UTE) have been shown to be an effective way to provide sufficient signal in lung, and encouraging results with isotropic spatial resolution, comparable to that of CT, have been reported ^1-3. Respiratory motion, however, still remains a major challenge for lung MRI. Although external devices can be used to monitor respiratory motion for gated image acquisitions ⁴, the scan efficiency is reduced and a cumbersome setup is required. Sparse imaging techniques have become a powerful approach for rapid MRI ⁵; and in addition to increasing imaging speed, sparsity can also be used to resolve respiratory motion by reconstructing an extra motion dimension ⁶. This technique, called XD-GRASP (eXtra-Dimensional Golden-angle RAdial Sparse Parallel MRI), combines the self-navigation properties of radial sampling and the acceleration capability of sparse sampling and reconstruction. The purpose of this work is to propose a continuous 4D (x-y-z-respiration) respiratory motion-resolved lung MRI framework combining XD-GRASP with a 3D golden-angle radial UTE sequence.

(i) Data Acquisition and Motion Detection: IRB-approved lung MRI was performed in 5 healthy volunteers (28±2.3-year) during free-breathing without external gating on a 3T clinical scanner (MAGNETOM Prisma, Siemens Healthcare, Erlangen Germany). Center-out half-projection radial k-space data were continuously acquired using a prototype 3D UTE sequence ³ with golden-angle rotation scheme based on the spiral phyllotaxis pattern ⁷ (Fig.1a). Imaging parameters included: TR/TE=3.3/0.05ms, FOV=(250mm)³, matrix size=256³, voxel size ~(1mm)³, and RF excitation angle=6^o. A total of 122,520 spokes were acquired in each subject in 8 minutes, including 2042 golden-angle interleaves. Each interleave started with a half-projection oriented along the superior-inferior (SI) direction (red lines, Fig.1a) for self-navigation and was preceded by CHESS fat saturation. In order to test the acceleration capability, a separate 5-minute acquisition was also performed in one volunteer with identical imaging parameters but with 1377 golden-angle interleaves only. The missing half of the SI half-projections were zero-filled to generate z-directional projection profiles, and respiratory motion was extracted (Fig.1b) using the approach described in ^{6, 8}.

(ii) Data Sorting and XD-GRASP Reconstruction: The continuously acquired k-space data were sorted into 4 respiratory phases, spanning from expiration to inspiration, using the estimated respiratory motion signal. XD-GRASP reconstruction was performed on the sorted undersampled 4D dataset (k_x-k_y-k_z-respiratory) by solving

$${d=\min_{d}\parallel{F\cdot{C\cdot{d}}-m\parallel}_2^2}+\lambda\parallel{S\cdot{d}\parallel}_{1}$$

where F represents the NUFFT operator, C coil sensitivities, d the 4D image to be reconstructed (size=256x256x256x4), and m the sorted radial k-space data. S is the sparsifying transform (first-order finite differences) applied along the respiratory dimension with regularization parameter λ (empirically selected). For comparison, standard NUFFT reconstruction without motion-sorting was also performed.

(iii) Image Quality Assessment: A chest radiologist blinded to the reconstruction scheme was presented a pair of datasets from each of the 5 subjects, one for XD-GRASP (end-expiratory phase only) and the other one for NUFFT, ordered in a random fashion. The radiologist evaluated which data set (left or right) produced higher overall image quality, lower motion artifact, and higher sharpness of diaphragm and pulmonary vessels. The assessment was performed on both transverse and coronal orientations on a 5-point scale: -2: left>>right; -1: left>right; 0: left=right; 1: left<right; 2: left<<right. The reported scores were rearranged to a new 5-point scale: -2: NUFFT>>XD-GRASP; -1: NUFFT>XD-GRASP; 0: NUFFT=XD-GRASP; 1: NUFFT<XD-GRASP; 2: NUFFT<<XD-GRASP. Note here >> corresponds to “much better than”, > to “better than”, = to “equal to”, < to “worse than”, << to “much worse than”.In all the subjects, XD-GRASP achieved systematic improvement in overall image quality and sharpness, and systematic reduction in motion artifacts comparing to NUFFT reconstruction without motion sorting (Fig.2). Fig. 3 shows a particular example, where a suspected lung nodule can be better detected in XD-GRASP (yellow arrows, Fig.3b). The yellow dash lines in Fig.4 show the displacement between consecutive respiratory states, suggesting that respiratory motion was resolved by reconstructing an extra motion dimension. The scan time may be further reduced to ~5 minutes without visible loss of image quality (Fig. 5). This work proposes a framework for continuous respiratory motion-resolved 4D lung MRI with 100% scan efficiency, high isotropic spatial resolution and scan times of ~5-8 minutes. The method allows for assessment of lung structure in arbitrary orientations at different respiratory motion states, and can be potentially used for simultaneous morphologic and functional assessment of the lung during respiration.