Motion-free Abdominal MRI using Manifold Alignment
Xin Chen1, Muhammad Usman1, Christian Baumgartner2, Claudia Prieto1, and Andrew King1

1Division of Imaging Sciences and Biomedical Engineering, King's College London, London, United Kingdom, 2Biomedical Image Analysis Group, Imperial College, London, United Kingdom

Synopsis

We present a novel method based on manifold alignment, which enables reconstruction of motion-free abdominal images throughout the respiratory cycle to better capture respiratory intra- and inter-cycle variations. The proposed method was evaluated on both simulated and in-vivo 2D acquisitions. Based on virtual navigator measurement, the reconstructed dynamic sequence achieved Pearson correlation coefficient of 0.9504 with the ground truth of the simulated dataset. The proposed method enables much richer profile data to be used for self-gating, resulting in less blurring when compared to conventional central k-space self-gating method for the in-vivo acquisition.

Introduction

Respiratory motion remains a problem in free-breathing abdominal MRI, introducing artefacts in the reconstructed images. Self-gating techniques have been proposed to estimate respiratory motion and reconstruct data acquired within a respiratory gating window. Most of these techniques estimate simple 1D respiratory motion from the repeatedly acquired k-space centre [1]. However, respiratory motion is only approximately periodic, so the use of such simple signals for self-gating may only partially correct for respiratory motion. In this work we propose a novel method based on manifold alignment (MA) which permits richer data to be used to identify k-space data acquired at similar respiratory positions. This enables reconstruction of motion-free abdominal images throughout the respiratory cycle to better capture respiratory intra- and inter-cycle variations.

Method

The intuition behind our technique is that respiratory motion correspondences can be established in a reduced-dimensional space (or manifold space (MS)) [2] of the acquired k-space data using MA [3]. This is achieved by dividing the data into groups. Each group contains data that is directly comparable, but acquired at different respiratory positions. A common MS is formed in which data from different groups that were acquired at similar respiratory positions are close together. This alignment of the groups is performed using our recently proposed MA scheme [3], which is based on locally linear embeddings (LLE), and involves no inter-group comparisons of data.

An overview of the proposed technique is shown in Fig. 1. Acquisition is performed using a golden-angle radial trajectory. All k-space radial profiles are evenly divided into $$$L$$$ groups according to their profile angles (Fig. 1, left), where each group contains $$$N$$$ profiles ($$$N>L$$$). For forming the MS only, each profile is Gaussian-weighted ($$$\sigma_1$$$) based on its distance to the centre of the profile. The MA scheme is used to embed all $$$L$$$ profile datasets into the common MS in which data acquired at similar respiratory positions are close together (Fig 1, middle), by minimising a joint cost function [3]:

$$argmin_{\space Y_1...Y_L}(\sum_{l=1}^{L}\phi_{l}(Y_{l})+\mu\sum_{n=1,m=1,m\neq n}^L\sum_{i,j}^{N}U_{ij}^{(nm)}\parallel y_{i}^{(n)}-y_{j}^{(m)}\parallel ^2) $$

where $$$\phi_{l} $$$ is the LLE cost function [4], $$$y^{(n)}_i$$$ represents the MS coordinates of profile $$$i$$$ in group $$$n$$$, $$$Y_l$$$ is a matrix containing the $$$y$$$ for group $$$l$$$, $$$U^{(nm)} $$$ is an inter-group similarity kernel [3], and $$$\mu$$$ is a weighting parameter that balances the intra-group/inter-group terms.

Image reconstruction (at as many respiratory positions as k-space profiles acquired) is performed by selecting the $$$P$$$ highest weighted profiles from each of the $$$L$$$ groups, followed by non-uniform inverse FFT. The weight is a product of two terms: Gaussian-weighted profile distances in MS ($$$\sigma_2$$$) and time ($$$\sigma_3$$$) respectively.

Experiments/Results

Our method was evaluated on simulated and in-vivo datasets. For the simulated dataset, 100 2D sagittal dynamic liver images were acquired. A smooth synthetic dynamic sequence of 10000 images was simulated from this acquisition using B-spline interpolation. One k-space radial profile per image was simulated to mimic a realistic acquisition process. Parameter settings for reconstruction are reported in the caption of Fig. 2.

A 2D radial golden-angle acquisition was performed under free-breathing for ~45s on a healthy subject, resulting in 15000 k-space profiles. Data was acquired on a Philips 1.5T scanner using a 28 channel-coil (b-SSFP, 2x2x8mm resolution, flip angle 70$$$^o$$$, TR/TE=3.08/1.54).

A virtual navigator (VN) was used to measure the superior-inferior translation of the liver-lung boundary on the simulations. When visualising the MS, each profile was colour-coded using the VN value. The resulting MS (Fig. 2) shows that profiles with similar VN values (colours) are embedded close to each other. The VN values for the ground-truth (red) and reconstructed images (blue) of the entire simulated sequence are shown in Fig. 3 (Pearson’s r=0.9504).

In-vivo data was reconstructed with the proposed method. Example reconstructions at 4 different respiratory positions are shown in Fig. 4. A k-space centre self-gating reconstruction was also performed for comparison purposes at mid-inhale (Fig. 5). Less blurring can be observed with the proposed method, especially at small structures.

Discussion

We have presented a novel technique based on MA, which enables reconstruction of motion-free abdominal images throughout the respiratory cycle. The method allows reconstruction of images at as many respiratory positions as k-space profiles acquired and also provides information for retrospective motion correction. The proposed MA-based scheme enables much richer profile data to be used for self-gating, resulting in less blurring when compared to the central k-space gating method.

Acknowledgements

This work was funded by EPSRC grant EP/M009319/1.

References

[1] Stehning et al, MRM, 2005.

[2] Usman et al, MRM, 2014.

[3] Baumgartner et al, IPMI, 2015.

[4] Saul and Roweis, Science Magazine, 2000.

Figures

Fig.1: Overview of proposed manifold alignment technique for self-gating dynamic MRI. Left: all k-space radial profiles are evenly divided into $$$L$$$ groups according to their profile angles. Middle: all profiles are embedded in a lower dimensional manifold space (MS). Right: image reconstruction based on profile distances in the MS.

Fig.2: Manifold space (MS) of the simulated dataset. Each k-space profile is represented as a dot in MS. Profiles acquired at similar respiratory positions (colours) are embedded close to each other. Parameter settings for the proposed method: $$$L=99$$$, $$$N=100$$$, $$$\mu =0.01$$$, $$$P=3$$$, $$$\sigma 1 = 10$$$, $$$\sigma 2=0.005$$$, $$$\sigma 3=1000$$$.

Fig. 3: Virtual navigator values for the ground-truth (red) and reconstructed sequence (blue) of the entire dynamic for the simulated dataset. Pearson correlation coefficient is 0.9504.

Fig. 4: Example of reconstructed images from in-vivo acquisition using the proposed approach at end inhale, mid exhale, end exhale and mid inhale respiratory positions respectively.

Fig. 5: Example of reconstructed images from in-vivo acquisition. Central k-space self-gating (gating window=1/4 respiratory amplitude) reconstruction (left) and the proposed approach (right) at a mid-inhale respiratory position are shown. Less blurring can be observed with the proposed method, especially at small structures (arrows).



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
0784