Improving temporal resolution in fMRI using a 3D spiral acquisition and low rank plus sparse (L+S) reconstruction
Introduction
Functional magnetic resonance imaging (fMRI) is a dynamic neuroimaging method that utilizes endogenous Blood Oxygenation Level Dependent (BOLD) contrast produced by deoxyhemoglobin to observe neuronal activity non-invasively in the brain Ogawa et al., 1990, Turner et al., 1993. Rapid whole brain imaging is of particular interest in BOLD fMRI because of increased sensitivity in detecting functional activation by acquiring a greater number of samples Posse et al. (2012), reducing aliasing of physiological noise Frank et al. (2001), and reducing sensitivity to head motion. Higher sampling rates also allow for a finer characterization of the hemodynamic response function (HRF) leading to an increased statistical power and specificity of the BOLD signal. Furthermore, high temporal resolution is useful for various cognitive fMRI studies, such as estimating brain connectivity in resting state networks Beckmann et al. (2005), that are usually detected over a broad cortex during hundreds of milliseconds.
There are numerous proposed approaches for obtaining acceleration in fMRI. One method is to use parallel imaging where the inherent spatial sensitivity information in multiple receivers is used to recover an image from its undersampled aliased version. Application of parallel imaging techniques, such as SENSE Pruessmann et al. (1999) and GRAPPA Griswold et al. (2002), to fMRI with 2D echo-planar imaging (EPI) sequences demonstrate two- to four-fold in-plane acceleration Golay et al. (2000); de Zwart et al. (2002); Afacan et al. (2012). The use of segmented 3D EPI sequences can achieve further acceleration by using coil sensitivities in both in-plane and through-plane phase-encoding directions Poser et al. (2010). Highly undersampled single-shot 3D methods show whole-brain fMRI at extremely fast sampling rates of up to 100 ms Lin et al., 2006, Zahneisen et al., 2012. Simultaneous multi-slice (SMS) imaging has also emerged as a successful single-shot acceleration strategy in fMRI Zahneisen et al., 2014, Setsompop et al., 2012 by acquiring up to 12 overlapping slices the are later separated during reconstruction. Methods have also been proposed to decrease the fMRI acquisition time by exploiting data redundancy in both the spatial and temporal domains. One of the first such methods demonstrated in fMRI, known as UNFOLD Madore et al. (1999), exploits specific distribution of the sampling in relation to the signal aliasing patterns. Parallel imaging approaches have also been developed to incorporate information across temporal frames such as GRAPPA Huang et al. (2005), for example, and have demonstrated improved spatial resolution without significant penalties on temporal resolution with accelerations up to a factor of four.
Compressed sensing (CS) Lustig et al. (2007) techniques are another method for increasing imaging speed in MRI. CS uses sparsity in image space or an appropriate transformed domain with incoherence in the acquisition to reconstruct undersampled data. CS has been successfully applied for accelerating a variety of dynamic MRI applications, such as myocardial perfusion imaging Adluru et al., 2009, Ge et al., 2009 and MR angiography Mistretta (2009), and even resulted in the development of the specific techniques, such as SPARSE Lustig et al. (2006) or FOCUSS Jung et al. (2009). SPARSE exploits both spatial and temporal sparsity of dynamic MRI image sequences using l1-norm minimization of transformed dynamic data (using, for example, wavelet transform along the spatial direction and Fourier transform along the temporal direction) subject to data fidelity constraints. The method is based only on a sparse representation of the dynamic data and does not require an a-priori known spatio-temporal structure or a training set. On the other hand, the FOCUSS utilizes a CS framework using prediction and residual encoding where the prediction approximately estimates the dynamic images and the residual encoding takes care of the remaining signals from a small number of samples. Application of CS to fMRI has been shown to provide improvements in sensitivity Jung and Ye, 2009, Holland et al., 2013, temporal resolution Jeromin et al., 2012, Zong et al., 2014 and spatial resolution Nguyen and Glover, 2014, Fang et al., 2015, Nguyen and Glover, 2014. Like parallel imaging, exploiting temporal sampling redundancy results in further acceleration gains. For example Fang et al. Fang et al. (2015) used a variable density spiral acquisition and a modified SPARSE method, HSPARSE, to achieve high spatial resolution fMRI with acceleration factor of approximately five. Unlike SPARSE the HSPARSE exploits both spatial and temporal redundancy.
Techniques based on low-rank matrix completion are also gaining attention in the scientific community in application to dynamic MRI Zhao et al., 2010, Haldar and Liang, 2010. Often the dynamic information in the imaging matrix can be captured in a few singular values. Minimization of the nuclear norm of the matrix enables the recovery of missing or corrupted entries under low-rank and incoherence conditions. These methods favor large singular values of data that have a smaller number of degrees of freedom allowing for undersampling. Chiew et al. (2015) used undersampled multi-shot 3D EPI combined with low-rank matrix completion, FASTER, to accelerate resting-state fMRI by a factor of approximately four. In a later work, the authors used a 3D radial acquisition combined with SENSE and rank-constrained reconstruction to achieve variable acceleration factors in resting-state and task-based fMRI Chiew et al. (2016). While the application of low-rank matrix recovery techniques was shown to be beneficial for accelerating task-based fMRI, generally the assumption of a low-rank matrix might be too strong to capture the weak BOLD activation signals in addition to the relatively slow-evolving non-activation related functional brain networks.
Recent research proposes combining sparse and low-rank matrix completion techniques in application to dynamic MRI. One approach is to assume that the solution is simultaneously low-rank (L) and sparse (S) Lingala et al., 2011, Zhao et al., 2012. Another approach suggests decomposition of the data matrix as a linear combination of the L and S components, also known as robust principal component analysis (RPCA) Candès et al. (2011) or L+S decomposition Yuan and Yang (2009). Gao et al. applied this method to dynamic computed tomography Gao et al. (2011) and later to dynamic MRI applications, such as retrospectively undersampled cardiac MRI data Gao et al. (2012) and accelerated diffusion-weighted MRI data Compressive, 2013, Compressive, 2013. The method was further developed by Otazo et al. who initially demonstrated the L+S reconstruction approach for dynamic contrast enhanced imaging Otazo et al. (2013) and later for a variety of other clinical MRI applications, such as cardiac perfusion, cardiac cine, time-resolved angiography, and abdominal and breast perfusion MR imaging Otazo et al. (2015a). In parallel with Otazo's work, Tremoulheac et al. proposed an L+S model decomposition based on an alternating direction method of multipliers for cardiac MRI, called RPCA Trémoulhéac et al. (2014).
Most recently, Otazo et al. demonstrated the feasibility of L+S decomposition for separation of subsampled physiological noise in resting-state fMRI Otazo et al. (2015b). Unlike using low-rank matrix completion alone for modeling fMRI data, L+S assumes that activation is modeled as periodic and sparse and the low-rank component models the slowly correlated brain background. The L component can then potentially serve as a means of removing unwanted background signal from the BOLD activation. Otazo et al. showed that L+S decomposition is comparable with methods for retrospective removal of physiological noise performed by linear regression of nuisance signals due to cardiac pulsation and respiratory motion. Singh et al. (2015) also applied the L+S model to recover retrospectively undersampled block-design fMRI signal and showed that the low-rank matrix captures the temporally static T2*-weighted images while the sparse component captures the pseudo-periodic BOLD signal for an acceleration factor of three. However, the application of L+S model to prospectively accelerated fMRI data was not evaluated.
In this work we present accelerated fMRI using L+S reconstruction in both retrospectively and prospectively undersampled data acquisitions. We propose to accelerate fMRI using a partially acquired 3D stack of spirals (SoS) sequence where undersampling is performed in the domain by fully sampling the central z-phase encoding partitions and randomly excluding remaining partitions. We demonstrate that the L+S model is a good representation for fMRI data using blocked and potentially slow event-related designs. In these experiments, the time series of the activated voxels is a convolution of the BOLD HRF with a periodic stimulus. As a result the BOLD signal will be sparse in the Fourier transformed temporal domain. Additionally, non-task related signals that change slowly over time will be predominantly low-rank allowing for the possibility of removal from the BOLD time course. The results for retrospectively and prospectively undersampled fMRI experiments demonstrate that the L+S algorithm achieves good reconstruction of the aliased functional images with accurate decomposition into temporally correlated brain background in the L component and dynamic information with underlying BOLD signal in the S component with a reduction factor of four without using parallel imaging.
Section snippets
Low-rank plus sparse matrix decomposition
The fMRI time-series of images is modeled by a matrix M where each column represents a temporal frame. The L+S algorithm then divides this matrix M into two distinct matrices, a low-rank matrix (L) characterized by few non-zero singular values and a sparse matrix (S) with a few non-zero entries. This decomposition is well-posed if two conditions are satisfied: 1) L does not have a sparse representation and 2) S does not have low-rank Candès et al. (2011). These conditions are referred to as the
Retrospectively undersampled fully sampled fMRI data
Fig. 3 shows L+S reconstruction results of 12 representative slices from the fully sampled fMRI data (image B, top row) of a healthy adult volunteer in the vicinity of visual cortex that were also retrospectively undersampled with and the three different sampling patterns. The B images in the and undersampled data sets represent the zero-filled reconstructions of the selective temporal frames where aliasing artifacts are clearly seen as blurring. The reconstructed images are
L+S reconstruction of undersampled fMRI
In this work we demonstrated an undersampled SoS fMRI acquisition using a low-rank plus sparse (L+S) decomposition model. As mentioned earlier, Otazo et al. established a precedent for using the L+S model in resting-state fMRI and Singh et al. demonstrated successful application of the L+S model to block-design task fMRI. However, both of these studies were limited to retrospective undersampling. In this research we show an undersampled fMRI acquisition with L+S reconstruction using true
Conclusion
We show that L+S model can be successfully applied to accelerate fMRI acquisition for up to a factor of without using parallel imaging. The method exploits inherent temporal low-rank and sparsity features of fMRI data. The results indicate the promising ability of the method to uniquely decompose the problem into slow temporally-correlated brain background images in the L component and capture BOLD signal in the S component by exploiting sparsity in the temporal Fourier transform domain.
Conflict of interests
The authors have no conflict of interests. All authors agree to submission.
Role of the funding source
Work supported by the following grants: NIH (R01DA019912, K02DA020569), NCRR (G12-PR003061, P20-PR011091).
References (60)
- et al.
Analysis of fmri time-series revisited
Neuroimage
(1995) - et al.
Three dimensional echo-planar imaging at 7 tesla
Neuroimage
(2010) - et al.
Enhancement of temporal resolution and bold sensitivity in real-time fmri using multi-slab echo-volumar imaging
NeuroImage
(2012) - et al.
Simultaneous multi-slice fmri using spiral trajectories
NeuroImage
(2014) - et al.
Compressed sensing fmri using gradient-recalled echo and epi sequences
NeuroImage
(2014) - et al.
Acquisition and reconstruction of undersampled radial data for myocardial perfusion magnetic resonance imaging
J. Magn. Reson. Imaging
(2009) - et al.
Rapid full-brain fmri with an accelerated multi shot 3d epi sequence using both unfold and grappa
Magn. Reson. Med.
(2012) - Beckmann, C.F., Smith, S.M., 2004. Probabilistic independent component analysis for functional magnetic resonance...
- et al.
Investigations into resting-state connectivity using independent component analysis
Philos. Trans. R. Soc. Lond. B: Biol. Sci.
(2005) - et al.
Statistical methods of estimation and inference for functional mr image analysis
Magn. Reson. Med.
(1996)
A singular value thresholding algorithm for matrix completion
SIAM J. Optim.
Robust principal component analysis?
J. ACM (JACM)
k-t faster: acceleration of functional mri data acquisition using low rank constraints
Magn. Reson. Med.
Accelerating functional mri using fixed-rank approximations and radial-cartesian sampling
Magn. Reson. Med.
An iterative thresholding algorithm for linear inverse problems with a sparsity constraint
Commun. pure Appl. Math.
Application of sensitivity-encoded echo-planar imaging for blood oxygen level-dependent functional brain imagingü
Magn. Reson. Med.
High spatial resolution compressed sensing (hsparse) functional mri
Magn. Reson. Med.
Multiplexed echo planar imaging for sub-second whole brain fmri and fast diffusion imaging
PLoS One
Estimation of respiration-induced noise fluctuations from undersampled multislice fmri dataü
Magn. Reson. Med.
Robust principal component analysis-based four-dimensional computed tomography
Phys. Med. Biol.
Myocardial perfusion mri with sliding-window conjugate-gradient hypr
Magn. Reson. Med.
Presto-sense: an ultrafast whole-brain fmri technique
Magn. Reson. Med.
Generalized autocalibrating partially parallel acquisitions (grappa)
Magn. Reson. Med.
Compressed sensing reconstruction improves sensitivity of variable density spiral fmri
Magn. Reson. Med.
k-t grappa: a k-space implementation for dynamic mri with high reduction factor
Magn. Reson. Med.
Selection of a convolution function for fourier inversion using gridding
IEEE Trans. Med. Imaging
Cited by (13)
Subspace-constrained approaches to low-rank fMRI acceleration
2021, NeuroImageCitation Excerpt :Methods that do jointly consider k-space and time are known as k-t methods and can be broadly separated into three categories: methods that make a strong assumption about the spatiotemporal structure (Madore et al., 1999; Tsao et al., 2003; Huang et al., 2005; Yun et al., 2013), methods that make a strong assumption about sparsity within a pre-defined basis set (compressed sensing approaches) (Lustig et al., 2007; Holland et al., 2013; Jeromin et al., 2012; Zong et al., 2014; Chavarrías et al., 2015), and methods that assume the data fits a globally low-rank model (Liang, 2007; Chiew et al., 2015). There are also approaches which combine these methods (Chavarrías et al., 2015; Pedersen et al., 2009; Jung et al., 2009; Qin et al., 2019; Otazo et al., 2015; Petrov et al., 2017). By focusing on redundancies or structural features in k-t space, k-t methods have the potential for much greater degrees of acceleration than time-independent methods due to the extra dimension of shared information.
R-fMRI reconstruction from k–t undersampled data using a subject-invariant dictionary model and VB-EM with nested minorization
2020, Medical Image AnalysisCitation Excerpt :In contrast, our framework models the structure of the R-fMRI image through a learning-based approach, to learn a data-adaptive dictionary, a more effective linear (sparse) representation of the data than analytically-defined transforms like the discrete cosine transform or wavelet transform. Some reconstruction methods for R-fMRI and task-based fMRI use non-Cartesian undersampling patterns in k-space (Chiew et al., 2016; 2018; Fang et al., 2016; Graedel et al., 2017; Lazarus et al., 2019; Petrov et al., 2017) to speedup acquisitions, without proposing undersampling in time. Although non-Cartesian undersampling can give higher acceleration factors, it can be non-trivial to implement on clinical scanners and may be prone to specific artifacts in the reconstructions.
Improved statistical efficiency of simultaneous multi-slice fMRI by reconstruction with spatially adaptive temporal smoothing
2019, NeuroImageCitation Excerpt :Alternatively, low-rank (space x time) methods have been proposed for use in fMRI based on the use of low-dimensionality representations in fMRI analysis methods (Chiew et al., 2016, 2015). While low-rank approaches do not impose any specific constraint on the representation of spatial or temporal information, more recent low-rank plus sparse methods have also been used to additionally exploit sparsity in the temporal Fourier domain (Aggarwal et al., 2017; Petrov et al., 2017; Singh et al., 2015; Weizman et al., 2017) which requires smoothness or periodicity in voxel time-courses. Here, we introduce an improvement to SMS-EPI for fMRI by introducing time-varying sampling with a spatially adaptive temporally regularized reconstruction.
Recovering task fMRI signals from highly under-sampled data with low-rank and temporal subspace constraints
2018, NeuroImageCitation Excerpt :While simultaneous multi-slice imaging has emerged as a popular successor to multi-slice EPI, in recent years, a number of different strategies have been proposed for accelerating fMRI data acquisition, not solely dependent on coil-sensitivity encoding. Some examples that leverage compressible representations of fMRI data in some way include compressed sensing (CS) using spatial wavelet or temporal spectral sparsity (Jung and Ye, 2009; Jeromin et al., 2012; Holland et al., 2013; Zong et al., 2014), partially separable function (PS) modelling (Liang, 2007; Lam et al., 2013; Nguyen and Glover, 2014), low-rank modelling (LR) (Chiew et al., 2015; Chiew et al., 2016), and most recently low-rank and sparse decompositions (L + S) (Singh et al., 2015; Petrov et al., 2017; Aggarwal et al., 2017; Weizman et al., 2017). With the exception of the use of spatial wavelet CS, all these methods move away from time-independent reconstruction of 3D volumes, leveraging temporal structure in the fMRI data, as they effectively seek to fit reconstruction models with fewer free parameters to enable reconstruction in the presence of under-sampling.
Non-Cartesian Non-Fourier FMRI Imaging for High-Resolution Retinotopic Mapping at 7 Tesla
2023, 2023 IEEE 9th International Workshop on Computational Advances in Multi-Sensor Adaptive Processing, CAMSAP 2023