The Theory section described how prospective motion correction can maintain data consistency during scanning. Although the technique is conceptually simple, the difficulty often lies in the implementation details. Unlike many other MR methods, prospective motion correction involves real-time changes to the scanning process, which complicates the development and testing of the technique. This section describes successful implementations reported to date. The focus is placed on correction of head motion, but several other examples are mentioned.
Obtaining Tracking Data
For brain imaging and spectroscopy, various methods have been used to obtain the necessary head pose information. In this review, we classify these as optical methods, field detection methods, and navigator methods. This classification relates to three fundamentally different ways to obtain information describing the pose of the imaged object. Optical methods are completely independent from the MR sequence timing. Often, they use technology developed for a completely different purpose, such as motion capture for the movie industry. Field detection methods are based on a similar principle to spatial encoding, namely that different points in the scanner bore experience different magnetic fields. This principle can be used to encode the position of a marker. Navigator methods are a more traditional way of obtaining object pose, where the scanner is used for imaging, in one or more dimensions, and the resulting data are compared over time. The three methods defined above are described in more detail in the following paragraphs.
Optical methods include laser systems (14, 15), bend-sensitive optical fibers (23), and camera systems. Camera systems have recently become popular, due to technology improvements in both cameras and computing. Methods successfully used for motion correction include out-of-bore stereo camera systems (18, 24), out-of-bore single camera systems (25), in-bore single camera systems (26, 27), and in-bore systems with multiple cameras (28, 29). Generally, the optical approach began with out-of-bore camera systems [e.g., Zaitsev et al. (18)], but it has now moved to in-bore solutions. Cameras situated out of the MR scanner bore were useful for proof-of-concept studies, due to the lower requirements for MR compatibility, compared to in-bore cameras. However, they require optical line-of-sight and extremely high mechanical stability when the tracking marker is located several meters away from the camera. These practical considerations mean that in-bore tracking options are likely to be the better long-term solution.
All currently used optical systems require a marker. Although “markerless” head tracking would be ideal from a patient handling perspective, sufficient accuracy and speed have not yet been demonstrated. Examples of markers include reflective spheres (18), variations on traditional computer vision approaches [e.g., the “self-encoded” marker of Forman et al. (30)], or new technology such as moiré phase tracking (31), which generates moiré patterns allowing accurate determination of through-plane rotations [previously known as the retro-grate reflector (25, 32)]. The last example allows the use of particularly small targets (diameter 1.2 cm or less) with a single camera and has been shown to be a suitable alternative for conventional three-dimensional (3D) motion capture (33). Of course, in all of these examples, a marker must be rigidly attached to the head. This issue is discussed later in this review in greater detail.
Field detection methods are a completely different approach with a long history in MRI. The scanner gradient fields are measured to localize the object. The method requires the use of a short sequences of pulses to obtain position information from a small sample of MR-visible material fixed inside a miniature receive coil. This approach was first conceived in 1986, by Ackerman et al. (34) for catheter tracking. Dumoulin et al. (35) also pioneered developments in this area. A proof-of-principle study for slice-by-slice prospective motion correction using such a system was published by Derbyshire et al. in 1998 (16). More recent implementations, such as that of Ooi et al. (36, 37), refer to these as “active markers.” Active markers have been used for prospective motion correction in structural brain scans (36) and in echo-planar imaging (EPI) (37). A similar technique has been recently applied to measure gradient waveforms by Barmet et al. (38, 39), who decouple tracking from MR imaging by using RF-shielded probes and separate transmit/receive chains. Recently, they have also demonstrated the possibility of computing the probe position during simultaneous MR imaging by applying “tones” [10 and 13 kHz in Brunner et al. (40)] simultaneously with the conventional gradient waveforms. However, this approach notably perturbs the k-space trajectory, which has to be accounted for in image reconstruction.
Field detection methods require several probes or active markers to be attached to the subject (a minimum of three markers are required, positioned noncollinearly and connected in a rigid arrangement). In Refs.36 and37, marker fixation is achieved by attaching the coils to a headband worn by the subject. There is a slight disadvantage over optical methods here, as the active markers (and hence the subject) are connected to the scanner by wires, which makes patient handling more difficult and could perhaps increase patient anxiety levels. The presence of cables also enhances the difficulties with the rigid marker fixation.
MR navigators are the traditional means of obtaining position information during MRI (Fig. 2a). Recent examples used for motion correction include navigators operating in k-space, such as cloverleaf navigators (41), orbital navigators (42, 43), and spherical navigators (44) as well as image-based navigators, such as PROMO (45) or EPI navigators (46, 47). k-Space navigators repeatedly sample parts of k-space and quantify rotations and translations of the object by measuring rotations and phase shifts in the k-space data. Depending on the trajectory used, this can allow motion quantification in all 6 degrees of freedom. Image-based navigators use low-resolution images or volumes. These generally require longer to acquire than k-space navigators but allow the user to define the region of interest for motion quantification, thus avoiding nonrigid regions (e.g., the neck). Alternatively, it is possible to detect, but not quantify, motion by comparing the relative intensity of a free induction decay signal between multiple receive coils (48). Navigator methods with sufficient accuracy for prospective motion correction all require unused time in the sequence to obtain accurate motion information [e.g., about 48 ms for PROMO (45)], which makes them incompatible with some sequences. This spoils one of the main advantages of prospective correction, namely that the technique can be applied to most MR sequences. Nevertheless, if time in the sequence is available, as is often the case in spectroscopy, this method is very practical. Navigator methods have an advantage over optical tracking and field detection methods, in that they require no additional hardware and that there is no need for a marker to be attached to the subject. This is particularly important in terms of patient handling in clinical MRI.
Figure 2. Methods to obtain head pose information for prospective motion correction. a: Navigator methods use data from the MR scanner, rather than from an external source. Navigators can operate in both k-space (i) and image-space (ii). In both cases, data are repeatedly sampled and then compared between different time points to compute motion parameters. Alternatively, the relative change in signal intensity from multiple coils can be used to detect motion with a simple free induction decay (iii). b: Ideally, a method for use with prospective motion correction would meet all three requirements illustrated by the circles in the Venn diagram. Such a technique has not yet been devised, so the best choice currently depends on the imaging situation and the compromises that can be made in each case. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
Download figure to PowerPoint
Figure 2b provides a rough guide to the strength and weaknesses of different systems. No current approach is perfect, and the best choice of method will depend on the relative importance of the three criteria indicated in the Venn diagram for each imaging situation. A quantitative comparison between the different systems discussed above is not provided here, due to the many different ways of measuring and quoting the parameters that describe their performance (such as accuracy).
Finally, it should be noted that the above description applies mainly to brain imaging. Other tracking systems, not mentioned above, can be applied to track other objects of interest. One example is ultrasound imaging to track the position of organs inside the abdomen (49, 50).
Data Transfer and Transformation
Regardless of the tracking modality used (optical, field detection, or navigators), a key component of prospective correction involves the transfer of the pose estimation data to the imaging sequence. However, changing the tracking modality makes a significant difference to how this is performed. Navigator techniques, for example, often use the feedback facility made available by the scanner manufacturer; as this is vendor specific, it is not discussed further here.
In the case of external tracking systems, pose data are often computed on an external computer and sent to the scanner computer using a network connection. For this purpose, data are sent via user datagram protocol (UDP) [e.g., (18, 25)] or transmission control protocol (TCP) [e.g., (26, 49–51)]. UDP is perhaps better suited for real-time applications, as it puts more emphasis on timeliness than on reliability. Indeed, there is little purpose in resending missing packets, as a late packet will become redundant, due to new tracking information. The importance of minimizing latency is described later in this article.
For external tracking systems, the coordinates of the tracked object must be transformed into the coordinate system of the MR scanner. This process is trivial once the correct transform is known. Following the terminology introduced in Zaitsev et al. (18), we refer to the process of determining the transform as cross-calibration.
There are a number of ways in which this cross-calibration procedure can be performed. Aksoy et al. (27, 30, 52) use a 60 s cross-calibration procedure based on a precisely manufactured marker (30) that is visible to both the scanner and the camera. Two other common approaches involve recording motion of a phantom using both the tracking system and the MR scanner (using image registration). Depending on the exact implementation details, we call these approaches iterative or noniterative. The noniterative approach involves collecting numerous datasets and solving for the transform that best fits the data [e.g., as described by Kadashevich et al. (53)]. The iterative approach, as described in Ref.18, applies prospective correction using the latest version of the transform. If the transform is accurate, the resulting images will be perfectly aligned, due to motion correction. If the transform is inaccurate, then errors in the image alignment will result; these are used to fine-tune the transform. In our experience, calibration based on image registration can produce very good results, but there are several confounding effects to be aware of. These include field distortions (caused by rotating the phantom during calibration), gradient nonlinearities, and imperfect fixation of the tracking marker to the phantom. These issues are similar to general limitations of prospective motion correction, which are discussed later in this review.
In Ref.52, Aksoy et al. describe a hybrid prospective and retrospective correction method to mitigate the effect of cross-calibration errors. This involves retrospectively finding a transform by minimizing image entropy in a similar way to previous work by Atkinson et al. (54, 55). As k-space lines are rotated off the Cartesian grid, a gridding reconstruction (56) is used to resample the data. Results show that application of the retrospective stage significantly improves image quality by reducing artifacts caused by poor cross-calibration.
Imaging Volume Update
To perform prospective motion correction of a moving object, the gradient and RF fields are adjusted so that the imaging volume follows the observed motion. This process is nothing more than the position update that is already applied at the start of imaging to set the position of the field of view (FOV). Details concerning adjusting this “on the fly” are specific to the scanner used. For the three main manufacturers, more information can be found in the literature, for example, Siemens in Zaitsev et al. (18), Philips in Manke et al. (20) and Ooi et al. (36), and GE in Qin et al. (26). These examples all describe “inter-view” correction, where adjustments are made between spin excitations. For MR sequences where the time between excitation and signal readout is short relative to the motion expected, a coordinate update prior to each excitation pulse is usually sufficient. However, when additional signal encoding such as diffusion weighing is used, a more sophisticated correction scheme can make sense, as motion during the strong and enduring diffusion gradients leads to severe motion artifacts and signal dropouts. This “intraview correction,” was suggested, but not implemented, by Nehrke and Börnert in Ref.22. Recently, however, Herbst et al. (57) have demonstrated a practical implementation of intraview correction on a Siemens system and have shown it to prevent signal dropouts in DWI.
Peripheral nerve stimulation and technical limits of the gradients mean that clinical MR scanners typically impose “hard limits” on the gradient strength and slew rate. This is relevant to the real-time imaging volume update, as these limits might be violated after transformation of the gradient waveform. Normally, such a violation would lead to an abortion of the pulse sequence, so having a mechanism in place to prevent this is essential. One approach is by incorporating an extra safety margin when specifying the initial maximum gradient strength (22).
An alternative to updating the imaging volume is data reacquisition when motion is detected. This approach has been actively pursued by Kober et al., who use their “free induction decay navigators” (48) to detect head motion exceeding a predefined threshold. Motion correction by data reacquisition increases scan time and cannot be considered to be “prospective motion correction.” It also requires that the imaged object returns to its original position after motion, which is unlikely to be the case in brain imaging. However, a hybrid approach where detected motion triggers the acquisition of an extra EPI volume to quantify the motion parameters and apply prospective correction has also been developed (58). This appears to be a promising compromise.
Figure 3 shows a typical application of prospective motion correction. Data were collected at 1.5 T with a 3D gradient echo sequence, modified by the authors to allow prospective motion correction. Initially, the subject was asked to remain as still as possible. Then the subject was instructed move between two predefined positions, whenever prompted to by the scanner operator. In both cases, the subject was imaged twice: once without and once with prospective motion correction. Prospective motion correction improves image quality under motion and maintains it in the “no motion” case. Figure 4 shows the motion that occurred in each of the corresponding experiments in Fig. 3. The careful examination of tracking data is important in the evaluation of prospective motion correction techniques, as it can be difficult to ensure that motion is consistent between imaging experiments.
Figure 3. Prospective correction results obtained at 1.5 T as a demonstration for this article. Head motion was quantified in real time using the single-camera moiré phase tracking system reported in Ref.31. In (a) and (b), the subject did not perform any deliberate motion; in (c) and (d), the subject performed a series of repeatable head movements (predominantly left-right rotations). Prospective correction is off in (a) and (c) and on in (b) and (d). The motion occurring during each of the four scans in (a)–(d) is shown in Fig. 4.
Download figure to PowerPoint
Figure 4. The motion that occurred during the acquisition of the corresponding images is shown in Fig. 3. In (c) and (d), the scanner operator instructed the subject to rotate his head at particular points in the scan; this process ensured repeatable motion. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
Download figure to PowerPoint
Figure 5 demonstrates the application of intraview (or “continuous”) prospective correction in DWI, where additional correction updates are applied during the diffusion-encoding gradients [Herbst et al. (57)]. This technique adapts all gradients to motion occurring during the encoding process (Fig. 5a). Figure 5b shows results obtained using the method. The continuous correction of the diffusion gradients and the refocusing pulses significantly reduces motion artifacts, and signal dropouts are prevented even when strong motion occurs. Intraview correction might also have application in sequences with long echo trains, such as fast spin echo [RARE (59)] and related techniques, such as SPACE(Siemens)/CUBE(GE) or HASTE, as described in Ref.60.
Figure 5. Prospective motion correction can prevent signal dropouts in DWI. a: The “continuous correction approach” described by Herbst et al. (57), where correction updates are applied during the diffusion-encoding gradients. b: Results from four in vivo experiments showing magnitude images on the left and the respective phase images on the right. Row 1: no motion and no correction. Row 2: strong motion and no correction. Row 3: strong motion and slice-by-slice prospective motion correction. Row 4: continuous correction. These results are reproduced from Ref.57.
Download figure to PowerPoint
Many promising in vivo results involving brain imaging and spectroscopy of volunteers have been published in the last three years. In proton magnetic resonance spectroscopy, prospective motion correction has been shown to reduce spectral artifacts and quantitation errors in choline/creatine ratios (25) and reduce lipid contamination and increase spectral reproducibility (61). It has also been combined with interleaved reference scans (Ref.62) to correct for both motion and motion-induced B0 offsets (63), and, in the case of Hess et al. (64, 65), combined with first-order shim correction. A similar methodology to that used in Ref.63 has been applied to spectroscopic imaging (chemical shift imaging), where it can prevent data degradation under motion (66). Prospective motion correction has also been shown to be potentially useful in fMRI (24, 37) and diffusion tensor imaging (DTI) (27).
Volunteer studies with prospective motion correction are becoming increasingly common. One implementation of prospective motion correction, PROMO (45), has been evaluated in healthy children (2, 67), where it was shown to improve the quality of 3D anatomical imaging. However, little research has been performed in patient populations. Even less validation has been performed with prospective correction using an external tracking system. This translational step is an obvious goal for future work.