Diffusion tensor imaging (DTI) with retrospective motion correction for large-scale pediatric imaging

Abstract

Purpose

To develop and implement a clinical DTI technique suitable for the pediatric setting that retrospectively corrects for large motion without the need for rescanning and/or reacquisition strategies, and to deliver high quality DTI images (both in the presence and absence of large motion) using procedures that reduce image noise and artifacts.

Materials and Methods

We implemented an in-house built GRAPPA-accelerated diffusion tensor (DT)-EPI sequence on 1600 patients between 1 month and 18 years old at 1.5T and 3T. To reconstruct the data, we developed a fully-automated tailored reconstruction software that selects the best GRAPPA and ghost calibration weights; does 3D rigid-body realignment with importance weighting; and which employs phase correction and complex averaging to lower Rician noise and reduce phase artifacts. For select cases we investigated the use of an additional volume rejection criterion and b-matrix correction for large motion.

Results

The DTI image reconstruction procedures developed here were extremely robust in correcting for motion, failing on only 3 subjects, while providing the radiologists high quality data for routine evaluation.

Conclusion

This work suggests that, apart from in the rare instance of continuous motion throughout the scan, high quality DTI brain data can be acquired using our proposed integrated sequence and reconstruction that uses a retrospective approach to motion correction. In addition, we demonstrate a substantial improvement in overall image quality by combining phase correction with complex averaging – which reduces the Rician noise that biases noisy data.

INTRODUCTION

In the past decade, diffusion-weighted MR imaging (DWI) and diffusion tensor imaging (DTI) have emerged as powerful contrast mechanisms in characterizing tissue physiology and disease. When combined with conventional MR, DWI can narrow the differential diagnoses by helping to enable the detection of subtle lesions: while DTI has increased our understanding of white matter structure and connectivity. DWI and DTI play a particularly important role in pediatric neuroradiology, where there are multiple pathologic processes sensitive to diffusion abnormality - including primary or secondary energy failure, toxic-metabolic disorders, seizure, trauma, demyelination, tumors, and infection.

A common problem with all pediatric MR imaging is motion. Often, motion increases as the scan progresses - and thus even cooperative subjects move during extended scan sessions. This is particularly a problem for lengthy DTI protocols, and the situation is exacerbated in small children or neonates. The motion problem of DWI/DTI emanates from the fact that diffusion imaging is highly sensitive to the effects of both plastic and rigid-body motion - even a small degree of motion can cause artifacts and signal loss in the final image. For large motion, the resulting images can be severely degraded, often leading to limited or non-diagnostic studies that need to be repeated. For this reason, at our institution general anesthesia (GA) is used during MRI scans for children between 6 months and 5 years (in some cases even up to 15 years old). For younger infants where GA poses a health risk, the ‘bundle-and-feed’ technique is used, while children older than 5 years are generally asked to remain as still as possible – sometimes to no avail. The GA approach is often associated with considerable logistical effort and costs. Specifically, a pediatric anesthesiologist needs to be available, MR-compatible anesthesia equipment is required, and the child needs to be fasting and cannot have a cold or any other condition that precludes administration of GA. Moreover, DTI is frequently applied to infants born prematurely and who are imaged at term age. In these cases, due to lung immaturity, physicians are often reluctant to administer GA.

Motion artifacts in diffusion imaging using EPI manifest in several ways and include blurring (motion between volumes), signal ‘drop-outs’ (motion during the diffusion preparation) (1,2), aliasing (motion during the calibration scan used to obtain the ghost calibration and GRAPPA weights), and inaccurate FA values (head rotation causing an inaccurate b-matrix) (3). Most diffusion studies therefore use motion-insensitive acquisitions, almost exclusively single-shot EPI. A number of navigator echo-based approaches have been tested (1,2,4,5), though the custom data acquisition and processing overheads have limited their adoption clinically. There have been prospective approaches in the literature which uses the reacquisition of corrupted data with the use of a navigator (6–8). In this work, we implemented and tested a simple and reliable un-navigated GRAPPA-accelerated EPI DTI sequence and retrospective motion correction that could be used on a large cohort of pediatric patients.

This retrospective approach consists of a number of stages. First the selection of GRAPPA and ghost calibration weights from the set of b=0 scans may reduce the chance of calibration errors due to motion. 3D rigid-body realignment with importance weighting is aimed towards reducing inter-volume motion artifacts, along with a volume rejection method to discard corrupted volumes. Finally, b-matrix correction may be necessary in the case of large motion. Our aim was to investigate how much the proposed pseudo-multishot acquisition and reconstruction method reduced the motion sensitivity as well as aided the retrospective motion correction process. The acquisition and each stage of the motion correcting reconstruction was tested using a very large pediatric patient cohort of 1600 subjects at both 1.5T and 3T.

MATERIALS AND METHODS

Patients

The study proposed in this work was approved by the institutional review board of our institution. Between January 2009 and March 2011, 1600 subjects between the age of 1 day and 18 years old were prospectively enrolled into the study after informed consent/assent was obtained. Specifically, 100 patients were scanned at 1.5T and 1500 patients were scanned at 3T. Since the images were to be used diagnostically, the DTI sequence and reconstruction discussed below was incorporated into the standard pediatric brain protocol and the final DTI images were sent to the hospital's patient database for routine radiological assessment. While there was not a dedicated procedure put in place for performing the assessments of each patient, every image was viewed by two readers, one of whom was a pediatric radiologist viewing them in their routine practice. For those patients flagged by either reader as having discernible motion artifacts post-correction, our radiologist classified them as to whether or not they were diagnostic.

MR Pulse Sequence

Our GRAPPA EPI sequence was installed on a 1.5T Twin-speed GE Excite and a 3T GE DVMR750 MR system (GE Healthcare, Waukesha, WI, USA) at our pediatric hospital institution. Both scanners used an 8-channel head coil. DTI data was acquired with an EPI sequence using a GRAPPA acceleration factor of R = 3 and R = 3 interleaves (9), with the following imaging parameters: FOV = 20 – 24 cm, acquisition matrix = 128 × 128, 3 mm slice thickness, 0 mm gap, partial Fourier encoding with 24 extra lines, 25 isotropically distributed diffusion directions with b=1000 s/mm², 5 T₂-w (b=0) images interspersed throughout the acquisition, and TR = 4 – 6s (depending on the slice coverage) corresponding to a scan time = 6 – 9 min. The sequence used here can be described as a single-shot technique with R averages (NEX) -- whereby the averages are different interleaves. Here, the interleaves are separated in the same way as traditional interleaved EPI, however, in this case, each interleave is reconstructed using parallel imaging (using a GRAPPA kernel derived from the fully-sampled R=NEX=3 b=0 image) and acts as a single fully sampled k-space.

Image Reconstruction

Figure 1 outlines in detail the workflow used to reconstruct all GRAPPA DTI data. All image reconstruction was performed using compiled and threaded MATLAB code (version 7.8.0; Mathworks, Natick, MA, USA) installed on the vendor’s multiprocessor reconstruction hardware. Except for the rigid-body correction (which was performed using core routines in the SPM5 toolbox (http://www.fil.ion.ucl.ac.uk/spm/ (10)), all of the processing steps described below used MATLAB code developed in-house .

Figure 1.

Schematic showing the GRAPPA DTI image processing steps as described in the text. The b-matrix correction was performed only on select patients and is therefore not shown in the figure. For a single DTI dataset acquired on a patient (25 diffusion directions/5 b=0 images/3 NEX and approximately 40 slices) this reconstruction chain takes approximately 10 mins to execute including transfer to the hospital patient database (PACS).

i) GRAPPA and Nyquist Ghost Calibration

The image processing begins with parallel imaging and Nyquist ghost calibration performed using the five acquired T₂-w volumes - however, note that, for simplicity, the schematic (Fig. 1) depicts the acquisition of only two T₂-w volumes. Since all volumes (T2-w and DW volumes alike) are acquired with the number of interleaves being equal to the parallel imaging acceleration factor prescribed (R=3 for this study), k-space is fully sampled for each volume, and GRAPPA weights calculation (11,12) can be estimated individually (9,13) for each of the five T₂-w volumes.

Here, note that the acquisition of several interleaves for each T₂-w volume eliminates the need for an upfront calibration scan (and avoids extra scan time of about 30 seconds). Although this can reduce the risk of motion-induced artifacts that may occur between an upfront calibration scan and the following acquisition, motion can occur between the T₂-w interleaves themselves - thereby corrupting the estimated ghost and GRAPPA parameters, leading to an aliasing pattern that propagates through-out the entire dataset. In order to reduce the effects of potential motion during the interleaves used for calibration, the optimal GRAPPA weights were derived from the parameter set of the five T₂-w volumes which yielded the lowest least-squares GRAPPA fit error, ε, to the calibration data, as this has shown to correlate extremely well with the least amount of motion (14). The scalar fit error measure based on an error term, ε, which takes the product of the maximum error for a slice and the average fit error across all slices as follows:

$$\epsilon =\text{max}(\mathit{r})<\mathit{r}>;\{\begin{array}{c}{r}_{\mathit{\text{slice}}1}={(\mathit{y}-\mathit{ŷ})}^2\hfill \\ \mathit{r}=[r_{\mathit{\text{slice}}1},․\dots r_{\mathit{\text{sliceN}}}]\hfill \end{array},$$ [1]

where y is a vector containing all pixels. The optimal GRAPPA weights are then applied to each interleaf of each b=1000 and T₂-w volume.

In a similar manner, the EPI-ghost correction was derived from these T₂-w data sets using an entropy-based, iterative calibration scheme (15). In brief, a constant offset and linearly increasing delay between even and odd bipolar EPI readouts was determined that minimized the total image entropy for each of the five volumes. A single correction pair was chosen based on the error term defined in Eq. 1, and applied to each volume.

ii) Phase correction, coil combination, and partial Fourier reconstruction

Following the Nyquist ghost correction and GRAPPA reconstruction, a total of 90 fully sampled component images (5 × 3 = 15 T₂-w images + 25 × 3 = 75 DWI images) exist. Due to the small random motion that occurs during the application of the diffusion gradients, each of the 75 DW component images has a random spatially non-linear phase imprinted on top of the image phase (16). In order to correct for this and avoid severe artifacts that would arise from combining these complex images, one must either remove the phase completely from each DW image (that is, magnitude operation), or employ a phase correction technique. If the magnitude operation is applied, the final isotropic DWI (isoDWI) will be characterized as Rician noise, which biases the signal with decreasing signal-to-noise level (17–19). If the individual complex-valued DWIs are phase-corrected and combined thereafter, this Rician noise can be minimized.

Here we use a triangular-windowing phase correction (20), modified for partial Fourier data (21). A schematic showing how the triangular-windowing approach removes unwanted phase is shown in Fig. 2a. For this approach, the complex-valued k-space data for each slice is inverse Fourier transformed into two temporary images - one after the application of a triangular window function of radius r in k-space. The phase information content in the corresponding windowed image is subtracted from the non-windowed image. The radius of the triangular window will determine how much phase is removed. From our previous analysis of DWI data acquired on over 1,200 adult patients scanned at our hospital with a FOV of 24cm and acquisition matrix of 192 × 192, we found empirically that a triangular-window radius of 25% of the maximum k-space radius minimized the amount of Rician noise in the final isoDWI data without noticeable residual phase cancellations (22). Fig. 2b illustrates the effect of changing r an adult stroke patient, with r = 0.25 found as an appropriate balance between artifacts and image noise. Here we also used r = 0.25, however further work needs to be done to test the appropriate value of r for the resolution used in this work.

Figure 2.

(a) Simple schematic illustrating the triangular phase correction (20) applied to each DW EPI k-space. A window radius of r (1 being full k-space radius) is used to remove unwanted phase due to motion occurring during the DW gradients, leaving only the image phase from spins and receiver coils. (b) Isotropic DW images (b = 1000 s/mm²) of an adult stroke patient reconstructed with magnitude averaging, as well as complex averaging for various r. Note that here the FOV = 24 cm and the acquisition matrix = 192 × 192 -- both larger than that used in this pediatric study -- and as such, a rigorous analysis of the appropriate r to use for this pediatric study needs to be conducted.

In addition to the reduction of Rician noise, another advantage of performing phase correction is to help reduce image artifacts for partial Fourier data. We have shown that employing phase correction before partial Fourier reconstruction can help to avoid artifacts due to significant brain motion occurring during the diffusion gradients (21). However, the performance of the partial Fourier reconstruction is contingent on a robust measure of the image phase, which is dependent upon both the extent of brain motion and the number of overscans used. In addition, we recommend that one should use at least 16 overscans in EPI to help yield a reasonable measure of the image phase and reduce the extent of motionrelated artifacts (21). (Note that in this work 24 overscans were used). In order to leverage on the improved noise characteristics of complex averaging, one must use a phase-preserving partial Fourier reconstruction such as POCS (23,24), as used here - rather than homodyne reconstruction, which clears the phase altogether. A further advantage of performing POCS is that it performs better than homodyne reconstruction in the presence of brain motion (21).

Prior to partial Fourier reconstruction with POCS, the individual receiver coil images generated by the GRAPPA reconstruction are combined. To preserve the phase in each pixel we take the average phase across all coil elements for each image pixel, while the magnitude is defined by the normal sum of squares. Our phase preserving sum-of-squares formula is therefore defined as (22,25):

$$I_{\mathit{\text{cplxSoS}}}=\frac{1}{N_{\mathit{\text{coils}}}}\underset{\text{magnitude part}}{\underbrace{\sqrt{{\displaystyle \sum _{j=1}^{N_{\mathit{\text{coils}}}}}{\left|I_j\right|}^2}}}\underset{\text{phase part}}{\underbrace{\frac{{\displaystyle \sum _{j=1}^{N_{\mathit{\text{coils}}}}}{I}_j}{\left|{\displaystyle \sum _{j=1}^{N_{\mathit{\text{coils}}}}}{I}_j\right|}}}$$ [2]

where I_j represents each complex coil image, and N_coils is the number of receiver coils. Note that the phase between coil elements is likely to be different for each pixel location (even on ideal, non-diffusion gel phantom data), thus a residual phase remains post coil combination. However, the phase in the final diffusion images is known to be spurious (26) and is consequently discarded after the diffusion images are combined.

Diffusion Image Processing

iii) 3D rigid body realignment

At the stage where each individual volume has been fully reconstructed, phase corrected, and combined over coils - retrospective 3D rigid-body realignment is performed using core routines in the SPM5 toolbox (http://www.fil.ion.ucl.ac.uk/spm/ (10)). Firstly, the T₂-w and the DWI’s (magnitude images) are realigned separately (27). As reference volume for the T2-w and DWI image volumes, respectively, the volume with the highest mean signal was chosen. Secondly, a mutual information based 3D registration between the mean magnitude T₂-w data and the mean magnitude DWI data is conducted (28). Note that the least-squares minimization method was used to align the DW images, despite that they have different contrast due to the different encoding directions. In addition to being a much faster approach than mutual information, we have found empirically that this approach is suitably robust for aligning DW images acquired with the imaging parameters used in this work.

Because of the smaller head size of children and low SNR in the DWIs (Fig. 2a), for the successful realignment of the DWI data it was necessary to enable SPM's weighting option (27), which masks out pixels from being used in the cost function during the estimation procedure. In this work, we used the 3D T2-w image volume with the lowest image entropy, binarized into tissue and air pixels. Finally, the realignment parameters found were used to resample the complex-valued volumes.

iv) DWI volume rejection

The effectiveness of the 3D realignment procedure above relies on the absence of motion within a volume as could occur during either excessive head motion or elevated brain pulsation during the diffusion encoding period. While slice-to-volume image registration has been proposed (29), it is often of little use for DWI because it has less accuracy than 3D registration - and head motion can often create massive artifacts over the entire slice. Figure 3 shows two un-averaged realigned diffusion-weighted volumes on a moving patient. Fig. 3b shows that excessive intra-volume motion can cause several slices to ‘drop out’, which often results in the corruption of an entire volume.

Figure 3.

A single DW volume (without averaging) from a GRAPPA DTI R=3 dataset post image realignment. (a) DWI volume with no intra-volume motion. (b) DWI volume with intra-volume motion, whereby excessive motion results in the corruption of the entire volume. The data is acquired on a 1 month old patient (note the different use of acquisition parameters than described in the text, as follows: FOV = 20 cm, acquisition matrix = 192 × 192, 35 directions, 5 T2-w images, slice-thickness/gap = 5 mm/0 mm, b=1000 s/mm²).

To help deal with this slice ‘drop out’ effect, here we investigated the use of a volume rejection criterion based on the linear correlation coefficient. The mean signal intensities of the slices in the DWI volume in question is correlated with the DWI volume with the highest signal intensity on a per-volume basis. In order to find a ‘rejection threshold’, we selected 35 patients with the maximum detected motion found from the 3D alignment procedure, where up to 15 mm of motion occurred. For all 35 patients, each individual DW volume was reviewed – and those that were corrupted, as defined on visual inspection by significant image drop-outs, were tagged manually. The total number of corrupted volumes for all 35 patients was 105. In order to find a threshold value that rejected DW volumes without rejecting uncorrupted data, correlation coefficients ranging from 0.1 to 0.9 were tested.

(v) b-matrix correction

An additional challenge in DTI is that, in the presence of head rotation each DW volume is potentially encoded with an effective diffusion encoding direction that is different from the one initially prescribed. Since the diffusion encoding is ‘imprinted’ on the magnetization, a simple counter-rotating of the diffusion direction does not suffice. There have been a number of studies examining the effects of the alteration of diffusion encoding direction (i.e. the b-matrix) as a result of motion. For single-shot methods, b-matrix correction can be done simply by multiplying the diffusion direction by the rotational component of the motion (30,31), whereas b-matrix correction for multi-shot DTI requires non-linear methods (32,33). While we did not implement this correction as part of our automated reconstruction chain at our institution, we chose a select patient group (with large amount of detected motion) and employed a non-linear b-matrix correction in this data for demonstration purposes.

vi) Final diffusion processing

With the motion-compensated T₂-w images and DWIs aligned, complex averaging across the T₂-w and DWI volumes can be performed as follows:

$$T_2{w}_{\mathit{\text{mean}}}=\left|\frac{1}{N_{T_{2}w}\mathit{\text{NEX}}}{\displaystyle \sum _{n=1}^{N_{T_{2}w}}}{\displaystyle \sum _{j=1}^{\mathit{\text{NEX}}}}{I}_{\mathit{\text{cplxSoS}},n,j}\right|$$ [3]

$$\mathit{\text{isoDWI}}=\left|\frac{1}{N_{\mathit{\text{dir}}}\mathit{\text{NEX}}}{\displaystyle \sum _{d=1}^{N_{\mathit{\text{dir}}}}}{\displaystyle \sum _{j=1}^{\mathit{\text{NEX}}}}{I}_{\mathit{\text{cplxSoS}},d,j}\right|$$ [4]

where N_T2w is the number of T₂-w images, N_dir is the number of diffusion directions, and NEX is the number of averages. In our case this NEX is equal to 3 (equivalent to the number of interleaves and the GRAPPA acceleration factor). With this approach, the Rician noise in the final isoDWI will be reduced (compared to magnitude averaging) according to the triangular window radius used for the phase correction.

For DTI processing, the post-rejection interleaves (or NEX) for each direction were complex averaged, and the diffusion tensor calculation was performed. Upon completion, the isoDWI, isotropic apparent diffusion coefficient (isoADC), fractional anisotropy (FA), and color FA images were automatically sent back directly to both the scanner and the hospital’s PACS database. In addition, the corresponding 'uncorrected' series of DTI images were sent to PACS in the event that the radiologists would like to compare uncorrected vs. corrected datasets. The total reconstruction time (on the vendor’s multiprocessor reconstruction hardware) was approximately 10 minutes, noting that the reconstruction itself was timed to start the calibration phase as soon as each raw data file became available. Fiber tractography was performed on select patient datasets using the visualization and tracking software called SmartTrack that was built in-house (34). Euler’s method was used for fiber tracking (step size = 0.6 mm, curvature threshold = 40°, FA threshold = 0.15). A seed region was chosen in the splenium of corpus callosum.

RESULTS

The maximum motion found from the 3D realignment procedure on all 1,600 DTI datasets scanned at our pediatric hospital is summarized in Figure 4. We found that the majority (51%) of these patients remained still – either through co-operation (~10%) or general anesthesia (~41%). The remaining 49% of cases moved to varying degrees, as defined by 1 – 15 mm of motion. When using the motion processing chain described here, only 3 (~0.2%) of patients scans would have required a repeat scan.

Figure 4.

Maximum amount of motion detected from the 3D realignment procedure performed on DTI datasets acquired on 1,600 pediatric patients. 51% (816) of patients had minimal motion of 0 – 1 mm. Of those, approximately 41% of patients were administered general anesthesia (GA).

The results of the motion rejection threshold are shown in Figure 5. As shown in Fig. 5, the curves intersect at 0.83 – suggesting that this correlation coefficient is an appropriate threshold sufficient to reject corrupted DW volumes, while minimizing the rejection of uncorrupted data. Note that for each DW gradient direction there are three averages, so the chance that all averages for a given direction will be rejected is small. In the 35 worst moving patients, where up to 15 mm of motion was detected in the image realignment, 6.6% of volumes in total were rejected. For the 4 patients of these 35 with the worst motion detected, at least one diffusion direction was missing after volume rejection. For these particular patients, the number of missing directions (out of 25) and the proportion of volumes that were rejected overall were as follows: 5 (37% rejected), 3 (21%), 2 (19%), and 1 (8%). Excluding the patient with 8% volume rejection, the remaining 3 patients moved excessively and continuously throughout the scan resulting in a 'failed' exam.

Figure 5.

Determination of an acceptable correlation coefficient threshold for the rejection of DW volumes degraded by signal-dropouts (arising from large motion during the diffusion gradients). Here, the individual DW volumes from 35 of the 'worst' moving patients (selected from 1,600 patients using the maximum motion detected from the 3D realignment procedure) were inspected; a total of 105 volumes were tagged manually as significantly degraded by signal-dropouts and used as gold standard for the automatic rejection. For each patient, the correlation coefficient between each DW volume and the reference DW volume (based on the highest mean signal intensity) was calculated, and rejected (or not) according to various thresholds. We selected a threshold of 0.83 (the intersection between the two curves) to be an acceptable correlation threshold to use for rejecting DW volumes.

Figure 6 shows an example acquired on a 1 month old unsedated infant, whereby motion occurred during the T₂-w scans. One slice of each of the five T₂-w scans is shown in Fig. 6a. Inter-shot ghosting due to motion during the acquisition of the first T₂-w image prevents proper estimation of both ghost and GRAPPA parameters from the data. Applying these parameters on a motion free T₂-w interleave later in the scan produces severe artifacts (Fig. 6b). However, with the aid of Eq. 1, GRAPPA weights derived from the least motion hampered T₂-w volume are used to produce artifact free data upon unfolding each interleaf in the data set (Fig. 6c). Note that some parallel imaging calibration methods are also based on the use of an upfront reference scan – which in the presence of motion would lead to corrupted data as in Fig. 6b.

Figure 6.

(a) Based on the least squares GRAPPA fit error, ε (Eq. 1), GRAPPA calibration on motion corrupted T2-w data can be avoided. (b) Bad GRAPPA weights derived from the motion corrupted volume cause artifacts, while (c) using the motion-free calibration data with the lowest ε, does not. Data is acquired on the same 1 month old patient as in Fig. 3.

Figure 7 shows an example acquired on a 4 month old patient, whereby motion of up to 8.8 mm resulted in significant image degradation of the FA maps (Fig. 7a). After 3D realignment, the image quality is greatly improved, with the disappearance of artifacts appearing pathologic on the uncorrected images (Fig. 7b). Following the rejection of bad volumes (Fig. 7c) using a correlation coefficient threshold for rejection of 0.83, the cortical and subcortical white matter definition has been improved, with more normal-appearing cortices and a more clearly defined subinsular region. With the addition of b-matrix correction (Fig. 7d), an improvement in image quality is not visually apparent from the FA maps. Fiber tractography performed on these datasets (Fig. 7e–g) also shows that – while the fiber tracts become more consistent with the reference tracts (in this case, the motion-corrected, volume-rejected, b-matrix corrected tracts color-coded in yellow) – the tracts are relatively robust to erroneous diffusion-encoding due to rigid-body motion.

Figure 7.

4 month old premature infant with brain underdevelopment, where significant motion (up to 8.8 mm) caused both intra- and inter-volume motion artifacts. (a) No 3D realignment, (b) 3D realignment, (c) 3D realignment with bad DW volume rejection (two bad volumes are rejected), (d) 3D realignment with rejection and b-matrix correction. Note the poorly visualized and blurred appearance of the corpus callosum, forceps major and minor, and the capsular white matter (a) as compared to all motion corrected images (b–d). There is also left middle cranial fossa and parafalcine abnormal signal (yellow arrows) not present on motion corrected images. These areas appeared pathologic but were found to be artifacts on corrected scans. Compared to b), c) shows improved cortical and subcortical white matter definition on FA maps. Especially, high intensity which suggest high FA in the cortices of the left frontal operculum, temporal, and inferior frontal lobes (red arrows) show more normal-appearing cortices after rejection. The left external capsule/subinsular region is more clearly defined in c) (white arrows). The fiber tractography calculated for each case is shown in (e–g). Here the fiber tracks color-coded in red are overlayed with the ‘gold standard’ (in this case: motion correction with rejection and b-matrix correction) color-coded in yellow. For each subsequent motion correction step employed, the white matter tracts come closer to the expected paths.

Figure 8 depicts a one month old moving patient whose maximum motion detected was 15.3 mm. Here, isoDWI’s are shown without and with motion-rejecting GRAPPA weights as derived from Eq. 1 (Fig 8a and 8b, respectively). The difference in using complex averaging can be appreciated in Fig 8c versus 8b. After 3D realignment and motion-corrupted DW volume rejection in 8d, a diagnostic-quality DW scan emerges.

Figure 8.

Isotropic DWI of a one month old patient (same patient as in Fig. 3 and Fig. 6). (a) Using GRAPPA weights from the first (motion-corrupted) T2-w volume and magnitude averaging. (b) ‘Optimal’ GRAPPA weights and magnitude averaging. (c) ‘Optimal’ GRAPPA weights with complex averaging. (d) as in (c), including 3D realignment and bad volume rejection (displayed FOV cropped down to ~13 cm). The maximum amount of motion detected from the 3D realignment procedure was 15.3 mm.

Two further clinical examples where motion impaired the image quality are shown in Figures 9–10. Figure 9 shows DTI data acquired on an 8 year old patient with a hydrocephalus post shunt, with a maximum detected rigid-body motion of 1.1 mm. Here, the motion correction significantly improves image quality on both the isoDWI and FA maps. In addition, there appear to be two catheter tracts on the uncorrected isoDWI (white arrows), when in reality there is only one – as depicted on motion corrected images. Figure 10 shows isoDWI images and FA maps of a 6 year old male patient with a coccidiomycosis infection and strokes from infection (maximum detected rigid-body motion of 5.5 mm). Bright areas representing strokes from underlying fungal infection on isoDWI are much more clearly delineated on the motion corrected isoDWI.

Figure 9.

8 year old female patient with a hydrocephalus post shunt. Motion correction significantly improves image quality. The pathology is related to ventricle size, which was enlarged on prior studies. On current study, the ventricles are reduced, and there is a catheter coursing through the brain for shunting. Due to motion, on isoDWI, there appears to be two catheter tracts, when in reality, there is only one, as depicted on the motion corrected isoDWI (arrows).

Figure 10.

6 year old male patient with coccidiomycosis infection and strokes from infection. Bright DWI areas representing strokes from underlying fungal infection are much more clearly delineated on the motion corrected isoDWI.

DISCUSSION

Over the past two years, we have acquired GRAPPA DTI data on over 1,600 pediatric patients using the procedures detailed in this work. 49% of patients moved to varying extents (up to 20 mm). While 28% of patients moved only 1 – 3 mm, we found that this small amount of motion could cause notable blurring artifacts in the uncorrected isoDWI and FA maps. Note that in all cases where motion was detected, the maximum amount of motion did not necessarily reflect the final image quality. This is likely because the value reported makes no inference about whether the motion was a single (or continuous) event.

Here we detailed the workflow used to acquire high quality DTI data, both in the presence and absence of motion. To reduce scan time and improve the chance of reconstructing an alias/ghost-free DTI dataset, GRAPPA and Nyquist ghost calibration data was extracted independently for all T₂-w volumes and the best of each was used. The calibration for parallel-imaging is often carried out with the use of a separate reference scan - which not only increases the total scan time, but also relies on the patient remaining still during this single upfront acquisition. Another approach is to oversample the center of k-space so that the calibration information is contained within each volume - however, the different EPI distortion properties of the central and peripheral regions of k-space in this case causes artifacts in the final GRAPPA reconstructed images (9). Here, acquiring repeated multi-shot T₂-w scans and using this data itself can resolve the problems with both upfront and intrinsic calibration scans - as one can choose from a selection of calibration parameters using the multi-shot T₂-w data, while the extra T₂-w acquisitions can be treated as NEX to boost the SNR in the final DTI.

For all patient data, 3D rigid-body realignment with importance weighting was performed, followed by mutual information based 3D registration of the T₂-w and DWI data. For excessive motion where motion tended to corrupt entire volumes, a volume rejection criterion was used based on the linear correlation coefficient between each volume and that with the highest mean signal intensity. We found that a rejection threshold of 0.83 was sufficient to remove data corrupted from intra-volume motion. For patients who moved to a larger extent (totaling 21% of patients who moved between 3 mm – 20 mm), the prevalence of blurring artifacts and image dropouts arising from intra-volume motion was higher, making rejection of data for higher quality FA maps more important. For motion-free datasets, no volumes were rejected using this procedure.

The scalar metric used to reject data was calculated on a per volume basis – in another words, whole volumes were rejected if the correlation coefficient of a particular volume and the ‘best’ volume was smaller than 0.83. With our GRAPPA driven EPI approach, each diffusion direction is effectively acquired three times, giving R=NEX=3 'chances' to properly acquire a particular diffusion direction. Hence, unless the patient moves continuously throughout the scan, it is rare that a single diffusion direction needs to be excluded. An alternative to whole-volume rejection is the rejection of individual slices. This approach was tested initially, but it was found difficult to use a fixed and adequate correlation threshold across patients and slices. Moreover, following 3D motion correction, a single slice drop-out tends to spread to neighboring slices, further challenging a slice-by-slice rejection scheme (e.g. Fig. 3). Further work, however, is needed to fully test this hypothesis. An alternative approach to data rejection is to use a reacquisition approach, whereby motion parameters are obtained from a navigated acquisition and used to perform real-time updates of the gradient coordinate system (6–8).

Much less detrimental to the final image quality than massive intra-volume motion (causing whole volume drop-outs) is the effect of brain pulsations. Brain pulsation varies randomly from one diffusion direction to the next, particularly in the basal part of the brain. Further improvement to the reconstruction could be first to reject bad volumes that have clearly 'dropped out' from excessive head motion as used here, and then reject individual slices using an approach such as k-space entropy (21) as an alternative to cardiac gating.

We also employed b-matrix correction on select patients (Fig. 7), but found that the b-matrix correction only marginally affected the image quality. At the same time, we would argue that, in the absence of substantial motion artifacts, the marginal improvement of a b-matrix correction is higher and important for more complex diffusion models in the research arena.

In this work we have also outlined other procedures we use to, for example, reduce Rician noise and phase artifacts through the use of phase correction and complex averaging on partial Fourier EPI data. With this we reduce the Rician noise floor in the data, which both increases the image contrast in the isoDWI and, to a lesser extent, biases the quantitative ADC and FA.

Both motion corrected and uncorrected DTI data of each patient were sent back to the patient database for analysis. From visual inspection across hundreds of datasets (observed routinely by several of the co-authors and experienced radiologists who routinely used these data diagnostically) the 3D motion correction procedures applied here did not blur motion-free data. However, no independent measure of motion was used in this study to determine motion-free cases. For those patients who moved according to the realignment procedure, the motion correction steps employed here were robust in correcting for large head motion. In all but 3 (out of 1,600) cases our imaging processing employed here produced images with high quality, and -- compared with the vendor supplied SENSE-EPI -- in many cases improved lesion conspicuity and diagnostic confidence by more clearly delineating the underlying pathology. For the 3 cases that failed, the patients moved excessively and continuously throughout the scan, which renders the correction of these data extremely difficult - despite our proposed motion robust reconstruction chain - making a re-scan necessary.

In conclusion, correcting for different types of motion at different stages of the DTI acquisition is critical to provide DTI scans of diagnostic quality. Here we developed and implemented a clinical DTI technique that is suitable for the pediatric setting. This approach retrospectively corrects for motion, and delivers high quality DTI images using procedures that reduce image noise and artifacts. With the use of our in-house built GRAPPA-accelerated diffusion tensor DT-EPI sequence, 1600 pediatric patients were scanned, and the data were reconstructed using an integrated reconstruction software that selected the best GRAPPA and ghost calibration weights; performed 3D rigid-body realignment with importance weighting; and which employed phase correction and complex averaging to lower Rician noise and reduce phase artifacts. For select cases we investigated the use of an additional volume rejection criterion for inferior DW volumes and b-matrix correction for large motion. We found that a correlation coefficient of 0.83 was a sufficient to reject volumes corrupted from large rigid-body motion while minimizing the rejection of uncorrupted data. Apart from the very rare instance where the patient moved continuously (0.2%) and extensively throughout the scan, the workflow used here to acquire and process DTI data were shown to be extremely robust. In summary, we demonstrate that high-quality DTI imaging of the brain is possible in routine clinical practice, both in the presence and absence of motion using a retrospective approach to image reconstruction.