Real-time motion correction for MRI using fat navigators

FIELD OF THE INVENTION

The present invention relates generally to the area of medical imaging. More specifically the invention relates to real-time motion correction during magnetic resonance imaging (MRI), using MR navigators with signal content from fat.

BACKGROUND

Medical imaging is a very important field when it comes to medical examination, medical procedures and diagnostics. Medical imaging is the technique and process used to create images of the human body (or parts and function thereof) for clinical purposes (medical purposes seeking to reveal, diagnose, or examine disease) or medical science (including the study of normal anatomy and physiology). Some medical imaging techniques creates a two dimensional (2D) image of a thin "slice" of the body and are therefore referred to tomographic imaging techniques, such as Computer tomography (CT) and Magnetic resonance imaging (MRI). Magnetic resonance imaging, or nuclear magnetic resonance (NMR) imaging as it is also known, is one of the most important medical imaging techniques used today. The MRI scanner uses a powerful magnet to polarize and excite hydrogen nuclei (single proton) in water and fat molecules in human tissue, producing a detectable signal that is spatially encoded, resulting in images of the body. The MRI scanner emits an RF (radio frequency) pulse that specifically excites the hydrogen atoms in soft tissue such as water and fat. The system sends the pulse to the area of the body to be examined. The pulse makes the protons in that area absorb the energy needed to make them spin in a different direction. This is the "resonance" part of MRI. The RF pulse makes them (only the one or two extra unmatched protons per million) spin at a specific frequency, in a specific direction. The particular frequency of resonance is called the Larmor frequency and is calculated based on the particular tissue being imaged and the strength of the main magnetic field. MRI uses three electromagnetic fields: a very strong (on the order of units of Tesla) static magnetic field to polarize the hydrogen nuclei, called the static field; a weaker time-varying (on the order of 1 kHz) field(s) for spatial encoding, called the gradient field(s); and a weak radio -frequency (RF) field for manipulation of the hydrogen nuclei to produce measurable signals, collected through an RF antenna (or RF coil).

MR images may be created by applying currents to the gradient and RF coils according to known algorithms called "pulse sequences." The selection of a pulse sequence determines the relative appearance of different tissue types in the resultant images. Various properties of tissue may be used to create images with a desirable contrast between different tissues.

One technique that has been developed to accelerate MR data acquisition is commonly referred to as "parallel imaging" or "partial parallel imaging". In parallel imaging, multiple receive coils acquire data from a region or volume of interest, where the data is under-sampled, for example, in one direction so that only a fraction of k-space (i.e. the MR raw data) is acquired in an image scan. Thus, parallel imaging is used to accelerate data acquisition in one or more dimensions by exploiting the spatial dependence of phased array coil sensitivity. Parallel imaging has not only been shown to be successful in reducing scan time, but also reducing image blurring and geometric distortions, the latter also being important for the inventions described here. Moreover, parallel imaging can be used to improve spatial or temporal resolution as well as provide increased volumetric coverage.

MRI is widely used for creating images of many parts of the body, such as the abdomen and the head (brain), with over half of the examinations worldwide made in the Central Nervous System (CNS) including the brain and the spine. When scanning a subject using MRI it is very important that the subject remains motionless during the scan to get a sharp image. Motion of the subject being imaged may degrade image quality, for example in medical imaging. Thus, motion of the subject during the scan may corrupt the image and reduces the quality of the image, which is referred to as image/motion artifacts. Motion artifacts are produced by movement of the object and involuntary motion in mammals encountered in medical imaging systems is a common source of image artifacts. The involuntary motion may lead to errors, such as when a physician is determining the size of a lesion, determining the location of the lesion, or quantifying the lesion. Therefore, motion artifacts during the scan may lead to a missed or misinterpreted diagnosis. When performing brain MRI, head motion of the patient is problematic, especially due to the long scan times (2-7 minutes). To prevent motion, the clinical MR staff put small cushions on the sides of the patient's head. While this reduces the risk of left-right motion, it is more difficult to restrain motion in the "nodding direction" (sagittal plane). Since deep breathing also gives rise to nodding motion, it is the major source of motion artifacts in brain MRI scans.

Different methods for addressing the problem of motion artifacts during MRI have emerged. Some methods contain external objects to keep the patient more immobile and thus prevent motion, while other methods instead involve techniques that try to correct for motion during (prospectively) or after the scanning (retrospectively). One existing such motion correction imaging technique is PROPELLER [1], which is able to correct for head motion retrospectively. However, this correction is only performed within the image plane, and while most MR images are taken in the transverse (axial) plane, the sagittal plane/through-slice (nodding direction) motion becomes perpendicular to the image plane and can therefore not be corrected retrospectively using this method. A general, but still experimental, solution to this involves using additional external hardware such as video cameras inside the MR tunnel that observe a marker on the patient's head and continuously inform the MR system to update its coordinate system to follow the patient's head movements in 3D. However, with this hardware, there is the problem of extra hardware cost and maintenance, compatibility with other medical cables and equipment during data acquisition, and not the least this requires that the video camera always have a line of sight to the marker on the patient's forehead. This is not trivial with the RF coil that is placed around the patient's whole head with only limited gaps. Moreover, the use of markers on a patient's forehead is not ideal from a practical clinical point of view, and if the patient is changing its facial expression, the marker may move while the brain remains still. To correct for head motion during the MR scanning without the use of external hardware, another way of performing prospective motion correction exists - via acquiring extra MR-data (a.k.a. navigators or navigator echoes), which are used to estimate the amount of motion and continuously update the MR scanner. Such extra MR-data, navigators, in some way cover the center region of k-space. As known in the art, k-space is the standard name for the MR raw data space that corresponds to the Fourier transform of the image. One common type of navigator is a 2D navigator that samples the center portion of k-space completely, from which one can obtain a low-resolution 2D image of the object and may thereby use this information to "navigate" or "determine" where the object (e.g. the patient's head) is located. Another type of navigator is the ID navigator, or pencil beam navigator, which is often used for body and cardiac MRI applications to determine where the position of the edge of the top of e.g. the liver is located in the head-feet direction. Here a special type of RF pulse excites a narrow cylinder-shaped volume in the head- feet direction and ID spatial information along this cylinder is obtained by Fourier transforming the raw data read out from this cylindrical volume. ID navigators are of no use for brain MRI applications, since head motion is not one-dimensional and involve significant rotations. For brain applications, 2D or 3D navigators are therefore necessary to capture the motion of the head. However, known methods to acquire full 3D information requires prohibitively long additional acquisition times. One may perform 2D navigation in either k-space or the image space with the same intended result; to find one rotation and two translation parameters. However, by first Fourier transforming the raw 2D navigator (which is recorded in k-space) to the image domain, the motion estimation algorithm can also select spatial (anatomical) regions that are more important, or trustworthy, than others. Performing the motion estimation in k-space, the data cannot be spatially resolved or excluded and also unwanted anatomical regions residing in the 2D navigator plane are included in the motion estimation process.

One highly successful prospective motion correction method is a technique called "PROMO" [2]. The PROMO technique involves prospective motion correction in MRI by realtime adjustment of the imaging pulse sequence. Without using any additional optical hardware, the PROMO technique is an extra software module that is added to the imaging acquisition and consists of three RF excitations and spiral readouts to obtain three orthogonal low-resolution 2D images of the brain (snapshots). With these images, the amount of head motion is estimated and fed back to the MR scanner, so that the prescribed slice locations are updated in near real-time to follow the patient's head. The PROMO navigator requires 48 ms [2] of additional sequence time played out several times per second, which will reduce the scan efficiency to a certain degree depending on the main pulse sequence. Moreover, the three orthogonal planes excited by the PROMO module consume some longitudinal water proton magnetization, which, depending on the type of main pulse sequence, may affect the diagnostic images acquired.

A problem when taking snapshots during a diagnostic scan is that the resulting MR image could be affected. As the water magnetization is primarily important for the high-resolution and slow diagnostic MRI scan of the brain, it is necessary that the navigator image does not affect this magnetization. Using an RF pulse that excites water when acquiring the navigator image might thus lead to dark bands in a (semi-orthogonal) diagnostic high-resolution image. These darker bands may make the interpretation of the images more difficult, and ultimately lead to misdiagnosis. Therefore, there is a need for new methods for prospective motion correction during brain MRI, which does not affect the brain water magnetization and thus the diagnostic MRI image.

Using excitation of fat in different MRI applications is known. In the art of MRI, fat navigation is sometimes called FatNav. In the paper of Max O. Kohler et al. [3] a fat-selective pencil-beam navigator is proposed for real-time monitoring and compensation of through-plane motion in MR-guided high-intensity focused ultrasound therapy of abdominal organs. Herein is described a 1 D "pencil-beam" fat-navigator that compensates for motion along one direction only. Hence, it cannot handle any rotational motion or 2D translational ( in-plane) motion. In general, all uses of 1 D navigators (most of them carrying both water and fat signal, but here only the fat signal) are for body and cardiac applications where the primary purpose is to synchronize the acquisition ith the breathing cycle. One approximates the breathing motion normally to only occur in the head-feet direction, and hence applies the 1 D navigators in that direction. Typically, these 1 D navigators do not update the coordinate system to follow along with the anatomy of interest. Rather, these I D navigators for body applications are used for monitoring the breathing cycle and tissue motion, and in a periodic manner allow/disallow (aka triggering or gating) data acquisition (so that the anatomy will be recorded in the same phase of the breathing cycle) without updating any logical coordinate systems. In the case of brain MRI, 1 D nav igators would not be useful even if modified to update the coordinate system instead of being used to control the data acquisition window ("gating"). This is because of the rigid body motion that occurs in the brain, involving rotations and translations in more than one direction. Thus, the 1 D pencil-beam navigator mentioned here would not be applicable for brain RI.

Patent application WO09134820 discloses cardiac fat navigators that track the motion of the heart by using a spatial-spectral excitation of the epicardial fat. This work is similar to Kohler et al. [3], by the use of ID fat navigators, and hence would not be suitable for brain applications.

Kawaji et al. [4] shows a 6 mm thick 2D image based fat navigator for cardiac applications, to depict the coronary artery. They are using an imaged based navigator and can select a spatial region of interest, similar to PROMO, however for cardiac MRI. The drawback with their method is that only a single k-space line was acquired after each fat-selective RF pulse, resulting in a single 2D FatNav image acquisition time of 300-400 ms. This long navigator time implies that motion can occur during the acquisition of their 2D FatNav, but more importantly, that the use of this 2D FatNav will consume a very large portion of the total acquisition when used alongside a diagnostic main pulse sequence. Finally, while the method of Kawaji is prospective (real-time) in nature, it is only used to control when diagnostic image data should be acquired (similar to Kohler et al. [3]).

The abstract of N. Ari and R.A. Kraft [5] describes fat navigators for fMRI in the brain, and investigates if it is possible to track the head motion using the fat signal only. The authors compare to motion estimation using the water signal. The major drawback of this method is that they attempt to detect motion in k-space (the Fourier transform of the image = the raw data) and never form an image. While it is possible to detect motion in k-space (see e.g. [1]), the problem is that they are unable to exclude the fat signal from anatomical structures that are not moving rigidly as the brain does. The reason for this is that they are operating in k-space, where each and every data point is a small contribution to the entire image (the Fourier transform relationship), and hence it is not possible to spatially mask out signal that is of no interest - or even behaves like decoys in this regard, in particular so in the back of the neck that often contains more fat than the scalp, but yet is not undergoing the same motion as the brain.

Gallichan et al. [6] has shown that since fat signal in the brain is sparse in the image domain, it can be undersampled and reconstructed using Compressed Sensing (CS) techniques, provided that the data collection is semi- or pseudo-randomly acquired. Gallichan used a 10-minute high- resolution 3D scan, from which radial data was synthesized to mimic raw data from an actual 3D radial scan. This highly undersampled 3D radial synthetic fat navigator (had been acquired in real-time) would have taken 750 ms to acquire, leaving very little room for the diagnostic imaging portion. With CS -reconstruction, an acceleration factor of 50 was found possible. However, CS -reconstruction is also a slow and approximate process, not well suited for real-time applications.

In summary, for brain applications, any k-space based (in 2D or 3D) motion detection method will be biased by non-rigidly moving structures in the imaged area (patient), and ID navigation is not feasible due to the lack of ID type of motions in the brain.

SUMMARY OF THE INVENTION

The general purpose of the invention is to provide new and better methods for motion correction during medical imaging.

A primary object of the invention is to provide new and better methods for motion correction during brain MRI using new navigators.

The current invention relates to novel methods for real-time motion correction using only the MR signal from fat. In the head, the MR visible fat signal resides outside the skull, around the eyes, and at the back of the neck, the latter of which is not moving rigidly with the brain (and should not be included for motion estimation). The methods are based on the principle of acquiring low-resolution snapshot images, a.k.a. 2D or 3D navigators in this context, of the head in parallel with acquisition of the high-resolution diagnostic MRI scan, and using these snapshots to track motion of the head (change of head orientation) for real-time correction of the scan plane. As mentioned above, a problem when taking snapshots during a diagnostic scan is that the resulting MR image could be affected. To obtain a snapshot image, the entire low-resolution image data for one slice is read out very quickly using e.g. an echo-planar imaging (EPI) readout. The signal may contain data both from water and fat, or just one of water or fat, depending on the leading RF excitation pulse used. The current invention provides a solution to the above mentioned problems of using the signal from water when acquiring the navigator, by using new types of navigator images that rely only on the fat signal.

For the navigators described in the present invention the snapshots are based on a fat- selective RF excitation only. For brain MRI, fat is present only in the scalp (which is diagnostically uninteresting), and around the eyes (where the presence of fat signal around the optic nerve is undesired). Hence, these fat-selective RF excitations do not alter the diagnostic MR image or affect the diagnostic quality or confidence. The fat navigator is implemented as a software plugin to the existing MRI pulse sequence (i.e. image acquisition methods) without need for additional hardware, such as external cameras etc.

In the present invention, fat-selective two-dimensional (2D) or three-dimensional (3D) navigators are used, in which the snapshots are based on fat-selective excitation only. Tracking of head motion is thus performed using fat magnetization only. The method is called 2- dimensional or 3-dimensional Fat-Navigation, or 2D FatNav and 3D FatNav, respectively.

An object of the invention is a method for reducing motion related imaging artifacts in MRI acquired images, said method comprising the use of the new fat navigator. Another object of the invention is a method for generating a magnetic resonance image by acquiring MR data and applying a prospective motion correction method to the MR data, in which prospective motion correction method according to the new fat navigator, 2D or 3D FatNav, is being used.

Yet another object of the invention is to provide new navigators for medical imaging, such as brain imaging using MRI. These navigators could either be 2D or 3D fat-selective navigators.

A further object of the invention is the use of new navigators, for motion correction during MRI, preferably brain MRI. These navigators could either be 2D or 3D fat-selective navigators.

Another object of the invention is a motion correction tool to be used as a plug-in module for existing imaging techniques. The tool compensates for motion, thus improving the quality of the diagnostic MRI scans. The tool of the invention could for example be designed as an optional plugin for existing MRI software.

An additional object of the invention is an MRI system, comprising a resonance assembly, an RF transceiver system and a controller programmed to acquire MR data and applying a prospective motion correction method utilizing the new fat navigator, and generating an image. The invention is described in detail below. The examples and experimental details are disclosed to provide an improved understanding and guidance for those skilled in the art. Other objects and advantages of the present invention will become obvious to the reader and it is intended that these objects and advantages lie within the scope of the present invention. For the avoidance of doubt, the description of a feature as an Object' of the invention does not necessarily imply that the object is achieved by all embodiments of the invention. Also, it is to be understood that the phraseology and terminology employed herein are for the purpose of the description and should not be regarded as limiting.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 shows an overview of the proposed prospective use of fat navigator for motion tracking.

Figure 2 shows a combined pulse sequence diagram (a), with a leading fat navigation sequence module (left) followed by a main pulse sequence (right). The corresponding image planes for the 2D sagittal fat navigation module and high-resolution axial diagnostic data are shown in b).

Figure 3 shows the ability of successfully reconstructing data with high acceleration factors when fat-signal, instead of water- signal, is used. White arrow marks an area with image reconstruction errors due to an acceleration factor R = 8.

Figure 4 shows that using fat navigation with high acceleration factors can result in motion estimates with low errors, since geometric distortions are minimal.

Figure 5 shows the proposed 3D fat navigation and the large amount of raw data that can be removed from a fully sampled 3D fat navigation dataset (a) and still be able to reconstruct a FatNav image volume used for motion detection (b) by means of a standard 3D GRAPPA technique. In c), an excellent agreement of motion estimates is shown (full data vs. 32-fold reduced data), d) The 32-fold accelerated 3D EPI pulse sequence for 3D fat navigation.

Figure 6 shows the Collapsed fat navigation variant that can optionally use the fat-signal from the fat-sat RF pulse from the main sequence. This makes the Collapsed fat navigation module short in duration.

DETAILED DESCRIPTION

The general purpose of the invention is to provide new and better methods for motion correction during Magnetic Resonance Imaging (MRI).

To clarify the scope of the invention further, we hereby define some of the components in more detail.

The term "magnetic resonance imaging" refers to a diagnostic medical method and/or a device (magnetic resonance scanner) that is using a static magnetic field, magnetic field gradients and radio frequency (RF) pulses to excite and record the magnetized signal from the protons of Hydrogen nuclei residing in water and fat in the body to produce tomographic images (or slices) of the body. Magnetic resonance imaging is abbreviated MRI. The term "magnetic resonance image" refers to an image being acquired during an MRI scan using an MRI scanner.

The term "EPI readout" refers to one type of fast (or snapshot) MR image acquisition technique (Echo Planar Imaging), where within a fraction of a second enough MR raw data may be acquired to allow for the reconstruction of an MR image. 2D EPI collects raw data corresponding to one spatial plane or projection (c.f collapsed 3D variant), while 3D EPI collects raw data corresponding to a 3D image data volume covering the volume of interest.

The term "motion correction" refers to a software driven process using an algorithm that in a first step is comparing a plural of different datasets to calculate the amount of motion that has occurred between the datasets, and in a second step uses this information to either correct this data, or another set of datasets or images subject to the same motion, or this information can be used to inform some image acquisition process on the amount of motion that has occurred/is occurring.

The term "prospective motion correction" refers to a motion correction process that occurs while the motion and the data acquisition occurs, rather than after the data acquisition is complete. A prospective motion correction framework consists of a quickly repeated loop where received data is fed to a motion detection process that soon thereafter informs the image acquisition process about the amount of motion, after which the image acquisition process takes some action based on this information.

The term "navigator" refers to MR raw data that is being collected to measure some kind of motion occurring in the patient. ID-navigators are detecting motion along a single straight direction. 2D navigators correspond to MR-data that may form a 2D plane in k-space, which can be Fourier transformed to an image. From a 2D navigator, translational motion (2 parameters, x and y) within this plane and one rotational motion (around an axis perpendicular to this plane) can be detected or estimated, either in k-space domain or in the image domain. 3D navigators can address motion with 6 degrees of freedom (three translational directions and three rotational axes) corresponding to a rigid body movement.

The term "fat-navigator" refers to a navigator that uses the signal fat magnetization only in the body to achieve the purpose of navigation (motion detection). The term "fat-selective excitation" refers to an radio-frequency (RF) pulse in MRI that has a carrier frequency that is centered around the magnetic resonance peak of fat and has a transmit bandwidth that is so narrow that it only induces magnetic resonance (excites) in the fat peak without affecting the water magnetization. The term "fat magnetization" or "water magnetization" refers to the net magnetization of the hydrogen protons in fat and water molecules, viewed on a macroscopic scale.

The term "image domain" refers to the data domain that has units of meters in which the final MR image resides.

The term "k-space" refers to the data domain that is the Fourier transform of the image domain (the Fourier transform of the final MR image) and has units of 1/meter, and is the data domain in which MR raw data is being initially stored and recorded.

The term "brain MRI" refers to a field of research, or clinical use of Magnetic Resonance Images of the brain.

The term "software plug-in" or "plug-in module" refers to additional computer software that can be used together with some other computer software with the purpose of adding functionality, features to, and outcome of the software.

One approach to motion artifact correction is to modify the pulse sequence during the course of the acquisition itself, in real time. Such prospective motion correction methods attempt to keep the image slice locations fixed with respect to the patient throughout the scanning process. Prospective correction methods have commonly used MR navigators with a variety of k-space trajectory shapes, including linear, circular, spherical, and "cloverlea '-shaped navigators, with varying degrees of speed and accuracy for tracking sophisticated types of motion in one to three dimensions.

A primary object of the invention is to provide new and better methods for motion correction during brain MRI using new navigators. The new motion correction methods of the invention involve different new navigators based on fat signal from the body, as is described in the preferred embodiments below. The motion correction of the invention is executed by real-time adjustments of the imaging pulse sequence (i.e. MR acquisition technique). The methods are e.g. capable of quantifying rigid-body motion using real-time motion correction combined with new navigator tracking mechanisms. The methods are thus performing "prospective motion correction". This invention disclosure involves novel types of MR navigator data to drive the real-time correction process, also known as "motion navigation/correction method". These novel MR navigation data are acquired using a short extra pulse sequence module that can be combined with a variety of Magnetic Resonance imaging acquisition methods (pulse sequences). The invention thus solves the problem of creating sharp MR-images even though the scanned subject is moving with as little extra acquisition time as possible, stemming from the extra tracking or navigator MR data acquired in parallel with the main imaging pulse sequence. A low- resolution image is generated from the novel MR navigator raw data to track real-time movements of e.g. the patient's head, updating the position of the slice image plane to follow the movement of the head.

Figure 1 is a schematic overview for the proposed fat navigation method. Figure la shows cross-sections of the MR scanner, with the x-y-z magnetic gradient ramps that are used to spatially encode the MR signal. In Figure lb (left-most panel), operator-prescribed slice planes are shown, with a fixed mapping to the physical coordinate system. In the middle is shown that if the head is moved, the slice planes will correspond to other parts of the anatomy. On the right is shown that with prospective motion correction, such as the proposed 2D or 3D fat navigation technique, the slice locations are continuously updated to follow the patient's head, based on motion information from the fat navigation data. In Figure lc is illustrated an acquisition process utilizing some main pulse sequence that is modified to also include a short fat navigation sequence module, run in parallel to the main sequence acquisition. The collected data from the fat navigation sequence module is sent to a motion estimation computer, which calculates the amount of motion that has occurred and continuously sends this information back to the MR acquisition computer, which in turn updates the logical coordinates so that it follows the motion of the patient.

Like PROMO [2], the data used for tracking in the present invention is in the image domain, not in k-space. This allows one to spatially select the region that is believed to move rigidly and along with the brain, such as the fat in the scalp of the head, but excluding the non-rigidly moving fat signal in the neck.

A difference between the method of the present invention and PROMO is that it leaves the brain water magnetization unaffected, since only the fat magnetization is being used. The effect of this is that the slowly acquired high-resolution diagnostic image will look the same with or without the motion correction according to the present invention using fat navigators, removing the risk of darker signal bands in the diagnostic image. Another difference is that PROMO uses non-Cartesian spiral readouts, without a unidirectional frequency-encoding direction, while the preferred data readout for the present invention uses a 2D or 3D Cartesian Echo-Planar-Imaging (EPI) readout with a unidirectional frequency-encoding direction. With the frequency encoding direction placed along the Head-Feet direction, signal wraps (a.k.a. 'aliasing') from the neck and chest outside the imaged area can be avoided in the proposed implementation. Yet another difference is that the images of the present invention are sparse due to localized presence of fat in the head (mostly around the skull). This lends itself to higher compressibility, which allows for higher parallel imaging factors and therefore shorter acquisition times. This means that less percentage of the scan time is needed for these navigation images, making the overall scan time shorter in many cases. A shorter scan time will allow for more image series per examination of a fixed duration, or the total examination time may be reduced to allow for more patients examined per day at a clinic. Both due to the fat-excitation and the shorter duration time, the module of the present invention can therefore be more easily combined with a larger set of MR pulse sequences (i.e. acquisition methods) than PROMO, and without modifying the original water signal in the diagnostic MRI data. This means that one can make more types of MR image acquisition protocol compatible with the method of the present invention than with PROMO, and thus allow for prospective motion correction for more types of MR acquisition methods used in a typical examination, which in turn will lead to an overall reduction in motion corrupted images in a complete MR examination. Another benefit over PROMO, is that the longitudinal ("Tl- relaxation") magnetization of fat recovers -5-10 times faster compared to the water magnetization in the brain parenchyma or the cerebrospinal fluid (CSF). Hence, one can re-use this magnetization more often compared to a water-magnetization based navigator.

The fat navigation methods proposed in the present invention are image based, similar to PROMO, and can therefore like PROMO exclude signal from the back of the neck or the jaws, for example by automatic ROI (Region-Of-Interest) selections. Out of the several embodiments of the present invention, 2D sagittal, i.e. a two-dimensional plane in the center of the head resulting in a side view of the head, fat navigation was the first one implemented. Unlike PROMO, this variant only corrects the data (in real-time) in 2D (within the sagittal plane), which corresponds to the most problematic nodding motion. The duration was ~15 ms. The use of 2D sagittal fat navigation was a reasonable first step for two reasons: First, the nodding motion is the most prevalent and corresponds to the highest amplitude of motion, mainly due to the fact that it is very hard to use soft padding around the patient's head to prevent this type of motion and because breathing typically creates this motion pattern. Second, as the majority of the MR images are acquired in the transverse (axial) scan plane, one can use already existing techniques such as PROPELLER imaging to correct for motion in this axial plane, using the high-resolution images themselves, and letting the fat navigation correct for motion in the nodding direction with a very small scan time overhead.

Figure 2 is a graph showing fat navigation according to the present invention and a main pulse sequence. This figure is a pulse sequence diagram consisting of i) a leading 2D fat navigation module with RF excitation followed by an accelerated EPI readout in the sagittal plane, and ii) a commonly used main pulse sequence in MRI. Note how the slice, phase and frequency encoding axes differ for the fat navigation module and the main (diagnostic) pulse sequence, making the two planes intersect. Each time the 2D fat navigation and FSE sequence is played out, a low-resolution sagittal snapshot image is obtained (in < -20 ms), while only a few lines worth of data are obtained for the axial (or coronal) diagnostic image, requiring minutes of scan time to complete. Figure 2b illustrates the fat navigation and high-resolution MR images as acquired according to the figure. For the sagittal fat navigation snapshot image, notice the lack of signal inside the brain and that the fat scalp can be well visualized despite the low resolution.

The 3D variants of fat navigation proposed herein are nevertheless more general, and ultimately preferred. Our own experience of how much the spatially sparse fat navigation data can be accelerated using conventional Cartesian parallel imaging, enables e.g. 3D EPI acquisitions using the fat signal content in the head within as low as a -20 ms time window. This would have been prohibitive using a similar 3D EPI and water signal (which is not spatially sparse). In the current implementation, acceleration factors of up to ~32x have been possible without loss of motion estimation accuracy. As mentioned above, adding fat navigation to an existing imaging method (the main "pulse sequence") will not affect the indented (water) signal from the sequence. A combined pulse sequence is created by interleaving the fat navigation module with the main pulse sequence, the latter of which can be of a different length. This makes the combined sequence longer in duration, which after executing the sequence many times for the purpose of collecting enough data with the main pulse sequence also increases the total scan time (like the case for most navigators). However there are also certain diagnostic pulse sequences (MR acquisition methods), such as those involving e.g. an inversion pulse with a time delay of anything between 100-3000 ms, in which the fat navigation sequence could be played out multiple times within this time period without any penalty in terms of scan time. Each time the combined pulse sequence is played out (normally within -300 ms, of which -15-25 ms is the added fat navigation module), one or a few lines worth of the high-resolution diagnostic image is acquired alongside with one complete, but low-resolution, fat navigation image (2D) or image volume (3D). Hence, one may obtain many fat navigation images per second, and the short Tl -relaxation from fat allows for strong MR signal from fat despite the fast repetition time of consecutive RF excitations. In the current implementation, the fat navigation data is read out using an echo planar imaging (EPI) readout. As EPI is a snapshot type of acquisition, it is sensitive to geometric image distortions originating from the magnetic field inhomogeneities near anatomic areas that are adjacent to air cavities (such as the ear canals and the sinuses). A concern is therefore that the fat navigation image is deformed and therefore causes a bias in the motion estimates when aligning them to each other over time. However, as the fat signal is sparse in the image, it lends itself to high acceleration factors. Sparse data has for some time been subject to world-wide MR research using 'compressed sensing' (CS) techniques, where undersampled, and randomly sampled, sparse data can be reconstructed in a CS framework. The drawback with CS is its very long reconstruction times, which makes it not useable in a real-time feedback scenario such as prospective motion correction proposed herein. Nevertheless, we have found that it is possible to accelerate the fat navigation acquisition using plain Cartesian parallel imaging (e.g. GRAPPA [7]), far beyond what is possible for water signal. For 2D fat navigation, good image results are obtained with acceleration factor of R=8 using an RF coil with only 8 receiver channels.

Another advantage of the present invention over other image domain based navigators relying on water signal is 'g-factor' noise. G-factor noise is structured in nature and stems from the parallel imaging reconstruction. The g-factor noise increases with higher acceleration factors, and is predominantly located in the center of the image. Using water-based signal, the navigator data comes from the brain tissue, where there may be significant g-factor noise artifacts. These artifacts do not move rigidly with the head and may bias the motion estimation. Using fat-signal, such as in the proposed methods here, the use of high parallel imaging factors lead to less g- factor noise, and the g-factor artifacts stemming from the fat-signal end up in the brain region, not overlapping with the fatty scalp used for motion navigation.

Figure 3 shows the benefit of using the fat-signal instead of water signal for fast, i.e. highly accelerated, data acquisitions. An acceleration factor R used together with an EPI data readout means that only every R ^th line needs to be acquired, and where the remaining (R-1)/Rxl00% of the raw data is synthesized in the imaging reconstruction using parallel imaging [7]. The fat navigation embodiments proposed herein have been implemented with EPI readouts to facilitate simple image reconstruction, in part involving the parallel imaging reconstruction step, in order to be able to feed back data quickly to the acquisition process. With everything else constant, R=8 GRAPPA reconstruction using fat signal provides more accurate reconstruction compared to when water signal is used (white arrow). Figure 4 shows motion estimation accuracy tests for four different 2D fat navigation scans with a slowly acquired high-resolution gradient echo (SPGR) image as reference. Motion estimates are shown for 30 distinct head poses, where the five scans of Figure 4a were acquired in each pose. Figure 4b shows that higher acceleration factors such as R=8 (instead of R=2) provides more faithful motion estimates. The difference between the "FatNav+" and "FatNav-" labeled scans, is the direction of the geometric distortions, while the R-factor determines the amount of geometric distortions. Higher R-factors gives less geometric distortions, and hence less bias in the final motion estimates.

We have found that a single-shot 3D-EPI Cartesian trajectory is not only an option, but can with good image quality be used with acceleration factors around R=4-6 in both phase encoding directions, leading to a combined acceleration factor of R = 16-36. This is shown in Figure 5. This is a major breakthrough, and allows e.g. a 3D-EPI volume of 32x32x32 pixels to be acquired in ~20 ms, only slightly longer than the first 2D FatNav implementation using R=4 in one direction. Moreover, to further reduce the EPI readout time, partial Fourier acquisition techniques could be used using reconstruction methods such as e.g. homodyne or POCS. This does not further reduce geometric distortions, but does shorten the total EPI readout time.

While the main focus of the present invention is for brain MRI application, fat also exists in other body parts and the same implementations may be used also outside the brain region. A first embodiment is used for motion patterns mainly occurring within a plane. In this embodiment, a single 2D fat navigation plane is applied to capture such motion.

A second embodiment addresses rigid body motion, but one may also extend the model for the motion estimation based on the fat navigation data to involve elastic motions as well. While this is not a likely scenario for the rigidly moving brain, it may be of use for other regions of the body.

In a third embodiment, 2D fat navigation may be extended to a 3D fat navigation technique, at the expense of longer readout duration. Fat excitation is thus performed in three orthogonal 2D planes. Thereby, one is able to perform 3D corrections for arbitrary head motion, but without affecting the water magnetization of the brain. However, this requires three spectrally selective RF pulses, each of which typically requires at least 8 ms time, which is why such fat navigation module duration may need to be about 40-50 ms long. Alternatively, e.g. either a single non-slice selective, but spectrally (i.e. fat) selective, RF pulse, or, a Spectral-Spatial RF pulse, which excites the entire volume of interest to acquire 3D volume data can be used. Here, the 3D readout time is kept short by using parallel imaging techniques (or similar) in preferably two directions. Cartesian 3D-EPI is the current implementation using high parallel imaging factors in both phase encoding directions, leading to a full 3D volume with smaller pixel (voxel) sizes than PROMO, with less navigator time. Moreover, this navigator time will be shorter compared to when performing excitation in three orthogonal planes, due to the use of only one RF pulse.

Figure 5 shows the potential of highly accelerated Cartesian 3D-EPI acquisitions using fat signal - 3D-fat navigation according to the present invention. Despite a 32-fold reduction of the fat navigation raw data (seen in the difference between Figure 5a and Figure 5b), fast and simple Cartesian 3D parallel imaging techniques [7,8] can reconstruct the 3D fat navigation data sufficiently well. This massive reduction of required raw data is possible for 3D fat navigation since: i) the geometric distortions are small, despite a 3D readout covering a -32x32x32 k-space, ii) the strong acceleration makes the 3D-EPI readout portion only ~20 ms, and iii) the fat-signal is inherently sparse. Without any parallel imaging acceleration, the readout time would have been -640 ms. The differences between Fig. 5a and 5b are relatively small, but more importantly, the very high correlation in the estimated 3D rigid body motion parameters, (Fig. 5c), derived from the fully sampled (c.f 'x' markers) and 32x (R=32) accelerated (c.f. triangle markers) fat navigation data, respectively, shows the potential of the this embodiment of fat navigation. Fig. 5c shows the six panels corresponding to the six motion parameters estimated for each time point (i.e. each FatNav volume, x-axis) for a volunteer performing very large head rotations of the order of ~30 degrees. Note that multi-shot acquisition has been used in this example to collect data for the fully sampled dataset in Figure 5a. With a 32x32x32 volume matrix size and a combined acceleration factor of R=32 (Ry=8, Rz=4), Fig. 5d) shows the 3D fat navigation sequence module.

The data acquisition part of the fat navigation embodiments described above is an EPI k- space trajectory. In a fourth embodiment, the readout is a non-Cartesian readout, such as 2D/3D spiral, 3D cones, 3D twisted-projections, or 2D/3D radial trajectory, the choice of which will imply different sensitivity to geometric distortions and data acquisition and reconstruction efficiency.

For 2D fat navigation, the fat-selective RF pulse is a Spectral-Spatial (SPSP) RF pulse. In a fifth embodiment, other RF pulses spectrally (fat) selective, but not necessarily spatially (slice) selective. The latter entails that all fat in the entire head is excited and e.g. a sagittal fat navigation image read out following this excitation would correspond to a collapsed image in e.g. the left-right direction, and correspondingly for other projection views. The advantages of this is a shorter RF pulse, and higher fat-signal over larger regions including the fatty signal around the eyes, which might help the image registration process. Data can be read out in all three orthogonal planes (i.e. collapsed data viewed from three orthogonal directions) following a single RF pulse with duration of around 5-12 ms. This is followed by three highly accelerated EPI readouts, which sum up to ~9 ms duration. Hence the total duration for this '3-plane Collapsed FatNav' would be -15-20 ms including RF pulse, depending on the B ₀ field strength and acceleration. The best use of this variant is when the main pulse sequence (used for diagnostic imaging) already employs a leading non-slice-selective 'fat-sat' RF pulse for diagnostic purposes. In this case, one makes use of ('borrows') that fat-signal for motion navigation before it is intentionally destroyed by the main pulse sequence using its spoiler gradient (prior to the main RF excitation). Hence, one of the best implementations of this embodiment, is to only add the three orthogonal EPI-readouts after the already existing fat-sat- RF pulse and thereby only add as little as ~9 ms of extra time. Fat-sat RF pulses are today always used for the following clinical MR acquisition techniques/applications: Diffusion MRI, Perfusion MRI, and Functional MRI. Moreover a fat-saturating RF pulse is sometimes used by the main pulse sequence for e.g. Tl -weighted MR contrast; for example when looking for pathologies around the eyes where the fat signal obscures pathologies.

Figure 6 shows another efficient implementation of fat navigation for brain applications with ability to correct for motion in 3D, where collapsed views of the object are generated along three orthogonal directions. As an example, Fig. 6a shows a Spin-Echo EPI pulse sequence main pulse sequence (dashed), commonly used for e.g. diffusion MRI applications. For image quality reasons, EPI sequences need to avoid signal from fat, hence the leading fat-sat pulse (black arrow, Fig. 6a) followed by a spoiler gradient (here on the y-axis) that destroys the fat signal prior to the normal RF excitation. By inserting a block of three orthogonal accelerated fat navigation EPI readouts between the main sequence's fat-sat pulse and named spoiler gradient, this fat-signal can be used for 3D motion detection. As the fat-sat RF pulse is usually not spatially selective, the three orthogonal fat navigation readouts effectively become projections of the data from three different views, hence the name 'collapsed' FatNav. These are less visually appealing, but also these can evidently be accelerated by R=8 on an 8-ch RF coil (Fig. 6b), without introducing parallel imaging related artifacts. The collapsed FatNav images' ability to capture motion is shown in Fig. 6c, where the top row shows the case for no motion.

It should be noted that it is possible also for the third embodiment, i.e. full 3D fat navigation, to use the fat-signal from a fat-sat RF pulse in the main sequence if present. This would shorten the 3D fat navigation duration by 6-10 ms, similar to the fifth embodiment. As the time between the main sequence's fat-sat RF pulse and the following RF excitation pulse increases when inserting a 3D FatNav or a 3-plane Collapsed FatNav module, the nominal flip angle of 90 degrees for the fat-sat pulse should be increased slightly to compensate for Tl -relaxation such that its longitudinal magnetization is zero at the time of the RF excitation of the main sequence.

An object of the invention is a method for reducing motion related imaging artifacts in MRI acquired images, said method comprising a motion correction method utilizing the new image domain based 2D, 3D or collapsed 3D FatNav navigators.

Another object of the invention is a method for generating a magnetic resonance image by acquiring MR data and applying a prospective motion correction method to the MR data, in which prospective motion correction method the new navigator, image domain based FatNav, is being used.

Another object of the invention is to provide new navigators for MRI, such as brain MRI. These navigators could either be 2D, 3D or collapsed 3D fat-selective navigators.

A further object of the invention is the use of new navigators, for motion correction during MRI, preferably brain MRI. These navigators are fat-selective navigators. An object of the invention is a motion correction tool to be used as a plug-in module for existing imaging techniques. The tool of the invention could for example be designed as an optional software plugin compatible with multiple MR pulse sequences. The fat navigation plugins are effectively adding new waveforms and acquisition segments before, after or inside the main pulse sequence used to record the MR data for diagnostic use. The raw data recorded by the main pulse sequence are sent to the image reconstruction process on the intended reconstruction computer with the intent to produce diagnostic images. In contrast, the fat navigation raw data will be sent to a process on some computer, which reconstructs the raw data to low-resolution images typically many times per second with the purpose of obtaining motion estimates in the form of millimeters and degrees by comparing the fat navigation image data over time. The motion information data is semi-continuously sent back to the pulse sequence that is running on the real-time computer of the MR-system. The pulse sequence can therefore continuously update the prescribed slice location (or the logical coordinate system) using this motion information. Thereby, the prescribed slices dynamically follow the head as it is moving.

The software may be in various forms such as system software or application software. Further, the software may be in the form of a collection of separate programs, a program module within a larger program or a portion of a program module. The software may also include modular programming in the form of object-oriented programming. The processing of input data by the processing machine may be in response to user commands, or in response to results of previous processing, or in response to a request made by another processing machine. How to program the motion correction software is known for a programmer skilled in the art, and the software or software plug-in of the invention is simple for the person skilled in the art given the information disclosed herein, i.e. the use of fat navigators, where the input data is the raw data of the signal from fat in the head (or other body parts). Given this information, the person skilled in the art can use a fat navigator over some other alternative navigator together with the main image acquisition software, and essentially use the same motion processing and motion feedback pipeline. The processing software could be written in any computer language, and the acquisition software, as well as the particular implementation of the fat navigator plugin, is typically made in the same programming environment that the vendor is using for its MR-systems. I.e. the various fat navigation plugins must be written in different software environments depending on the vendor. Hence, a reimplementation would be necessary to transfer the implementation across vendors for the acquisition part of the software. The motion estimation part of the software could be partially reused across vendors. A further object of the invention is an apparatus to correct for head motion in MRI using a fat-selective excitation and readout, for 2D or 3D navigation and real-time correction. The apparatus may be a computer connected to an MRI-scanner. While acquiring an MRI-image of a slice of the brain, or a 3D volume of the brain using fat-signal, a motion correction software is used on the navigator data receiving computer, in which the fat navigation data is used to estimate the amount of motion for the purpose of prospectively compensate for head motion of the patient. The motion information data is semi-continuously sent back to the pulse sequence that is running on the real-time acquisition computer of the MR-system. The pulse sequence can therefore continuously update the prescribed slice location (or the logical coordinate system) using this motion information. Thereby, the prescribed slices dynamically follow the head as it is moving.

An additional object of the invention is an MRI system, comprising a resonance assembly, an RF transceiver system and a controller programmed to acquire MR data and applying a prospective motion correction method utilizing a new fat navigator and generating an image. The resonance assembly comprises gradient coil(s) and RF coil(s). The RF transceiver system is coupled to the RF coil(s) and configured to receive magnetic resonance data from the RF coil(s). The controller is programmed to acquire two types of MR data from the RF coil(s), one of which is used for the purpose of motion detection and to apply a prospective motion correction method (the motion correction method utilizing the new fat navigator) to the other type of MR data and to generate an image based on the motion corrected MR data. When a subject (a patient, healthy volunteer or an animal) is to undergo an MRI scan/examination, the subject will be placed e.g. on a bed that is moved into the MRI scanner. An MRI scanner consists of a short tunnel that is open at both ends. For the acquisition and image reconstruction equipment, any existing commercially available MR system may be used. A computer is used to operate the MRI scanner, which is often located in a different room to keep it away from the magnetic field generated by the scanner. Thus the radiographer who operates the computer will be in a separate room to the subject. For the motion processing computer handling the fat navigation raw data, this can be any computer running any operating system, as long it has a networked or other type of connection with the acquisition software to which it will send the motion information. The MRI computer will run one or more pulse sequence programs to control the MR scanner and to acquire one or more images of the subject, for example the brain. Involuntary motion of the subject during the data acquisition will render the MRI images blurry and less useful for diagnostics. By using a prospective motion correction software program on the computer used to operate the MRI scanner, the location of the chosen slices may be updated in real-time to the computer. Thus the pulse sequence receives the motion information in real-time during the acquisition process itself, to follow the movements of the subject. Thus, the gradual build-up of raw data will be performed in a scan plane(s) that, more or less, is fi ed in relation to the head rather than fixed with respect to the laboratory system. The motion correction software of the invention uses special fat navigators to obtain low-resolution images that give the computer the data it needs to provide feedback to the acquisition software to move along with the motion of the subject. A typical image slice thickness in a diagnostic high- reso hit ion image is 2-5 mm, with a typical in-plane resolution of (0.5- 1 )x(0.5- l ) mm ² . A 2D fat navigation slice (for those variants) is generally thicker for the purpose of receiving more signal for motion estimation (-20-100 mm), with an approximate in-plane resolution of ~(3-6)x(3-6) mm ² . The 3D fat navigation variant excites not just a slice but essentially the whole head, with a resolution of the volume pixels ('voxels') of about (5-8)x(5-8)x(5-8) mm ³ .

In the following a fat navigation embodiment is described. Patient head motion is one of the leading sources of artifacts in brain MRI. In particular, it is difficult to restrain motion in the 'nodding direction' - a direction where motion naturally occurs, in part due to the patient's breathing. Pulse sequences acquiring the raw k-space data in a propeller fashion have been proven clinically robust to motion. However, as typically only in-plane retrospective correction is performed using propeller-based acquisition techniques, nodding motion is left uncorrected. Moreover, when retrospectively correcting any 2D imaging scan, spin-history effects (i.e. undesired saturation of the magnetization in e.g. adjacent slices) cannot be addressed. Over the recent years, both hardware (video camera) and MR data (navigator) based prospective motion correction techniques have emerged. For example, the MR navigator based PROMO technique uses three orthogonal excitations and spiral readouts to obtain three low-resolution images in parallel to the actual image acquisition. The flip angle of the three RF excitation pulses in PROMO has been reported to be 8° to minimally affect the longitudinal magnetization used for imaging. Nevertheless, the continuous use of about fifteen 8° RF pulses per second may saturate the brain signal by up to about ten percent, in particular in the -lxlxl cm ³ intersection point of the three orthogonal navigator planes. In quantitative or contrast sensitive applications, the PROMO technique may therefore not always be ideal. In the current invention, we propose a fat- only (FatNav) navigator image for prospective correction of head motion.

Here a 2D/3D fat-image navigator is described. The 2D/3D FatNav raw data is recorded using e.g. an echo planar imaging (EPI) readout. The navigation duration is shorter than PROMO. To reduce geometrical distortions and navigator duration, a high parallel imaging acceleration is used. For FatNav, the navigation time per "snapshot image" of the invention may vary between 9 and about 30 ms, depending on pulse sequence parameters chosen such as image resolution, bandwidth, FOV, whether partial Fourier techniques are used, pre-existing fat-sat RF, and 2D vs. 3D FatNav. The upper useful bound for the FatNav module is constrained by the signal (T2*) decay of the transversal fat magnetization and depends also on the type of main sequence for efficiency reasons. Other MR image acquisition parameters may vary. For example, the FatNav slice thickness may range from a few mm (~2-3 mm) to the whole head. The slice thickness will therefore include more or less fat-based anatomy that in the FatNav image becomes "collapsed". Including more anatomy will increase the total signal in the FatNav image and will make more pixels to have significant fat-based signal. A parameter such as the Field-Of-View (FOV) may be increased to include larger head sizes and nearby anatomy and also to allow for larger motion levels and geometric distortions at the expense of larger pixel sizes (in mm units) that will reduce the accuracy of the found motion parameters.

The image matrix size may be reduced to make the data readout time (and hence the FatNav module) shorter, again at the expense of larger pixel sizes but with the benefit of lower geometric distortions. A matrix size of 16x16 is considered low and will likely make the motion estimates less accurate than required. A matrix size of e.g. 128x128 is considered high for navigation purposes and may likely (but not definitely) be too distorted and data handling and processing times may potentially increase too much to achieve the desired motion information update rate. Higher matrix sizes than -32-48 in each direction will be difficult for 3D FatNav, due to the resulting long duration.

The echo time TE should be as short as possible to make the FatNav module short and to maximize the signal, but partial-Fourier techniques may be used to reduce the EPI readout time to almost half, which also reduces the FatNav module duration. TE will increase as the matrix size is increased, but will in most cases be -5-15 ms.

The parallel imaging reduction factor R should be as high as possible to shorten both the FatNav duration and the geometric distortions. The highest usable R will be dependent on the RF coil used. An R-value equal to the number of coil elements is feasible for FatNav. The R-value does not necessary need to be an integer, (the acceleration factor could be for example 3.1 or any other decimal value).

The Flip Angle (FA) of the RF pulse in the FatNav module should be as small as possible within the limits of receiving enough signal, except for when the fat-sat pulse from a main pulse sequence is used (in which case it has to be near 90 degrees to perform the intended saturation). A small flip angle produces less RF heating and perturbs the fat-magnetization less, hence results in less overall impact of the fat navigation module on the patient and the overall magnetization and signal for the main pulse sequence.

Other data readout types than EPI may be used. For example, Spiral readouts could be used in a similar manner as in PROMO.

EXAMPLES

Example 1 : 2D sagittal fat navigation to correct for nodding type of head motion Methods

Data was acquired on a healthy volunteer on a 3T GE MR system using an 8-channel RF coil. A 2D fat navigator (FatNav) module, comprising of a Spectral-Spatial (SPSP) fat-selective excitation pulse followed by a sagittal Gradient-Echo Echo Planar Imaging (GE-EPI) readout was implemented as a standalone pulse sequence (Fig. 2a, left) with the future goal to add this module to e.g. an axial PROPELLER sequence (Fig. 2a, right). An SPSP RF excitation pulse had its center frequency shifted to the fat resonance peak. The flip angle of the SPSP RF pulse was set as low as 10°. The data is collected using an EPI readout. High GRAPPA acceleration factors to reduce the geometric distortions and consequently the amount of susceptibility related errors in the motion estimation. GRAPPA acceleration factors ranging from R = 1 to 8 were investigated. As R increases, the geometric distortions (in mm units) are reduced by a factor of R, and the total duration of the fat navigation module becomes shorter.

We hypothesize that the GRAPPA acceleration factor could be increased higher than in a normal, water-excited head acquisition for four reasons: a) the fat navigation data is sparse, which may make it easier to unalias compared to a regular MR image, b) g-factor noise in parallel imaging is building up in the center of the coil, which is a region that contains little fat signal and is therefore not affecting an image registration based on the fat signal, c) that the thick sagittal slice used for both GRAPPA calibration and for the scans makes the image noise very low, and d) the fat navigation data (in the skull) is located quite close to the coil elements. A stepwise motion experiment was conducted, consisting of 30 scan sets with a change of head pose between each scan set. The first scan was a distortion-free spoiled-gradient echo (SPGR) scan, acquired with a matrix size of 256x256 and used as gold standard. A fat-sat pulse was added for the SPGR scan to avoid fat signal in this reference image. The following two scans in each set were fat navigation scans with R=2, the first with positive phase encoding blips ('bottom-up' k-space sampling) and the second with negative phase encoding blips ('top-down' k-space sampling). Changing the polarity of the phase encoding blips negates the direction of geometric distortions, and it was hypothesized that potential bias in the motion estimates due to geometric distortions would manifest in opposite directions for the two scans. 2D retrospective image realignment across the 30 scan sets was performed using a simple sum-of-squares metric to determine two translation parameters and one rotation parameter for each time point.

Results

In Figure 3, image reconstructions using GRAPPA acceleration factors of R=4 and R=8 are shown using the 8-channel RF coil. The thick slices used (30 mm) implies very high SNR, why both the water and fat data is well reconstructed for R=4. Doubling the acceleration to R=8, the water image becomes corrupted with GRAPPA reconstruction artifacts (g-factor noise).

In Figure 4, the estimated motion parameters are shown for the step-wise motion correction experiment with 30 head poses conducted by a volunteer. The five images acquired in each set (for each head pose) are shown in Figure 4a. In Figure 4b, motion estimates are shown. The solid thicker lines correspond to the high-resolution distortion-free SPGR data used as gold standard. The blue and red lines differ in the phase encoding blip direction, where an opposing bias around the solid black reference line should be a sign of interactions between motion and susceptibility distortions. The dashed lines correspond to the R=2 scans, with its underlying FatNav images being four times as sensitive to off-resonances as the R = 8 scans shown as solid red and blue lines. While the R=8 curves are not entirely overlapping, the reduction in discrepancy compared to the R=2 scans is still noticeable, particularly for the translation estimates. For rotations, it is more difficult to distinguish between the four FatNav scans relative to the SPGR time course, suggesting that the EPI distortions do not as much translate into a bias of the rotation estimates. Discussion

A 2D fat navigator technique (FatNav) using only the fat signal in a 30 mm mid-sagittal slice of the patient's head has been presented for future prospective correction nodding motion. We have shown that the motion estimates using the FatNav images acquired with high R factors are quite close to those from the gold standard SPGR reference data. The mean-absolute-error is about 0.5 mm for the translation parameters and about one degree for the rotation parameter. The maximum useable R for our 8-channel RF coil was found to be the highest attempted, R=8. This value is high compared to what is normally used in regular GRAPPA accelerated 2D pulse sequences. This was possible largely thanks to a careful choice of the GRAPPA calibration data and the fact that the normal g-factor noise appears in the center of the coils where there is no fat- signal of interest to drive the image registration. Another key to successful reconstruction at this high acceleration factor is the high SNR in both the calibration and accelerated scans due to the low resolution and most importantly the 30 mm thick slice used. The philosophy behind 2D FatNav was to approximately address out-of-plane motion for axial or coronal imaging, especially for sequences such as PROPELLER where 2D (in-plane) motion correction is performed retrospectively. For classical Cartesian pulse sequences, lacking in-plane corrections, the 2D FatNav module could be extended to cover three orthogonal planes like PROMO, and thereby be able to correct for 3D motion at the expense of a navigator module duration of the same length as in PROMO. Example 2: fat navigation using 3D-EPI readout for full 3D volume of the head, and two directional parallel imaging acceleration with combined acceleration of ~30X

Methods

Data was acquired on a 3T GE MR scanner using a 32-ch Nova head coil (Nova Medical, MA, USA), using a 3D fat navigation module comprising a GRE 3D-EPI readout and a 10° non- slice- selective fat-selective RF excitation using a FOV of 280x280x280 mm ³ and an isotropic voxel size of 5.8x5.8x5.8 mm ³ . A total of 15 acquisitions were performed on a healthy volunteer, who was instructed to assume different head poses between scans. Each 3D k-space data was decimated in the ky-kz plane as to simulate GRAPPA acquisitions at various acceleration factors and then reconstructed using weights estimated from the first scan. Specifically, the considered configurations had overall acceleration factors R of 16, 24 and 32 factored as 4x4, 4x6 and 4x8 along the ky and kz directions. Also, note that the corresponding navigator modules have a duration of approximately 58 ms, 44 ms and 35 ms respectively, for 48x48x48 reconstructed data points. For a lower matrix size of 32x32x32 and R=32, the 3D fat navigation duration became ~20 ms. The fully sampled and the accelerated volumes were registered separately against its own first volume using a simple sum-of-squares metric, resulting in six estimated rigid-body motion parameters per volume.

Results In Figure 5a, coronal, axial and sagittal cross-sections of the first volume are shown, using multi-shot EPI instead of parallel imaging acceleration to yield the expected low geometric distortions. In Figure 5b, l/32 ^th of the raw data acquired in Figure 5a have been used, and the 3D image data is recovered via 3D GRAPPA reconstruction. The image quality is largely maintained. More importantly, the motion estimates from Fig. 5a) and 5b) are nearly identical, as shown in Figure 5c.

Discussion

From the results presented, it is apparent that the proposed acquisition scheme shows promise for prospective correction using 3D FatNav. With this extreme acceleration, using a simple 3D GRAPPA kernel for reconstruction, there is no need for slow compressed sensing solutions and random sampling of the raw data such as shown by Gallichan et al. [6]). By reducing the matrix size to 32x32x32, full 3D data can be achieved in about half the duration as three orthogonal planes in the PROMO navigator technique, with maintained motion estimates.

Example 3; Collapsed FatNav - 3 orthogonal projections using existing RF pulse from main pulse sequence.

Methods

A healthy volunteer was scanned on a 3T GE MR system using an 8-channel head coil. A test sequence was made, with the standard 'fat-sat' RF-pulse followed by the three orthogonal EPI readouts using the following parameters: FOV = 28 cm, matrix = 48x48, resolution 5.8x5.8 mm ² , and TR = 40 ms. GRAPPA acceleration factors of 4, 6 and 8 were attempted. For R=8, the echo-times for the three orthogonal projections became 4.1, 6.7, and 8.7 ms. Separate GRAPPA calibration scans were performed using the same test sequence, but with the acceleration factor replaced with the same number of EPI shots. This allowed the GRAPPA calibration data to match the scan data both with respect to data content (fat signal) and geometric distortions (k- space phase accruals). For R=8, four experiments were carried out, were the volunteer was first instructed to lie as still as possible, followed by pitch, yaw, and lastly roll motion. 400 repetitions were performed for each of the four cases. The images were motion corrected relative to the first image in the series using a sum-of-squares metric.

Results

Figure 6b shows the navigator images for three acceleration levels, R = 4 and 8. The projection (collapsed) views makes the images less appealing compared to a full 3D acquisition, but what is important is the ability to correct for motion and avoidance of phase-induced signal cancellations, and image ghosting due to parallel imaging reconstruction. No apparent parallel imaging artifacts (ghosting) can be seen even at R=8, despite the use of only 8 receiver channels. Figure 6c shows motion estimates from the four different experiments, the first of which is without motion (top), followed by rotational motion along each axis (bottom). Discussion

When a (non-slice-selective) fat-saturation module is added to a (diagnostic) main sequence, the fat signal is spoiled to null prior to the RF excitation in the main sequence. Here we propose to use this fat signal for prospective motion correction before it is spoiled, by adding three orthogonal 2D-EPI readouts in between. As the fat-sat pulse is non-slice selective, reading out three orthogonal 2D k- spaces results in projection images. The addition of the three EPI readouts, using R=8, adds only ~9 ms of total sequence time, but will not affect the TE for the main sequence. EPI based pulse sequences, often used for diagnostic diffusion/perfusion/functional MRI, need fat-saturation. Common for these applications is that the main sequence duration is -50-120 ms long, making the additional Collapsed FatNav readout duration small in comparison. Through-plane, 'collapsed'/'projection', navigators has also recently been proposed (though using water signal) [9], but required the entire stack of 2D slices of the main sequence to form a single collapsed navigator. Despite recent experience of attaining good image quality for sparse 2D FatNav images at very high reduction factors, it was not obvious that it would translate to this collapsed FatNav data. Yet, with a successful acceleration factor of eight on an 8-channel coil, we will now explore yet higher acceleration factors using 32-channel head coils. Most importantly, we will next implement this as a sequence plugin module with a real-time feedback loop for real prospective motion correction of diagnostic pulse sequences. Bibliography

1. Pipe JG (1999) Motion correction with PROPELLER MRI: application to head motion and free-breathing cardiac imaging. Magn Reson Med 42: 963-969.

2. White N, Roddey C, Shankaranarayanan A, Han E, Rettmann D, et al. (2010) PROMO:

Real-time prospective motion correction in MRI using image-based tracking. Magn Reson Med 63: 91-105.

3. Kohler MO, Denis de SenneviUe B, Quesson B, Moonen CT, Ries M (2011) Spectrally selective pencil-beam navigator for motion compensation of MR-guided high-intensity focused ultrasound therapy of abdominal organs. Magn Reson Med 66: 102-111.

4. Kawaji K, Spincemaille P, Nguyen TD, Thimmappa N, Cooper MA, et al. (2013) Direct coronary motion extraction from a 2D fat image navigator for prospectively gated coronary MR angiography. Magn Reson Med.

5. Ari N, Kraft RA. FAT Navigators for Functional MRI: Is There Sufficient Amount of Fat Signal in The Human Head for Accurate Motion Detection with Fat Navigator Echoes?; 2005; Miami, FL, USA. pp. 1546.

6. Gallichan D, Marques JP, Gruetter R. FatNavs: Exploiting the Natural Sparsity of Head Fat Images for High-Resolution Motion-Navigation at Very High Acceleration Factors; 2013; Salt Lake City, UT, USA. pp. 309.

7. Griswold MA, Jakob PM, Heidemann RM, Nittka M, Jellus V, et al. (2002) Generalized autocalibrating partially parallel acquisitions (GRAPPA). Magn Reson Med 47: 1202- 1210.

8. Breuer FA, Blaimer M, Mueller MF, Seiberlich N, Heidemann RM, et al. (2006) Controlled aliasing in volumetric parallel imaging (2D CAIPIRINHA). Magn Reson Med 55: 549-556.

9. Lin W, Nielsen T, Qin Q, Mostofsky SH, Duensing R. Real-Time Motion Correction in 2D Multi-Slice Imaging with Through-Plane Navigator; 2013; Salt Lake City, UT, USA. pp. 3751.