Minimizing Susceptibility-Induced BOLD Sensitivity Loss in Multi-Band accelerated fMRI using Point Spread Function Mapping and Gradient Reversal

Abstract

Objective:

Interleaved reverse-gradient fMRI (RG-fMRI) with a point-spread-function (PSF) mapping-based distortion correction scheme has the potential to minimize signal loss in echo-planar-imaging (EPI). In this work, the RG-fMRI is further improved by imaging protocol optimization and application of reverse Fourier acquisition.

Approach:

Multi-band imaging was adapted for RG-fMRI to improve the temporal and spatial resolution. To better understand signal dropouts in forward and reverse EPIs, a simple theoretical relationship between echo shift and geometric distortion was derived and validated by the reliable measurements using PSF mapping method. After examining practical imaging protocols for RG-fMRI in three subjects on both a conventional whole-body and a high-performance compact 3T, the results were compared and the feasibility to further improve the RG-fMRI scheme were explored. High-resolution breath-holding RG-fMRI was conducted with nine subjects on the compact 3T and the fMRI reliability improvement in high susceptibility brain regions was demonstrated. Finally, reverse Fourier acquisition was applied to RG-fMRI, and its benefit was assessed by a simulation study based on the breath-holding RG-fMRI data.

Main Results:

The temporal and spatial resolution of the multi-band RG-fMRI became feasible for whole-brain fMRI. Echo shift measurements from PSF mapping well estimated signal dropout effects in the EPI pair and were useful to further improve the RG-fMRI scheme. Breath-holding RG-fMRI demonstrated improved fMRI reliability in high susceptibility brain regions. Reverse partial Fourier acquisition omitting the late echoes could further improve the temporal or spatial resolution for RG-fMRI without noticeable signal degradation and spatial resolution loss.

Significance:

With the improved imaging scheme, RG-fMRI could reliably investigate the functional mechanisms of the human brain in the temporal and frontal areas suffering from susceptibility-induced functional sensitivity loss.

1. Introduction

Gradient-echo (GE) based echo-planar imaging (EPI) has been widely used for functional magnetic resoance imaging (fMRI), which plays an essential role in basic neuroscience and clinical practice. Due to its high sensitivity to field inhomogeneity, it is prone to signal dropout and geometric distortion, especially in brain regions of rapid susceptibility changes. The signal loss can cause inadequate blood-oxygen-level-dependent (BOLD) signal and compromise fMRI studies in the inferior frontal, the medial temporal, and the inferior temporal lobes (Visser et al., 2010; In et al., 2020). In addition, local geometric distortions can lead to misalignment with the anatomical image, which deters fMRI interpretation.An adequate solution is needed to improve fMRI reliability.

EPI geometric distortion can be effectively reduced with several approaches including image reconstruction-based method (Yarach et al., 2017), field mapping (Jezzard and Balaban, 1995), and point-spread-function (PSF) mapping (Zaitsev et al., 2004; In and Speck, 2012). However, these methods cannot correct signal loss in areas with pronounced susceptibility. A reverse gradient (RG) approach using PSF-based distortion correction (In et al., 2015) was recently proposed as an alternative approach to minimize geometric distortion as well as signal loss. In this method, a pair of EPI with the opposite (forward and reverse) phase-encoding gradients were measured and combined after distortion correction (In et al., 2017). It takes advantage of the fact that the areas of stretching and compression in EPI images and related signal dropouts complement each other in the forward and reverse EPI scans. As the temporal resolution of the interleaved RG scheme is reduced by half, it has only been applied in animal study, which the brain coverage is relatively smaller (In et al., 2017). This study improves the temporal spatial resolution of the RG scheme to make it feasible for whole-brain fMRI in humans.

Local background field gradients cause the k-space echo peak to deviate from the pre-defined center along the phase-encoding direction and result in signal dephasing (Chen et al., 2006; De Panfilis and Schwarzbauer, 2005). Local echo shift measurements can provide useful information to optimize the imaging parameters for high susceptibility regions. To effectively compensate the signal loss with the improved RG scheme, it is essential to understand the relationship between the echo shift and geometric distortion and reliably measure both.

Without considering the echo shift in local areas, traditional partial Fourier (PF) acquisition used for high-resolution imaging by omitting the early echoes could cause signal dropouts in regions with stretched distortions (Chen et al., 2006) and undermine the prerequisite of the RG scheme. In contrast, a reverse PF acquisition omitting the late echoes can alternatively be used to avoid this phenomenon, which will be evaluated here.

A compact 3T (C3T) MRI scanner equiped with high-performance gradients (700 T/m/s, 80 mT/m) (Foo et al., 2018; Lee et al., 2016; Weavers et al., 2016; In et al., 2020) has a great potential for advanced head imaging such as high-resolution fMRI. The high-performance gradient system can reduce the distortion by up to 40% compared to the conventional whole-body 3T (WB3T) with a standard gradient system (200 T/m/s, 50 mT/m) (Tan et al., 2016; Kang et al., 2020). Nevertheless, the RG scheme can still be useful to further minimize the signal loss and geometric distortion for high-resolution fMRI even on high-performance scanner, which will also be demonstrated in this study. In addition, to investigate the benefit of reverse PF acqusition for high-resolution fMRI on the WB3T, the C3T data was utilized for PF simulation and validation as the corresponding high-resolution fMRI data is achievable without partial Fourier acquisition.

In this work, we demonstrated the feasibility of the multi-band (Setsompop et al., 2012) RG scheme for whole-brain fMRI study in humans on both the WB3T and C3T scanners. Firstly, the PSF mapping method was used to correct geometric distortions and also to estimate signal dropouts in forward and reverse EPIs. After collecting the EPI pair and corresponding PSF data with practical imaigng protocols, the signal dropouts and echo shift maps were compared, and the feasibility of further improving the RG scheme was explored. Breath-holding was chosen as a model fMRI stimulus and in-vivo high-resolution task fMRI with RG EPI (RG-fMRI) was conducted on the C3T. Cortical surface-based analysis was utilized to demonstrate the effectiveness, in terms of the improvements on image quality and BOLD signal detection at both individual and group levels. Finally, reverse PF acquisition was also assessed to demonstrate its benefit for RG-fMRI.

2. Theory

2.1. Local echo shift and geometric distortion in GE-EPI

Local susceptibility gradients lead to echo shift (or effective echo time). With a given phase-encoding gradient, G_y, and local susceptibility gradient, G_s, the relationship between the desired or pre-defined echo time (TE), $T{E}_{\text{des}}$, and obtained effective TE, ${TE}_{\text{eff}}$, can be obtained from Eqs. 9 and 13 in a previous study (Chen et al., 2006).

$$\gamma {\int }_{T{E}_{\text{des}}}^{T{E}_{\text{eff}}}{G}_{\text{y}}\left(t\right)dt+\gamma G_{\text{s}}T{E}_{\text{eff}}=0,$$ (1)

where $\gamma$ and t are gyromagnetic ratio and time, respectively. When a constant phase-encoding gradient over the time between ${TE}_{\text{des}}$ and ${TE}_{\text{eff}}$ is assumed, Eq. 1 can be simplied to:

$$T{E}_{\text{eff}}=\frac{G_{\text{y}}}{G_{\text{y}}+G_{\text{s}}}T{E}_{\text{des}}.$$ (2)

This equation demonstrates that $T{E}_{\text{eff}}$ is shorter than $T{E}_{\text{des}}$ if the polarity of the applied phase-encoding and local susceptibility gradient is matched, and vice versa.

In addition, ${\Delta \text{k}}_{\text{y}}$ defines the image field of view (FOV), i.e., $FO{V}_y=1/\Delta k_{\text{y}}=1/\left(\gamma G_{\text{y}}\Delta t\right)$. When the effective ${\Delta \text{k}}_{\text{y}}$ exceeds the targeted ${\Delta \text{k}}_{\text{y}}$ due to the additional G_s with polarity matched with G_y, the local effective FOV is reduced, which results in signal stretching. In contrast, local signal compression is observed with the G_s polarity mismatched with G_y.

2.2. Echo shift map using PSF mapping

Although the echo shift can be directly calculated from the phase data (Deichmann et al., 2002), it is not possible to estimate the effect in signal dropout areas. In PSF mapping method (Zaitsev et al., 2004; In and Speck, 2012), an additional spin-warp (i.e., distortion-free) phase-encoding gradient is added to understand how an undistorted image is transformed to a distorted image in both the phase-encoding dimensions and thus to correct the EPI distortion. In this work, since PSF mapping is used, the corresponding phase map can be calculated based on the shift (or distortion) map in the phase-encoding dimension (In, 2012).

$$\varphi \left(r\right)=\frac{2\pi \cdot \Delta y\left(r\right)\cdot T{E}_{\text{des}}}{N_{\text{y}}\cdot T_{esp}},$$ (3)

where r indexes a particular voxel, $\varphi \left(r\right)$ is the unwrapped phase for each voxel at $T{E}_{\text{des}}$, $\Delta y\left(r\right)$ is the voxel shift in the phase-encoding direction, $T_{esp}$ is the echo spacing time, and N_y is the number of voxels in the phase-encoding y direction. Then, $\Delta \varphi \left(r\right)$ can be calculated as the phase deviation between two adjacent lines along the y direction (Deichmann et al., 2002), which can be translated into a local effective TE map, $T{E}_{\text{eff}}$, and echo shift map in k-space, ${\Delta }_{k_y}$, respectively, with

$$T{E}_{\text{eff}}=T{E}_{\text{des}}+\left(T{E}_{\text{des}}-T_0\right)\cdot \frac{\Delta \varphi }{\pi }$$ (4)

and

$${\Delta }_{k_y}=\frac{N_y}{2}\cdot \frac{\Delta \varphi }{\pi }$$ (5)

where $T_0$ is the time delay from the isocenter of radio frequency excitation pulse to the beginning of the EPI readout. While the maps calculated from EPI data alone (Deichmann et al., 2002) can measure the echo shift ${\Delta }_{k_y}$ within the EPI readout acquisition window, in our method the measuring range is twice as wide due to the additional spin-warp phase-encoding gradient for PSF mapping, which can measure the echo shift ${\Delta }_{k_y}$ even when it is beyond the readout acquisition window.

3. Material and methods

3.1. Data acquisition and reconstruction

All volunteer exams (4 male, 8 female, age 24~57) were performed after obtaining written informed consent using a protocol approved by Institutional Review Board on both a conventional WB3T (GE 750, GE Healthcare, Waukesha, WI) with 200 T/m/s and 50 mT/m gradient specifications, and a high-performance (700 T/m/s, 80 mT/m) C3T scanner (GE Global Research Center, Niskayuna, NY) (Foo et al., 2018; Lee et al., 2016; Weavers et al., 2016; In et al., 2020) using a 32-channel brain coil (Nova medical, USA).

To examine signal dropouts in EPI and imaging protocols for RG-fMRI, PSF mapping sequence and corresponding EPI pair were acquired at 5 different isotropic resolutions: 2.0, 2.2, 2.5, 2.7, and 3.0 mm with a high multi-band (Setsompop et al., 2012) factor of 6 and no in-plane acceleration. The detailed protocols are provided in Table 1. Three healthy subjects were scanned on a conventional WB3T scanner. One subject was scanned for comparison on both the WB3T and the C3T scanners (Foo et al., 2018; Lee et al., 2016; Weavers et al., 2016; In et al., 2020).

Table 1.

Imaging protocols for PSF mapping and EPI scans with reverse gradient approach. Partial Fourier factor was identical for both EPI and corresponding PSF scans, except for 2.2 mm isotropic resolution imaging on whole-body 3T scanner.

Scanner	Whole-body 3T					Compact 3T
Resolution [mm]	2.0	2.2	2.5	2.7	3.0	2.0	2.2	2.5	2.7	3.0
PSF, TR [ms]	1000	1200	850	770	700	934	800	596	548	500
PSF, TE [ms]	17.1	36.5	30.5	27	23.6	24.4	23.1	20.4	18.3	16.2
PSF, reduced FOV	3	3	3	2	2	3	3	3	2	2
PSF, scan time [mins:secs]	0:40	0:43	0:27	0:34	0:28	0:34	0:29	0:20	0:24	0:20
EPI/fMRI*, TR [ms]	1150	1050	850	770	700	991/1000*	842	684	660	552
EPI, TE [ms]	30	30	30.5	30	30	30	30	30	30	30
EPI, repetitions	100	100	100	100	100	100/220	100	100	100	100
EPI/fMRI*, scan time [mins:secs]	1:55	1:45	1:25	1:17	1:10	1:39/3:50*	1:24	1:08	1:06	0:55
Matrix	120	108	96	88	80	120	108	96	88	80
Slices	79	72	63	58	52	79	72	63	58	52
Echo spacing [μs]	644	600	552	524	492	388	364	356	340	324
Receiver bandwidth [kHz]	250	250	250	250	250	250	250	250	250	250
Multiband factor	6	6	6	6	6	6	6	6	6	6
Partial Fourier in EPI (PSF)	5/8	5(8)/8	8/8	8/8	8/8	8/8	8/8	8/8	8/8	8/8

^*Note that 2.0 mm isotropic resolution protocol was adapted for breath-holding RG-fMRI on the C3T scanner after changing the TR from 991 to 1000 ms in EPI.

Breath-holding task RG-fMRI was performed on nine healthy volunteers only on the high-performance C3T scanner to enable high- (i.e., 2.0 mm isotropic) resolution imaging without PF acquisition. As a model fMRI stimulus, breath-holding was chosen to oberve functinoal contrasts over the entire brain areas. Clinical breath-holding fMRI protocol used at our institution was adopted to demonstrate its feasibility for clinical fMRI studies. The block paradigm consisted of 5 repeats of rest (20 s) and breath-holding (20 s) cycle and ended with a rest period of 20 s. The task timing was cued by visual stimulus, and both the timing and the extent of breath-holding were monitored using a respiration belt. Keeping TR within 1 s, the highest spatial resolution leading to strongest signal distortion was chosen to demonstrate the effectiveness of the RG scheme. The phase-encoding polarity of each EPI repetition was altered in an interleaved order, which makes effective fMRI sampling interval as double TRs (i.e., 2 s). Before RG-fMRI scan, a pair of PSF mapping scans with opposite phase-encoding polarities (In et al., 2015; In et al., 2017) was performed with identical imaging parameters as the RG-fMRI, except for TR/TE = 934/24.4 ms. Reduced FOV by skipping the spin-warp phase-encoding line in PSF scan was applied for acceleration, which is applicable unless remaining FOV is larger than the level of geometric distortions in EPI (Zaitsev et al., 2004).

All PSF and fMRI data were acquired only in the axial plane. Additionally, for all the in-vivo exams, a T1-weighted anatomical image was acquired with the following parameters on the compact 3T scanner: TR/TE/TI = 7.6/2.5/900 ms, flip angle = 8°, 1.0 mm isotropic resolution, scan time = 5:04 minutes.

The data reconstruction was performed offline using MATLAB (The MathWorks, Inc. USA). After PSF mapping-based distortion correction (In et al., 2015; In et al., 2017; In and Speck, 2012), a total of five different variants of the EPI series were compared, which included individual forward and reverse phase-encoded EPIs without (F and R) and with distortion correction (F_C and R_C) and the weighted combination of the distortion-corrected EPI pair (C_C). The square root of the sum of the squares of the two distortion-corrected images resulted in a combination weighted by the image intensity at that location (In et al., 2017). For qualitative comparison of EPI distortion correction, a T2*-weighted gradient echo image was reconstructed from the PSF data as the reference (Zaitsev et al., 2004; In and Speck, 2012).

3.2. Data analysis

3.2.1. Spatial and temporal resolutions for multi-band RG-fMRI

The effective temporal resolution and signal recovery performance on both scanners were evaluated qualitatively and quantitatively to explore optimum imaging protocol for task fMRI. Temporal SNR (TSNR), mean cortical coverage ratio, and Dice similarity coefficient were computed. After co-registration between the functional and the anatomical images with rigid-body transformation using AFNI (Analysis of Functional NeuroImages; http://afni.nimh.nih.gov) software package (Cox, 1996; Cox and Jesmanowicz, 1999), all functional data were interpolated to 2.0 mm resolution, and masked based on the TSNR map. Then, the Dice similarity coefficients (i.e., the Sorensen-Dice index) were calculated using AFNI. After labelling brain regions based on anatomical image using FreeSurfer software (Reuter et al., 2012), the cortical coverage ratio between EPI and anatomical volume were calculated and evaluated, as performed in a previous study (Kang et al., 2020).

3.2.2. Echo shift maps from EPI and PSF mapping

Echo shift and corresponding signal dropout in the EPI pair were compared using estimated echo shift maps from EPI (Deichmann et al., 2002) and PSF mapping using Eq. 5. An additional echo shift map was obtained from PSF mapping in the spin-warp phase-encoding (i.e., non-distorted) dimension to predict the signal dropouts in the non-distorted space. Since the distoriton correction is considered as a conversion from distorted space to non-distorted space, this echo shift map was compared with signal dropouts in the distortion-corrected EPI pair (F_C and R_C).

3.2.3. Breath-holding RG-fMRI

The RG-fMRI processing pipeline was performed separately for the forward and reverse EPI series using the AFNI software (Cox, 1996; Cox and Jesmanowicz, 1999), which included spike removal (3dDespike), slice timing correction, motion correction (3dvolreg), spatial smoothing (3 mm FWHM), and co-registration between functional and anatomical images. Additional slice timing correction was included only in the reverse EPI series to address the slice timing difference (i.e., one TR or 1 second) between the pair series before combination. The PSF-based distortion correction (In et al., 2015; In et al., 2017) was applied after motion correction of EPI time series. BOLD signal changes in response to breath-holding were predicted with beta and t-statistics derived by generalized linear model (GLM). To clearly demonstrate the improvements in local BOLD contrasts arising from the improved signal redistribution rather than the increased image SNR due to image combination, signal percentage changes between the breath-holding and resting period were also calculated, which were not directly associated with the image SNR due to the identical number of signal baselines and stimulation blocks in all of the distortion-corrected EPIs including F_C, R_C, and C_C (In et al., 2017).

For the surface-rendered group analysis, each cortical boundary derived by FreeSurfer (Reuter et al., 2012) was used to project functional data onto a 3D standard-mesh cortical surface model (Saad et al., 2004; Kang et al., 2020). The signal coverage, the BOLD contrast, and the signal percentage change were demonstrated and compared with the surface model. To evaluate the fMRI detection improvements in the regions with strong susceptibility, group functional contrast maps from t-score and percentage change calculations were compared by two-way analysis of variance (ANOVA) tests across all of the distortion-corrected EPI series including F_C, R_C, and C_C. A p-value <0.05 was considered to be significant.

3.2.4. Reverse PF acquisition for RG-fMRI

In contrast to standard/forward PF acquisition, which can cause additional signal dropouts by omitting echoes deviated from the k-space center in the high susceptibility areas, reverse PF acquistion can be useful to further minimize the EPI acquistion window without observable image degradation. To test this, simulation of standard and reverse PF was performed based on the breath-holding RG-fMRI data obtained on the C3T with full k-space acquisition and the results were compared. Three different PF factors including 5/8, 6/8, and 7/8 were applied and corresponding unsampled k-space areas were zero-filled before reconstructing the images. In high-susceptibility areas with echo shifts, the loss of the image intensity and functional contrast with different PF scheme and factors was compared with the full acquisition data.

4. Results

4.1. Spatial and temporal resolution for multi-band RG-fMRI

Figure 1A shows that the quantitative metrics used to evaluate F_C, R_C, and C_C images. The combined images (C_C) perform best among all image datasets regardless of spatial resolution. TSNR tended to increase with the voxel size as expected, and high-resolution imaging up to 2.5 mm isotropic resolution offered a better coverage ratio. Due to the shorter TR on the C3T, TSNR was conceivably lower than that of the counterpart on the WB3T (Fig. 1B-a). In addition, since TSNR map-based mask was used to calculate the coverage ratio and Dice coefficient, the mask volume increased with higher TSNR on the WB3T scanner consistently. Although slightly higher in mean coverage ratio (Fig. 1B-b), the data on the WB3T scanner had lower Dice coefficient (Fig. 1B-c), compared to those on the C3T. For finer than 2.5 mm resolution imaging, however, the ratios and coefficients were noticeably reduced on the WB3T as a 5/8 PF acquisition was needed to keep TE at 30 ms, which could not cover the effective echo shifts in high-susceptibility areas and resulted in signal dropout. In comparison, the use of high-performance gradients on the C3T greatly reduced the echo spacing (from 644 to 388 ms), which advantageously allowed 2.0 mm resolution data acquisition with a shorter TR (<1 second) and without PF acquisition (Table 1 and Fig. 1B). Higher cortical coverage ratio was achieved with high-resolution imaging up to 2.0 mm isotropic resolution. With a multi-band factor of 6 without PF acquisition, an effective temporal resolution less than 2 s was achieved, for whole-brain 2.5 mm isotropic resolution fMRI on the WB3T and 2.0 mm for the C3T scanners, respectively (Table 1).

Figure 1.

Quantitative evaluation of the RG approach with different isotropic spatial resolutions on the whole-body 3T (WB3T) scanner (A) and comparison with the compact 3T (C3T) scanner in one subject (B): (a) TSNR), (b) Dice coefficients, and mean coverage ratios (c) for the entire brain cortices and (d) only for the frontal and temporal cortices.

4.2. Echo shift maps from EPI and PSF mapping

Figure 2 depicts the difference in compressed areas between the echo shift maps based on both EPI and PSF methods, ESM_EPI and ESM_PSF. As indicated by Eq. 2, a shorter effective TE was measured in the stretched area and a longer one was measured in the compressed areas of EPI images regardless of the phase-encoding polarity. This corresponds to negative echo shift in Fig. 2B and positive shift in Fig. 2C. In the ESM_PSF of the forward EPI (Fig. 2B), strong echo shifts exceeding half of the k-space acquisition window (i.e., >±48) were measured in strongly compressed regions, which resulted in regional signal dropouts. In contrast, the estimated values from the ESM_EPI is limited to within the acquisition window, even in the regions with complete signal loss (Fig. 2B). In reverse EPI, however, the measured echo-shifts were very similar in both the ESM_EPI and the ESM_PSF (Fig. 2C) because stretched distortions appeared dominantly. Moreover, the PSF mapping method provided the echo shift map in the non-distorted (i.e., spin warping phase-encoding) dimension (Fig. 2A).

Figure 2.

Images and distortion maps from the non-distorted (A) and the distorted forward (B) and reverse (C) phase-encoding space are shown. Echo shift map calculated from the PSF mapping data (ESM_PSF) or directly from EPI pair (ESM_EPI) are also shown. It is noted that the scale of ESM_PSF is ranging from −96 to 96 as shown in (A) but is matched to the scale of ESM_EPI (−48 to 48) to clearly demonstrate their difference in (B) and (C). Note that 2.5 mm isotropic resolution data on the WB3T scanner were chosen for this demonstration.

A qualitative evaluation of the proposed reverse gradient scheme in the temporal lobes of human brain is presented in Figure 3. Severe geometric distortions appeared in the measured forward (F) and reverse EPI (R) were corrected in distortion-corrected forward (F_C) and reverse (R_C) EPI. After combining the corrected pair, signal loss was greatly minimized in the combined EPI (C_C). Both the image intensity and the geometry were very similar to the undistorted reference image obtained from the PSF mapping data. The distortion correction on the WB3T scanner and the comparison between the WB3T and C3T scanners are shown in supplementary figure 1. The overall pattern of the intensity difference between the distortion-corrected EPI pair matches well with ESM_PSF (see contours in Fig.3).

Figure 3.

Difference of signal dropouts in the EPI pair with opposite phase-encoding polarities: forward (F) and reverse EPI (R), distortion-corrected forward (F_C) and reverse EPI (R_C), combined EPI (C_C), undistorted reference (Ref.), subtraction image between F_C and R_C (F_C-R_C), and PSF-based echo shift map (ESM_PSF). Green- and red-colored contours calculated from the ESM_PSF indicate the areas of significant signal loss in the either F_C or R_C, respectively and are overlaid onto the F_C, the R_C, the F_C-R_C, and the ESM_PSF. The threshold value (±19.3) for generating the contours is manually chosen by visual inspection. Note that 2.5 mm isotropic resolution data on the WB3T scanner were chosen.

4.3. Breath-holding RG-fMRI

Figure 4 demonstrates that the BOLD contrast improvement by minimizing signal loss in temporal and frontal cortices using RG-fMRI. Since the breath-holding period was considered as an onset for task fMRI, all BOLD contrasts calculated from both the GLM and signal percentage change were negative. The susceptibility-induced signal dropout was considerably different between the R_C and F_C in regions of temporal (Fig. 4A) and frontal lobes (Fig. 4B) and varied across the slices, which resulted in mismatched activation maps in the affected areas (see arrows). Nevertheless, the signal dropout and corresponding BOLD contrast losses were well recovered in the combined data (C_C). Furthermore, the effectiveness can be appreciated in the local susceptibility areas in both the t-score and BOLD signal percentage maps.

Figure 4.

BOLD contrast improvement with the proposed reverse gradient approach in temporal (A) and frontal areas (B) of human brain from a single subject. In 1^st row, distortion-corrected forward (R_C, left column), reverse (F_C, middle column), and combined EPIs (C_C, right column) are shown. The corresponding t-score and signal percentage maps from the breath-holding fMRI are shown in 2^nd and 3^rd row, respectively. Note that 2.0 mm isotropic resolution data on the WB3T scanner were chosen.

Similar to the individual result shown in Fig. 4, a signal dropout pattern was also observed in the group averaged coverage-ratio map (Fig. 5A). Either partial or zero overlap of the signal dropout pattern was observed, particularly in frontal and temporal areas due to the similarity of the signal dropout pattern across the subjects. In the frontal lobes, including orbital-medial olfactory and subcallosal areas, the signal loss in the combined data was greatly reduced with the forward EPI (Fig. 5B). In temporal lobes, more apparent signal loss was observed in the anterior and the posterior parts, respectively in the forward and reverse EPI. However, the signal loss was noticeably reduced in the combined data. Consequently, both the global group and the regional coverage ratio were improved in the combined data, as demonstrated by the improved mean coverage ratio with reduced standard deviation (Fig. 5B). With the proposed scheme, the signal loss was minimized in not only brain cortices, but also other local areas such as cerebellum, pallidum, accumbens, ventral diencephalon, and brain stem areas.

Figure 5.

Comparison of cortical coverage ratio between the distortion-corrected forward (F_C), the reverse (R_C) and the combined EPI images (C_C). (A) The cortical coverage for nine subjects is visualized on the inferior view of 3D inflated surface model. Anterior-to-posterior (A/P, upper row) and inferior-to-superior (I/S, bottom row) views are shown. Regions with apparent coverage ratio differences are indicated by arrows and dashed-rectangular boxes in (A). (B) The coverage ratios are present as the mean and standard deviation at local areas of both hemispheres, except for the entire brain stem area, the values of which show in the left hemisphere plot. Note that only the regions (21 out of 74 locations in total) where the coverage ratio is smaller than 0.95 (i.e., 95%) for F_C, R_C or C_C volumes, are plotted in (B). Structural abbreviations (Destrieux et al., 2010) in (B) are: ctx: cortex, G: gyrus (or gyri), S: sulcus (or sulci), inf: inferior, ant: anterior, med: medial, lat: lateral, collat: collateral, transv: transverse, oc-temp: occipito-temporoal, frontomargin: fronto-marginal, frontopol: frontopolar, olfact: olfactory, and Parahip: parahippocampal, fusifor: fusiform, VentralDC: ventral diencephalon.

Figure 6A shows comparisons of group functional contrast maps among the F_C, R_C and C_C images. In group level analysis, it was difficult to obtain robust t-score maps from either the forward or the reverse EPI data alone, especially in the high-susceptibility regions where the signal coverage was missed or the coverage variations across the subjects were high. Since the coverage ratio of the C_C data was improved in each individual and its variation across the subjects was reduced, the GLM-based fMRI t-score map of the C_C data was visually more robust (Fig. 6A-i). In addition, the overall activation pattern of the t-score maps was very similar to the signal change percentage map that is relatively independent from the image SNR (Fig. 6A-ii).

Figure 6.

Group functional contrast maps (A) of F_C, R_C and C_C and the statistical differences (B) on a 3D inflated surface model. In (A), group average maps of GLM t-score (i) and signal percentage changes (ii) are shown for the comparison. Arrows and dashed boxes indicate the activation mismatch between F_C and R_C and the well-preserved activation in C_C. In (B), corresponding statistical difference maps are shown.

Figure 6B presents statistically significant improvements of BOLD contrasts in frontal (dashed rectangular boxes) and anterior temporal areas (red arrows) in the combined C_C data, compared to either F_C or R_C data alone. Rather than using GLM maps (Fig. 6B-i) to show the improvements due to both the signal average (i.e., improved SNR) and redistribution in the combined data, the percentage change maps showed dominant improvements only from signal redistribution (Fig. 6B-ii). When only F_C and R_C were compared, statistically significant differences (p-value<0.05) in frontal and anterior temporal areas were clearly observed in both the GLM and the percentage change maps. In posterior temporal areas (yellow arrow), however, there were no obvious statistically significant differences between F_C and R_C due to partial loss of signal, which cannot be addressed by geometric distortion correction alone. Nevertheless, the BOLD contrasts dominantly observed in F_C were well preserved in C_C.

3.3. Reverse PF acquisition for RG-fMRI

Figure 7 demonstrates the amount of signal intensity loss with the standard and reverse PF acquisition and its reduction in the combined EPI image. With standard PF omitting the early echoes, the signal loss was observed primarily in the stretched areas in both F_C and R_C, which compromised the goal of minimizing the signal loss in the combined EPI, C_C (Figs. 7A and 7B). With the reverse PF omitting the late echoes, however, dominant signal loss was observed in the compressed areas, where the spatial information was either partially or entirely lost. When a very high partial PF of 5/8 was used, the mean signal loss in the combined EPI was about 10.0% vs. 46.4% for reverse vs. the standard PF (Fig. 7C). While the TR remained the same for the standard PF method as TE is fixed at 30 ms, the TR for the reverse PF method was supposed to reduce from 934 ms to 861 ms (7.9%), 860 ms (15.7%), and 788 ms (23.6%), with PF factors of 7/8, 6/8, and 5/8, respectively for the 2.0 mm resolution fMRI on the compact 3T.

Figure 7.

Signal intensity loss with the standard and reverse PF acquisition and its minimization in the RG approach. For a clear demonstration, a represented axial slice across the temporal lobe is shown. While all three data including F_C, R_C, and C_C with various PF factors are shown in (A), image subtraction with the corresponding full acquisition data (i.e., 8/8) is presented to demonstrate the signal losses due to PF acquisition in (B). In (C), the signal loss was quantitatively calculated from nine subjects in a brain mask with high susceptibility effects. Note that only the areas, where the signal difference between F_C and R_C without partial acquisition was higher than 40%, were included in the brain mask (i.e., high susceptibility areas).

Figure 8 demonstrates BOLD contrast changes due to signal loss and image blurring associated with the PF acquisition. While the local signal was entirely lost with a high PF factor, not surprisingly, the corresponding BOLD contrast was also lost, as indicated by arrows on the zoomed image with a 5/8 PF factor. These effects appeared mainly in the stretched (white arrows) and compressed areas (yellow arrows, reverse PF), respectively with the standard and reverse PF acquisition. However, while the combined EPI with the standard PF only partially recovered the signal loss, the signal was fully restored in those with the reverse PF. Therefore, there was no noticeable difference in the BOLD contrast map even with a 5/8 reverse PF factor compared to that without PF.

Figure 8.

Effects of BOLD contrast with PF acquisition: F_C, R_C and the C_C data. As contoured in a yellow-colored dashed box in the full FOV images (A), an enlarged section of the temporal lobe is shown to access the changes of the signal intensity (B) and corresponding function contrast (C). Two color contours demonstrate the boundary of the images without (red) and with (yellow) different PF factors. Arrows indicate the areas, in which the signal intensity (B) and corresponding functional contrast (C) vary with the PF factors noticeably.

5. Discussion

In this study, practical application of RG GE-EPI is evaluated in whole-brain human fMRI. Multiband imaging is adopted so that the temporal and spatial resolution of the interleaved RG scheme can be improved for whole-brain human fMRI studies on both the conventional WB3T and the high-performance C3T scanners. With in-vivo breath-holding RG-fMRI, it was demonstrated that the proposed RG method improves image quality and thus maximizes BOLD signal detection at both the individual and group level. In addition, when PF is need for high spatial resolution, reverse PF instead of standard PF can be used to minimize loss in resolution and signal. The proposed method was based on understanding the relationship between the echo shift and geometric distortion and reliably measuring both effects using the PSF mapping.

Multi-band imaging (Setsompop et al., 2012) was used to achieve a reasonable effective temporal resolution for whole-brain RG-fMRI study and to shorten the calibration time of PSF mapping scan. Rather than using in-plane acceleration, which penalizes the image SNR substantially due to undersampling, a relatively high through-plane acceleration of 6 was applied in this study since it can minimize the TR and the SNR loss given fixed total acceleration factor (Smith et al., 2013). Without in-plane acceleration, however, noticeable in-plane distortions were observed for high-resolution imaging on both the conventional WB3T and the high-performance C3T scanners. Nevertheless, both the geometric and signal distortions were effectively minimized with the PSF-based distortion correction scheme. Furthermore, multi-band factor of 6 shortened the PSF scan.

This study demonstrated that the echo shifting leading to EPI signal loss which can be either further exacerbated or reduced depending on the type of geometric distortion or the phase-encoding polarity. When the regional signal is compressed in EPIs, the local field gradients were concentrated within the nominal imaging grid (or voxel), which leads to shifting of the echo peak out of the k-space acquisition window and resulted in complete signal dropout in the affected area, as shown in the effective echo-shift map (Fig. 2B). In distinction, in regions with signal stretching, the local field gradients were distributed over the stretched voxels. Therefore, the echo shifting in the stretched area is less than in the compressed area and still can be captured within the EPI acquisition window (Fig. 2C). This is the reason why the RG scheme can minimize signal loss in the combined data. In contrast to the EPI-based calculation (Deichmann et al., 2002), the PSF mapping method was able to reliably assess the echo-shift effect even in compressed areas. Since the patterns of signal dropouts between the distortion-corrected EPI pair were well-matched with the effective echo-shift map in the non-distorted space, a direct estimate of the signal dropout in EPI is possible with PSF mapping, even without any signal dropout simulation (De Panfilis and Schwarzbauer, 2005) or distortion correction. Consequently, the proposed effective echo-shift map in the distorted space can be evaluated to guide imaging parameter optimization for minimizing the signal loss in a target ROI for fMRI, especially when the RG scheme is not available.

The advantage of the RG scheme for fMRI can be maximized by increasing the EPI acquisition window, i.e., increasing the k-space coverage or using higher spatial resolution, as indicated by the improved Dice coefficient and mean coverage ratios in high-resolution imaging without PF (Fig. 1). In terms of minimizing the signal loss on the WB3T scanner, 2.0- and 2.2-mm isotropic resolution whole-brain imaging would benefit from full acquisition without PF, but TEs need to be shifted to 42.7 and 36.5 ms respectively, which were still acceptable for fMRI at 3T. When seeking to increase both the EPI acquisition window and fMRI temporal resolution, the use of the reverse PF acquisition becomes necessary on the WB3T. For the RG distortion correction, spatial information is more reliable in the stretched regions rather than the compressed regions due to loss of spatial information (In et al., 2017). As demonstrated in this study, the local effective TE in the stretched regions is always shorter than the desired TE. As reverse PF preserves the signal from stretched region better than the compressed region, in this regard, it results in less signal loss and suits the RG method better.

It’s important to note that breath-holding as a physiological stimulus could also induce local field gradient changes leading to echo shift variations. However, previous studies (Arja et al., 2010; Hagberg and Tuzzi, 2014) showed that the percentage signal change in the phase data is low, compared to that of the magnitude, just like other unwanted phase effects induced by subject motions and hardware imperfections. In addition, its spatial deviation, which leads echo shift would be even smaller. Therefore, the echo shift variation induced by this effect could be considered negligible in this study.

Although the effectiveness of the RG scheme was demonstrated mainly in brain cortices by performing breath-holding fMRI, this approach would also be beneficial for fMRI in other brain regions such as pallidum (Gibson et al., 2017; Hancu et al., 2019), accumbens (Gibson et al., 2017; Hancu et al., 2019), cerebellum (Marvel et al., 2012), and brain stem areas (Beissner, 2015), which also suffered from susceptibility artifacts and were not well investigated yet. As shown in this study, the measured coverage ratio difference related to the phase-encoding polarity can be useful guidance to minimize the signal loss by choosing the appropriate phase-encoding polarity in the specific target brain areas. For example, while more observable signal loss in cerebellum and brain stem areas was shown in forward EPI, reverse EPI suffered from higher signal loss in pallidum and accumbens areas. To cover the entire brain volume, the proposed RG scheme would still be the optimum option in investigating functional mechanism.

A simple zero-filling PF reconstruction is used in this simulation, which can cause image blurring in the reconstructed image. Our major interest in this study, however, is to minimize the spatial resolution loss with the RG scheme in the high susceptibility areas, where the echo signal is strongly deviated from the k-space center. In more advanced PF methods (Chen et al., 2008; Koopmans and Pfaffenrot, 2021), the required image background phase information is estimated from the small portion of the central k-space and the low-resolution phase data still may not provide accurate phase estimation for such high susceptibility areas, which often resulted in additional artifacts. To avoid these potential confounding factors, this investigation was conducted with the zero filling PF reconstruction. This can be considered a limitation of the study, and the use of an improved PF reconstruction method (Chen et al., 2008; Koopmans and Pfaffenrot, 2021) could be beneficial to minimize any PF related blurring in the areas with no or lower susceptibility effect.

The RG approach minimizes the signal loss induced by in-plane spin dephasing, but not through-plane spin dephasing in the slice direction (Deichmann et al., 2002; In et al., 2017). To minimize through-plane signal loss, EPI acquisitions with a compensation gradient in slice direction (Deichmann et al., 2002) and optimum slice angle selection (Deichmann et al., 2003; Weiskopf et al., 2006) were applied previously. Even in the areas with strongest  through-plane signal loss such as rectus, however, at least greater than 60% were still recovered with the proposed RG scheme in the group coverage map from nine subjects (Fig. 5). It is expected that the proposed RG approach combined with the correction methods for through-plane signal loss could further minimize the signal loss in a target ROI.

6. Conclusion

In summary, a PSF mapping-based reverse gradient approach was applied for human brain fMRI to minimize the loss of the susceptibility-induced functional contrast, especially in the frontal and temporal lobes. This study demonstrates that the proposed method can improve the fMRI reliability in both individual and group data analysis. In conjunction with the multi-band imaging and reverse partial Fourier acquisition, the temporal resolution of the RG approach and the calibration scan for distortion correction are practical for whole-brain human fMRI. PSF mapping method allowing local echo shift measurements can also be useful as a tool for optimizing the imaging protocol. With the improvements, the proposed technique is particularly useful for investigating the functional mechanisms of the brain in the temporal and frontal areas which have not been intensively explored yet due to conventional GE-EPI limitations.