Block‐interleaved segmented echo‐planar imaging for improved activity detection in submillimeter high‐resolution functional MRI at 7 T

Abstract

Purpose

Isotropic submillimeter high‐resolution functional MRI (fMRI) facilitates noninvasive investigation of neuronal activities at the mesoscale level, including cortical columns and laminae. However, the task‐evoked functional activity becomes less detectable when using single‐shot echo planar imaging (EPI) with parallel imaging techniques such as generalized autocalibrating partially parallel acquisitions (GRAPPA), owing to a reduction in temporal signal‐to‐noise ratio (tSNR). Although conventional multishot EPI (msEPI) enhances tSNR, it reduces temporal resolution, making it less suitable for fMRI studies involving short‐duration stimuli. To overcome this problem, a novel msEPI‐based fMRI acquisition and reconstruction method, called block‐interleaved segmented EPI (BISEPI), was proposed.

Methods

This technique uses timing information of a block design paradigm during acquisition and reconstruction to preserve both temporal resolution and tSNR. Furthermore, a k‐space‐based motion correction method is incorporated to mitigate head motion artifacts.

Results

Results from human studies showed that the proposed method reduced g‐factor penalties typically observed in GRAPPA‐accelerated EPI. It provided high tSNR, improved motion robustness, and enhanced sensitivity to blood oxygen level–dependent signal responses.

Conclusion

The proposed method circumvents both conventional msEPI‐related and GRAPPA‐related problems and enables the detection of blood oxygen level–dependent signal responses at submillimeter spatial resolution. It is effective for paradigms using short‐duration, low‐power stimuli at 0.7‐mm isotropic resolution, and standard‐power stimuli at 0.4‐mm isotropic resolution.

1. INTRODUCTION

Blood oxygenation level–dependent (BOLD) functional MRI (fMRI) ¹ , ² remains the most widely used modality for noninvasive investigation of brain function during tasks or resting states. To capture BOLD responses of certain brain regions at the mesoscale, imaging of functional microstructures must achieve submillimeter resolution with high functional contrast. Ultrahigh‐field (7 T or higher) fMRI, combined with parallel imaging techniques, has enabled challenging investigations targeting cortical functional units, such as columns and depth‐resolved laminar responses. ³ , ⁴ However, both the signal‐to‐noise ratio (SNR) and spatial specificity of fMRI signals remain critical for detecting small neuronal activity at columnar or laminar resolutions. Compared with the widely used gradient‐echo BOLD fMRI, spin‐echo BOLD and vascular‐space‐occupancy (VASO) fMRI offer higher localization specificity near neuronal activity in gray matter but suffer from reduced SNR and sensitivity, limiting their functional detectability. ⁴ , ⁵ , ⁶

Ultrahigh‐field MRI provides substantial SNR gains and increased BOLD contrast relative to lower field (≤ 4 T) systems. ⁷ , ⁸ When paired with array coils, this SNR improvement expands the potential for designing pulse sequences that extract fine‐scale anatomical and physiological information from the brain tissue, ⁹ , ¹⁰ , ¹¹ enabling submillimeter‐resolution fMRI with isotropic voxels. ¹² Echo‐planar imaging (EPI), typically implemented as a single‐shot technique (ssEPI), is the fastest acquisition method ¹³ and is used widely in fMRI owing to its strong sensitivity to BOLD signal. However, acquisition with large matrix sizes, issues such as rapid signal decay during echo trains, gradient system limitations, and extended readout durations arise due to long acquisition windows, ultimately restricting its practical application to high‐resolution studies. Although partial Fourier and parallel imaging approaches such as generalized autocalibrating partially parallel acquisitions (GRAPPA) ¹⁴ and sensitivity encoding (SENSE) ¹⁵ shorten readout time, they reduce SNR and introduce artifacts related to k‐space undersampling, which can compromise detectability and spatial specificity of BOLD signal. Additionally, ssEPI is prone to geometric distortion due to prolonged readout time. The SNR of GRAPPA‐reconstructed images is further affected by static magnetic field (B₀) inhomogeneity, which indirectly reduces SNR through mechanisms such as T₂ signal loss, phase errors, and g‐factor noise amplification. Sensitivity encoding is even more susceptible to phase errors from B₀ inhomogeneity, leading to greater noise amplification in affected regions. ¹⁶ , ¹⁷ , ¹⁸ In both methods, the SNR is significantly reduced when a high acceleration factor is applied due to g‐factors substantially exceeding 1. ¹⁹ , ²⁰ Therefore, the use of high acceleration factor in ssEPI is generally discouraged for fMRI applications. ²¹ , ²² , ²³

Multishot EPI (msEPI) sequence offers shorter readout times per shot (or segment) than ssEPI and is therefore less susceptible to phase errors during k‐space filling, which accumulate and result in distortions and signal loss in the reconstructed images. In msEPI, images are reconstructed from separately acquired k‐space segments, allowing a reduced readout time while improving spatial resolution. However, increasing the number of segments to achieve high spatial resolution comes at the cost of reduced temporal resolution. Thus, the number of segments cannot be increased indefinitely, as excessive segmentation may obscure the BOLD signal response, particularly when using short‐duration stimuli. ²⁴ , ²⁵ Furthermore, longer acquisition durations, along with factors such as Nyquist ghosting, physiological noise, and sensitivity to system instabilities, pose significant challenges in msEPI.

In this study, we introduce a novel msEPI method called block‐interleaved segmented EPI (BISEPI) for fMRI. This approach realigns k‐space trajectories using stimulus timing information derived from a block design paradigm during both acquisition and reconstruction, thereby preserving temporal resolution and temporal SNR (tSNR). Head motion artifacts arising from long acquisition durations in BISEPI are corrected in k‐space based on motion estimates derived from GRAPPA‐reconstructed images by treating BISEPI per‐segment data as ssEPI data. We applied BISEPI to perform submillimeter high‐resolution fMRI at isotropic spatial resolutions of 0.8, 0.7, 0.6, 0.5, and 0.4 mm, combined with temporal resolutions of 1.5 and 1.0 s across multiple blocks/cycles, in healthy participants. To reconstruct activation maps and their corresponding BOLD response curves, we explored various combinations of GRAPPA acceleration factors and number of segments. This study demonstrates the feasibility of using the proposed BISEPI technique to detect brain activation elicited by both standard and short‐duration, low‐power visual stimuli in submillimeter high‐resolution fMRI at 7 T.

2. METHODS

2.1. BISEPI

In msEPI, temporal resampling strategies are essential in fMRI to balance temporal resolution, spatial resolution, tSNR, and image artifacts. These strategies are particularly critical for accelerating acquisition, improving motion correction, and optimizing BOLD signal detection. The BISEPI pulse sequence was designed with reference to the k‐space filling pattern used in DIANA. ²⁶ Instead of consecutively acquiring all k‐space segments for a single volume as in conventional msEPI (Figure 1A), BISEPI repeats acquisition of each k‐space segment during a single fMRI stimulation unit, which is a minimum set of stimuli being repeated (Figure 1B). By repeating this approach across segments, we can obtain a “functional unit” of k‐space dataset, a minimum analyzable set of k‐space time series in which all segments in all volumes are filled. This functional unit is, in most cases, further repeated within the run to achieve sufficient statistical power in fMRI analysis. In EPI‐based time‐series acquisition, repetition time (TR) can be defined in several ways, depending on which part of the acquisition cycle is considered. In conventional msEPI, a volume TR (TRvol), which is the time required to complete a single volume acquisition, can be defined as the time interval between successive volume excitations. Similarly, a segment TR (TRseg), which is the time needed to complete a single‐segment acquisition, can be the time interval between successive excitations for successive acquisition of segments. TRvol is thus given by TRseg × Nseg, where Nseg is the number of segments. In BISEPI, TRseg can be similarly defined as the time interval between successive excitations for consecutive two segments, regardless of whether they belong to the same volume. However, TRvol in BISEPI has a completely different definitions that is in close association with the block design paradigm of the stimulus. When the acquisition of each segment has to be repeated Nrep times to comprise a single functional unit, the TRvol is given by TRseg × Nrep × Nseg. In addition, it may be helpful to define TRstim as the time required to present a minimum set of stimuli that is repeated Nseg times to obtain a single functional unit dataset. It should be noted that, although TRvol in BISEPI is extremely longer than in conventional msEPI, effective temporal resolution in BISEPI is TRseg, and therefore shorter than in conventional msEPI. This strategy allows us to avoid the loss of temporal resolution that inevitably occurs in conventional msEPI, enabling submillimeter high‐resolution fMRI studies. BISEPI can be implemented using either a two‐dimensional or three‐dimensional (3D) acquisition scheme. In this study, we used a 3D acquisition to improve SNR. In the 3D implementation, k‐space sampling along the slice‐encoding direction is performed during the acquisition of each segment. The acquired signals were realigned in k‐space time series according to the corresponding segments and volumes for reconstruction. BISEPI with its high SNR and increased spatiotemporal resolution is suitable for fMRI studies, including those involving short‐duration stimuli used in earlier studies. ¹⁹ , ²¹ BISEPI incorporates three navigator lines acquired before each segment to estimate shot‐to‐shot phase variations and frequency drift for phase correction in the reconstruction stage. A fat saturation pulse is applied before each excitation pulse, and a crusher gradient follows each readout train.

FIGURE 1.

Comparison of conventional multishot echo planar imaging (EPI) versus block‐interleaved segmented EPI (BISEPI) in functional MRI (fMRI). TR, repetition time.

2.2. GRAPPA and shot‐wise k‐space motion correction

Offline GRAPPA reconstruction was implemented following the approach described in https://github.com/marcsous/parallel, using a kernel size of 3 × 4 to match the Siemens product GRAPPA reconstruction. Segmented (Experiments 1–4 in Table 1; experiment numbers hereafter refer to those in Table 1) and FLEET (Experiment 5) GRAPPA reference schemes were used, with 36 autocalibration signal lines. Because BISEPI is sensitive to both physical and physiological motion due to the extended TRvol, we perform k‐space motion correction (k‐MoCo) to mitigate motion artifacts. For k‐MoCo, a conventional msEPI single volume is acquired for reference images before the BISEPI acquisition (Figure 2). Each BISEPI segment is then reconstructed using the GRAPPA scheme, treating BISEPI data as ssEPI data. Motion parameters are extracted using Statistical Parametric Mapping (SPM12; Wellcome Centre for Human Neuroimaging, University College London) software by comparing these ssEPI images reconstructed from BISEPI data with the reference images. These parameters are used to reconstruct images from the ssEPI k‐space dataset via the iterative nonuniform Fourier transform implemented in the MIRT toolbox ²⁷ to correct motion artifacts. Finally, obtained images are transformed back to k‐space using the fast Fourier transform before final image reconstruction, as illustrated in the procedure shown in Figure 2.

TABLE 1.

Imaging parameters used in each experiment.

Exp	Voxel (mm)	R	Nseg	Nrep	TRseg/TRvol (s)	TE (ms)	FOV (mm)	Matrix	Slices	FA (deg)	BW (Hz/px)	Offline postprocess
1	0.7	1	3	60	1.5/90	26	87	64 × 128	16	25	1002	M
2	0.6	1	3	20	1.5/90	26	87	144 × 144	16	18	846	NA
	0.6	3	1	NA	NA/1.5	26	84	144 × 144	16	18	846	NA
3	0.4	1	4	20	1.5/120	31	77	192 × 192	16	20	636	S, M + N
4	0.4	1	4	30	1.0/120	26	64	160 × 160	12	23	694	NA
	0.4	2	2	15	1.0/60	26	64	160 × 160	12	23	694	NA
	0.4	4	1	1	NA/1.0	26	64	160 × 160	12	23	694	NA
5	0.5	1	3	NA	1.0/90	26.3	72	144 × 144	12	20	790	M + N
	0.5	3	1	30	NA/1.0	26.3	72	144 × 144	12	20	790	M + N
	0.6	1	3	NA	1.0/90	22	78	128 × 128	12	20	952	M + N
	0.6	3	1	30	NA/1.0	22	78	128 × 128	12	20	952	M + N
	0.8	1	3	NA	1.0/90	21	102	128 × 128	12	20	952	M + N
	0.8	3	1	30	NA/1.0	21	102	128 × 128	12	20	952	M + N
	0.7	1	3	NA	1.0/90	24	89	128 × 128	12	20	952	S, M, M + N
	0.7	3	1	30	NA/1.0	24	89	128 × 128	12	20	952	S, M, M + N

Note: Offline postprocess: S = Standard, M = k‐MoCo, N = NORDIC; NA = Not applicable.

Abbreviations: BW, bandwidth; Exp, experiment number; FA, flip angle; FOV, field of view; Matrix, matrix size; NA, not applicable; Nseg, number of segments; Nrep, number of repetitions; R, GRAPPA acceleration factor; Slices, number of slices (partitions); TE, echo time; TRseg, segment TR; TRvol, volume TR; Voxel, isotropic voxel size.

FIGURE 2.

Diagram of motion correction and reconstruction. BISEPI, block‐interleaved segmented echo planar imaging; GRAPPA, generalized autocalibrating partially parallel acquisitions; k‐MoCo, k‐space motion correction; msEPI, multishot echo planar imaging; ssEPI, single‐shot echo planar imaging.

2.3. fMRI experiments

2.3.1. fMRI data acquisition and stimuli

The fMRI data were acquired using a MAGNETOM 7T scanner (Siemens, Erlangen, Germany; Terra for the last two experiments, Experiments 4 and 5) with a 32‐channel phased‐array head coil (Nova Medical, Wilmington, MA, USA). We used a cosine‐modulated variable flip‐angle scheme to reduce blurring and improve tSNR: FA(p) = FA_norm × (1 – 0.45 × cos[2πp/N]), where FA(p) is the flip angle at partition p, FA_norm is the nominal flip angle, and N is the total partitions. The cosine modulation can maximize the signal at k‐space center (FA[p] ≈ 1.45 × FA_norm if p ≈ N/2) while minimizing saturation at outer k‐space (FA[p] ≈ 0.55 × FA_norm, if p ≈ 0 or N − 1). A bite‐bar stabilization system was used during scanning to minimize subject motion. Additionally, an in‐bore display was positioned 50 cm from the subject's eyes to present visual stimuli. Scanning was conducted using our in‐house 3D BISEPI sequence, as described previously, at an oblique coronal slab orientation (Figure 3A; head‐to‐foot readout direction, right‐to‐left phase‐encoding direction, anterior‐to‐posterior slice‐encoding direction). The acquisition parameters are listed in Table 1. A phase partial Fourier factor of 6/8 was applied to all scans. Different subjects were scanned in each experiment, with a total of 12 subjects scanned overall. Slices were oriented in the coronal plane and covered the visual cortex, including the striate cortex (V1). Across experiments, various stimuli were used to evaluate the performance of BISEPI‐based fMRI. The stimuli were commonly used visual paradigms, including weak or standard ring stimuli, drifting gratings, and flickering checkerboards. The findings are expected to demonstrate BISEPI's potential in high‐resolution fMRI studies involving these visual paradigms. Two types of GRAPPA reference schemes were used (segmented reference in Experiments 1–4 and FLEET reference in Experiment 5) for comparison.

FIGURE 3.

Visual stimuli. (A) Slice position. (B,C) Visual stimuli. (D,E) Interstimulus interval (ISI) and non‐ISI stimuli.

2.3.1.1. Ring stimuli at 0.7‐mm isotropic resolution (Experiment 1)

Subjects were instructed to fixate on a point presented at the center of the field of view while undergoing an fMRI experiment using a block design with a 6‐s ON/24‐s OFF visual stimuli (Figure 3B). Two types of visual stimulus patterns were used to assess differences in the BOLD signal response (Figure 3D,E): Pattern 1 (Figure 3D) consisted of a small white ring presented continuously for 6 s, and Pattern 2 (Figure 3E) began with a small ring shown for 0.05 s, followed by a randomly determined blank interval (interstimulus interval [ISI]) of either 0 or 0.05 s, then a large ring for 0.05 s, and finally a blank interval adjusted to ensure the total duration from onset to the end of the final blank reached 1.5 s. This 1.5‐s visual stimulus set was repeated 4 times, forming a 6‐s block containing small and large rings interleaved with blank. Compared with Pattern 1, which delivers a continuous 6‐s stimulus without blanks (non‐ISI stimuli), Pattern 2 includes only 0.4 s of visual stimulus within the 6‐s ON period (ISI stimuli). This introduces a key distinction: Pattern 2 (ISI stimuli) is weaker (lower power) than Pattern 1 (non‐ISI stimuli). A similar block design was used in an fMRI study on visual object recognition processing. ¹⁷ Three different visual stimuli (one Pattern 1 and two Pattern 2) were randomly presented to subjects. Each block set was repeated 3 times to align the data with BISEPI using three segments, resulting in nine ON/OFF blocks (3 stimuli ON/OFF blocks × 3 segments) that composed a single functional unit. The acquisition time for one functional unit (i.e., TRvol) is calculated as TRseg × Nrep × Nseg = 270 s, where TRseg = 1.5 s, Nrep = 60, and Nseg = 3. Five functional units were repeated per run, yielding a total of 45 ON/OFF blocks in 1350 s (22.5 min). Each subject completed multiple runs (two runs for Subject 1 and three runs for Subject 2) on different days using the same protocol, with randomized stimulus presentation during each run.

2.3.1.2. Flickering checkerboard/drifting gratings stimuli at 0.6‐ and 0.4‐mm isotropic resolution (Experiments 2–4)

Subjects were scanned using 18‐s/12‐s flickering checkerboard (Figure 3C) or drifting grating ON/OFF visual stimuli. For the 0.6‐mm isotropic resolution acquisitions (Experiment 2), the total scan duration was 12 min (flickering checkerboard ON/OFF repeated 24 times) for one run. One BISEPI run and one independently acquired ssEPI with GRAPPA (R = 3) (hereafter referred to as ssEPI‐R3) run were performed to compare imaging performance and detectability of BOLD signal response. The acquisition parameters, except for BISEPI‐specific ones, were matched between the two runs, with the TRvol of ssEPI‐R3 set to be equal to the TRseg of BISEPI. The same acquisition parameter matching strategy was applied in all subsequent comparisons between BISEPI and ssEPI‐R3. For the same acquisition time of 12 min, each BISEPI run yielded realigned data consisting of 160 volumes, whereas the ssEPI‐R3 run with a segmented GRAPPA reference scheme acquired 480 volumes. The 0.4‐mm isotropic resolution acquisitions were performed twice (Experiment 3 and 4), each with different condition. For the first acquisition (Experiment 3), one BISEPI run was performed over 8 min, during which the drifting gratings ON/OFF block was repeated 16 times. The second acquisition (Experiment 4), different combinations of GRAPPA acceleration factors (R) and segment numbers (S) (hereafter referred to as BISEPI‐R‐S format), BISEPI‐R4‐S1, BISEPI‐R2‐S2, and BISEPI‐R1‐S4, were used under the same acquisition time (6 min), and the flickering checkerboard ON/OFF block was repeated 12 times for comparison.

2.3.1.3. Flickering checkerboard stimuli at 0.5‐, 0.6‐, 0.7‐, and 0.8‐mm isotropic resolution (Experiment 5)

To identify the source of BISEPI's advantages over ssEPI with GRAPPA (hereafter referred to as ssEPI‐GRAPPA), we conducted the flickering checkerboard–stimulus fMRI experiments (Experiment 5) on multiple subjects using BISEPI with Nseg = 3 (hereafter referred to as BISEPI‐S3) and ssEPI‐R3 at 0.5‐, 0.6‐, 0.7‐, and 0.8‐mm isotropic resolutions. In all runs, a 15‐s ON/15‐s OFF block was repeated 12 times, resulting in a total scan duration of 6 min. The number of volumes acquired was 360 for ssEPI‐R3 and 120 for BISEPI‐S3, forming the same scan duration. The FLEET GRAPPA reference scheme was used to reduce the effects of head motion during GRAPPA reconstruction.

2.3.2. MRI data preprocessing and analysis

Data processing and analysis were performed using MATLAB 2024a (MathWorks, MA, USA), SPM‐12, and BrainVoyager QX (Brain Innovation, Maastricht, Netherlands). The raw data from Experiments 1, 3, and 5 were transferred from the MRI system to a computer for offline preprocessing using in‐house software. The images from Experiments 2 and 4 were reconstructed using the product reconstruction system provided with the MRI system, and no k‐MoCo or noise reduction with distribution correction (NORDIC) (see below) processing was applied to allow for a direct comparison of inherent sequence performance. The data from each run were realigned but not normalized to a standard space. The preprocessing pipeline was designed to preserve the data in its native space.

The data from Experiments 1, 3, and 5 were phase‐corrected using navigator lines acquired before each readout train and reconstructed followed by k‐MoCo. The reconstructed complex image time series for each coil channel in Experiments 3 and 5 were denoised using NORDIC denoising (https://github.com/SteenMoeller/NORDIC_Raw). NORDIC is an advanced denoising method for MRI, particularly effective for high‐resolution functional imaging. It reduces thermal noise while preserving spatial and temporal structure, enhancing image quality and SNR without compromising data fidelity. Based on principal component analysis, it leverages the statistical characteristics of noise in k‐space. Following k‐MoCo and NORDIC denoising, the data were converted into standard NIfTI format and temporally high‐pass‐filtered using a general linear model Fourier basis set of two sine/cosines, along with temporal Gaussian smoothing using an full‐width at half‐maximum kernel of two data points. No additional spatial smoothing was applied.

For BOLD response analysis, general linear model analysis was performed by modeling the ON/OFF stimulation as a regressor convolved with a canonical hemodynamic response function. Activated pixels were identified using BrainVoyager based on the presence of a BOLD response, independent of stimulus type.

2.4. Ethics approval

This study was approved by the institutional review board. Written informed consent was obtained from all participants.

3. RESULTS

3.1. Ring stimuli at 0.7‐mm isotropic resolution (Experiment 1)

We focused on two regions of interest (ROIs), striate and extra‐striate cortex, to evaluate the detection of BOLD signal responses to the aforementioned visual stimuli. During the experiment, three types of visual stimuli were presented in randomized order. In the analysis, we assessed whether any of them elicited BOLD responses. We consistently observed a higher response amplitude for non‐ISI stimuli than for ISI stimuli in the striate cortex; however, no significant difference was observed in the extra‐striate cortex, regardless of imaging run, subject, or stimulus order during the experiments. These results are presented in Figure 4 for Subjects 1 (Figure 4A) and 2 (Figure 4B), where the activity map is displayed in the left column, while the stimulus‐specific BOLD signal response is shown in the right column. The BOLD signal responses within ROIs are consistently higher in amplitude for the non‐ISI stimuli than for stimuli with ISIs of 0 ms and 50 ms in ROI 1 (striate cortex, green rectangle), as illustrated in the leftmost graphs of the right column. The responses in ROI 2 (extra‐striate cortex, orange rectangle) differ slightly in amplitude, as indicated in the rightmost graphs. We do not interpret the response further. However, the detected activities in our BISEPI experiments at 0.7‐mm isotropic resolution, including low‐power stimuli, were highly reproducible across runs, days, and subjects.

FIGURE 4.

Comparison of interstimulus interval (ISI) versus non‐ISI responses and their reproducibility at 0.7‐mm isotropic resolution. (A) Subject 1. (B) Subject 2. Left column, activation maps. Center and right columns, blood oxygen level–dependent signal responses within Regions of Interest (ROIs) 1 and 2, respectively.

3.2. Flickering checkerboard stimuli at 0.6‐mm isotropic resolution (Experiment 2)

Figure 5 presents the results obtained from data collected in two subjects (Subjects 3 and 4). In each panel, the top row displays BOLD activation maps for ssEPI‐R3 (left column) and BISEPI‐S3 (middle column). The estimated BOLD signal responses from the activated regions in these slices are shown in the right column. The bottom row of Figure 5 shows the averaged BOLD signal time series from these activated regions alongside the motion estimates. BISEPI‐S3 can detect BOLD responses in both superficial and deep cortical layers, whereas ssEPI‐R3 can detect BOLD responses at the pial surface. Although the BOLD responses at the pial surface were strong even with ssEPI, these signals are likely attributed to blood in veins rather than in tissue compartments. Although the motion estimates with BISEPI‐S3 were larger than those with ssEPI‐R3, they were still within ±1 mm or ±1°, indicating that the bite‐bar stabilization system effectively limited head motion to submillimeter level.

FIGURE 5.

Comparison of acquisition/reconstruction strategies in activation maps, blood oxygen level–dependent (BOLD) signals, and motion estimates at 0.6‐mm isotropic resolution. Top row, activity maps and corresponding averaged BOLD signal responses. Bottom row, BOLD signal time courses and their corresponding motion estimates. BISEPI, block‐interleaved segmented echo planar imaging; ssEPI, single‐shot echo planar imaging.

3.3. Flickering checkerboard/drifting gratings stimuli at 0.4‐mm isotropic resolution (Experiments 3 and 4)

We confirmed the performance improvement of BISEPI with NORDIC at 0.4‐mm isotropic resolution by comparing BOLD signal responses (from drifting gratings stimuli, Experiment 3) obtained from BISEPI with and without NORDIC‐denoised data. Figure 6A shows tSNR maps for one subject's data reconstructed directly (no offline postprocessing; hereafter referred to simply as standard), using k‐MoCo and using k‐MoCo followed by NORDIC denoising (hereafter referred to simply as NORDIC), respectively. These data were not subjected to spatial smoothing or masking. The tSNR was more than four times greater for the data processed with NORDIC compared with the standard or k‐MoCo data. Figure 6B presents activation maps from three subjects (Subjects 5–7; standard reconstruction in the top row, NORDIC in the bottom row). Functional images, rather than structural images like MPRAGE (magnetization‐prepared rapid acquisition with gradient echo) images, were used here to overlay to illustrate noise level differences between standard and NORDIC reconstructions. Although no activation in deep cortical layers was detected with standard reconstruction, highly precise BOLD activation was observed in NORDIC images, including deep layers. These results demonstrate that stimulus‐evoked signal changes at 0.4‐mm isotropic resolution were clearly visible using BISEPI with NORDIC.

FIGURE 6.

The effect of offline postprocessing on temporal signal‐to‐noise ratio (tSNR) and blood oxygen level–dependent (BOLD) activation maps at 0.4‐mm isotropic resolution. (A) tSNR maps from one subject. (B) BOLD activation maps with standard (without any offline postprocessing) and noise reduction with distribution correction (NORDIC) (with k‐space motion correction [k‐MoCo]) reconstruction.

Figure 7 shows results from Experiment 4 (with Subjects 8 and 9) using flickering checkerboard stimuli and three different combinations of GRAPPA acceleration factors (R) and segment numbers (S). No offline postprocessing, such as k‐MoCo or NORDIC, was applied, allowing evaluation of the pulse sequence's inherent BOLD signal detectability. BISEPI‐R2‐S2 and BISEPI‐R1‐S4 detected more activated pixels than BISEPI‐R4S1. The BISEPI technique can be effectively combined with GRAPPA, like the combination of standard msEPI with GRAPPA.

FIGURE 7.

The impact of combinations of acceleration factors and number of segments on block‐interleaved segmented echo planar imaging (BISEPI) at 0.4‐mm isotropic resolution. BISEPI‐R4‐S1, BISEPI‐R2‐S2, and BISEPI‐R1‐S4 indicate combinations of generalized autocalibrating partially parallel acquisitions acceleration factor (R) and number of segments (S) in BISEPI.

3.4. Flickering checkerboard stimuli at 0.5‐, 0.6‐, 0.7‐, and 0.8‐mm isotropic resolution (Experiment 5)

As shown in Figure 8, experimental results from two subjects (Subjects 10 and 11) demonstrated that BISEPI‐S3 detected a greater number of activated pixels than ssEPI‐R3 across three isotropic spatial resolutions (0.5‐, 0.6‐, and 0.8‐mm). Furthermore, ssEPI‐R3 either failed to detect activity or exhibited signal loss in certain regions. Experimental results for the third subject (Subject 12), acquired at an isotropic spatial resolution of 0.7 mm, are presented in Figure 9. Here, the effect of postprocessing methods, k‐MoCo and NORDIC, on BOLD signal detection is displayed. The results indicate that although both k‐MoCo and NORDIC enhanced BOLD detection for both ssEPI‐R3 and BISEPI‐S3 data, they did not substantially change the relative performance between these acquisition methods. In other words, the observed advantages of BISEPI are not attributable to k‐MoCo, NORDIC, or both, but are more likely associated with the scanning and image‐reconstruction strategies. One plausible explanation is that B₀ inhomogeneity deteriorated image quality (especially in ssEPI) though GRAPPA reconstruction process, potentially compromising its performance.

FIGURE 8.

Blood oxygen level–dependent (BOLD) activation maps at isotropic resolution of 0.5, 0.6, and 0.8 mm. BISEPI, block‐interleaved segmented echo planar imaging; ssEPI, single‐shot echo planar imaging.

FIGURE 9.

The effect of offline postprocessing on block‐interleaved segmented echo planar imaging (BISEPI) and single‐shot echo planar imaging (ssEPI) at 0.7‐mm isotropic resolution. k‐MoCo, k‐space motion correction; NORDIC, noise reduction with distribution correction.

4. DISCUSSION

In this study, we demonstrate that BISEPI can enhance spatial resolution and reduce geometric distortions. However, its intershot instability, caused by physical and physiological motion, can be more pronounced compared with ssEPI. Image‐based motion correction may not fully compensate for head motion occurring during BISEPI acquisition, especially in cases of abrupt movements exceeding several voxels. Such sudden motion can introduce artifacts or residual errors that are difficult to correct using standard postprocessing, potentially compromising the accuracy of functional activation maps and BOLD signal detection. This limitation underscores the importance of minimizing subject motion during fMRI acquisition, particularly in high‐resolution studies in which even small displacements can significantly affect data quality. A bite‐bar stabilization system can constrain head motion to below 1 mm and 1°, and intershot instability can be mitigated using the proposed k‐MoCo, with additional improvements achievable through NORDIC denoising. The effectiveness of k‐MoCo depends on the quality of the ssEPI‐GRAPPA images reconstructed from the BISEPI data. Therefore, if the quality of these images is insufficient, BISEPI data acquired across several segments may not be accurately corrected for motion in k‐space. An upper limit on the number of segments can be set for an experiment to ensure acceptable quality of the ssEPI‐GRAPPA reconstructed images for k‐MoCo. Although physiological noise in BISEPI is more difficult to eliminate than in ssEPI‐GRAPPA, BISEPI reduces the g‐factor dependency of the SNR compared with ssEPI‐GRAPPA (with both segmented and FLEET reference schemes). Consequently, BISEPI maintains an acceptable SNR for submillimeter high‐resolution fMRI while leveraging SNR gains to enhance spatial resolution.

In comparison with ssEPI‐GRAPPA, a disadvantage of BISEPI is the reduced number of volumes that can be acquired within the same scan time. Although the number of volumes acquired during the specified acquisition time is reduced in BISEPI, increased tSNR achieved by BISEPI may be more critical in submillimeter high‐resolution fMRI, as demonstrated by our results, which are consistent with previous findings. The effect of fewer BISEPI volumes compared with ssEPI‐GRAPPA can be further mitigated by increasing the total imaging time series.

BISEPI segmentation can be performed along one or both directions of the in‐plane and through‐plane for 3D acquisition. The results of this study show the performance of in‐plane BISEPI. If multiple slices are needed, through‐plane 3D BISEPI may improve image quality and activity detection. Combining GRAPPA and BISEPI enables balancing tSNR and the number of volumes to achieve suitable activity detection by appropriately setting imaging parameters, including spatial resolution, GRAPPA acceleration factor, and the number of segments according to specific requirements. In this study, we empirically found that a tSNR of approximately 20, combined with 300 volumes, enables reliable detection of BOLD signal responses elicited by two types of stimuli in the primary visual cortex.

As an EPI acquisition and reconstruction technique, BISEPI can be integrated with other imaging methods, such as slice‐saturation slab‐inversion VASO, for submillimeter high‐resolution fMRI in challenging studies involving simultaneous BOLD and VASO activations. BISEPI also provides high structural contrast even in functional images, allowing them to serve as anatomical references for activity overlay. This utility, along with their improved signal quality, may be further enhanced by NORDIC denoising. Indeed, our results demonstrate that BISEPI improves the detection of active voxels, estimation of stimulus‐specific signal changes, and localization of functional activity in submillimeter high‐resolution fMRI. Our findings also suggest that optimal dedrifting ²⁸ , ²⁹ , ³⁰ may further improve the overall performance of functional imaging. It should be noted that a fundamental limitation of the proposed method is its incompatibility with tasks involving intertrial variability, such as learning, plasticity, adaptation, and many event‐related designs. Different segments will exhibit varying activations across trials, resulting in ghosting artifacts in the images.

It is important to clarify that the temporal resolution of BISEPI should not be misconstrued as increased acquisition efficiency. For submillimeter high‐resolution fMRI with a short‐duration stimulus design, BISEPI offers a viable solution for achieving high temporal resolution within a block. The experimental results demonstrate that BISEPI acquisition improves activity detection at 0.4‐mm isotropic resolution with a TRseg of 1.0 s, whereas the TRvol for conventional msEPI would be 4.0 s. Our findings, along with previous studies, ³¹ , ³² , ³³ , ³⁴ , ³⁵ suggest that BISEPI is advantageous for future submillimeter high‐resolution fMRI studies focusing on columnar, laminar, and other specialized imaging targets.

5. CONCLUSION

We propose a modified msEPI‐based imaging technique, termed BISEPI, that realigns k‐space trajectories during acquisition and reconstruction based on the timing information of stimuli from a block design paradigm. When combined with other imaging and postprocessing techniques, such as GRAPPA, k‐MoCo, and NORDIC, BISEPI functions in a manner similar to standard msEPI, while providing improved spatiotemporal resolution and functional/structural contrast.

This proposed method avoids GRAPPA‐related image deteriorations caused by B₀ inhomogeneity and high g‐factor penalty, enabling the detection of BOLD responses, including from low‐power stimuli, at submillimeter isotropic resolutions. By addressing the challenges posed by aforementioned problems, the method further improves the reliability and sensitivity of high‐resolution fMRI, particularly for capturing subtle neural activations at submillimeter scales. These findings may facilitate future studies investigating fine‐grained functional organization in the brain. In other words, the proposed BISEPI method aligns with the current trend in submillimeter fMRI sequence development toward advanced segmentation approaches. ³⁶ , ³⁷ , ³⁸ , ³⁹ Within this trend, the segmented approach presented here may contribute to ongoing developments aimed at improving the resolution of mesoscopic fMRI (e.g., layer‐fMRI) beyond the conventional 0.8‐mm limit.

Liu G., Ueguchi T., and Ogawa S., “Block‐interleaved segmented echo‐planar imaging for improved activity detection in submillimeter high‐resolution functional MRI at 7 T ,” Magnetic Resonance in Medicine 95, no. 1 (2026): 86–100, 10.1002/mrm.30569.