Figure 1. Diagram for the proposed approach.

Figure 5. (a) displays the scanning protocols. Radial-like corresponds to the undersampled radial trajectory, SPARKLING trajectory initialized with the undersampled radial trajectory, and BJORK trajectory initialized with the undersampled radial trajectory. (b) showcases one example from the in-vivo experiment. The first row is the CS-based reconstruction and the second row is the unrolled neural network-based reconstruction. For fully-sampled data, conjugate phase reconstruction is displayed.

Fig. 4. Dynamic cardiac valve images acquired using epi-SPEEDI (A) and epi-SPEEDI-kt (B). Each image corresponds to a specific time point during the aortic valve movement process. The temporal resolution was 0.6 ms. Images 6-30 show the opening status of cardiac valve (annotated by the yellow arrow) and the cardiac valve is closed in the other images. The red arrows indicate artifacts.

Fig. 2. Comparison of phase-encoding schemes used in epi-SPEEDI and epi-SPEEDI-kt sequences (A) and the k-space sampling pattern in epi-SPEEDI-kt (B). Nnav phase lines (yellow lines) in the central k-space (k_c) region are sampled in all the time blocks (TBs) while the outer k-space (k_o) regions are randomly and sparsely sampled (green lines). The k-space sampling patterns are the same at the echoes in one TB but different between TBs.

Sketch of the FID-ECCENTRIC sequence. 4-pulses WET water suppression precedes the excitation pulse. After the excitation, cartesian encoding along z-axis and gradient ramp along x and y axes are played before the acquisition simultaneous to the sinusoidal gradient waveform. Right, the parametrization of the circle position and the Fourier domain trajectory of ECCENTRIC for a 64x64x9 encoding matrix.

Choline (Cho), total NAA (tNAA) and ratio maps resulting from ECCENTRIC FID-MRSI acquisition performed on a brain tumor patient. Results from retrospective acceleration AF=2 and AF=3 show that the tumor lesion and contrast was preserved through acceleration. Spectra from the tumor location (1) and healthy tissue (2) are shown for all AFs. Bottom left, histograms of the Cramer-Rao Lower Bound (CRLB) for tNAA and Cho LCModel fit from the entire slab are presented.

Figure 3: Movie showing every fourth image frame from a k-t SWEEP acquisition acquired in a fetus aged 30⁺² weeks gestational age.The k-t SWEEP acquisition smoothly sweeps across the field-of-view (compare against the k-t M2D movie in Fig 2). In the k-t SWEEP data a stable signal is reached after an initial transient period, which manifests as noisy, fluctuating image frames at the beginning of the movie. Note: image quality not fully representative of scan quality due to the need for extreme gif compression.

Figure 1: Simulations of M2D and SWEEP [6] acquisition schemes. In both, the same number of images are acquired and coverage in the z-direction is equivalent. For the M2D acquisition (A), multiple dynamics are acquired for each slice z-location. (B) For each slice, 100s of RF pulses are required before the steady-state signal is reached. With SWEEP, (C) the excitation frequency of the RF pulse is slowly linearly increased resulting in a gradually shifting slice profile across the FOV. (D) After the initial transient period, a stable signal results throughout the rest of the acquisition.

Figure 4. Comparison of in-vivo images acquired using the proposed SPRING-RIO TSE method and standard Cartesian TSE. From top to bottom are trajectory- and off-resonance-corrected images from SPRING-RIO TSE with one signal average and with two signal averages, and images from standard Cartesian TSE. The red arrows point to structures where residual signal loss or artifacts exist, likely due to strong susceptibility, concomitant gradients or flow effects. The white circles indicate the spots where the image contrast is visually better in SPRING-RIO TSE than that in Cartesian TSE.

Figure 1. Pulse sequence diagram showing the sampling strategy, which includes fat saturation, field-map acquisition, and data acquisition using annular spiral rings. The center of k space is sampled by a self-retraced spiral in-out arm. The spiral-in rings in the orange box are designed to cover the outer k space, while the spiral-out rings in the blue box retrace the corresponding spiral-in rings. The refocusing RF pulse angles are set to 150° for reducing SAR.

Figure 2 Imaging sequences of T2-FSE and T2-DIADEM (A) and corresponding images (B). The flow compensation gradients on the DIADEM sequence are shown in red-colors in (A). For demonstration purpose, four slices among 38 are shown in (B).

Figure 3. Advantages of T2-DIADEM over T2-FSE: T2-DIADEM provides better depiction of internal details of left temporal lobe multi-cystic lesion (a), falx calcifications (white arrows in b), small T2 hyperintensities (c). T2-DIADEM identifies small developmental venous anomaly, as demonstrated in MPRAGE (green arrows in b)

Figure 3: (Animated gif, please click) Mid-ventricular short-axis cine acquired with phase-modulated bSSFP with BIPS and a long TR of 15ms enabling more efficient spiral readout and water-only excitation pulses. 60 cardiac frames were reconstructed with an in-plane resolution of 2.2mm; images from frequency bins were combined with RMS.

Figure 5: Diastolic cine frames in short-axis orientation (A,C) and orthogonal to right coronary artery (B,D) acquired with conventional Cartesian bSSFP (A,B) or the proposed phase-modulated bSSFP with BIPS method (C,D). Yellow arrows point to dark band artifacts in conventional bSSFP, and green arrows point at the coronary arteries that are obstructed by Indian ink artifacts in conventional bSSFP without fat suppression. Images from frequency bins were combined with complex sum and deblurred.

Figure 1: (a): A section of an ETGAR-II sequence consisting of two repetition times (TRs or shots). Blue, yellow and green gradients are the readout gradient pulses in the echo train. Red and purple gradients are the steering gradient pulses. Brown gradients are for phase-rewinding, as required by bSSFP. (b): k-Space trajectory of the sequence segment in (a). The dashed curves illustrate the effect of the steering gradient pulses or the rewinding gradient pulses on the k-space trajectory. The k-space rotation angle between two consecutive TRs is set to golden angle (111.24…°).

Figure 5: Human brain images obtained with single-echo golden-angle radial bSSFP (left) and ETGAR-II bSSFP with (middle) and without (right) phase corrections. With echo-train bSSFP, the scan time was reduced by 36% without substantial reduction in image quality as compared to single-echo bSSFP. In the middle image, the susceptibility/banding artifacts were slightly elevated in the frontal area (red arrow) due to the increased sensitivity to off-resonance in echo-train acquisitions.

Fig. 4. (a) compares overall image quality of 4 techniques and (b) shows a special case in the presence of metal. In (a), ssEPI shows strong signal pileup, which is reduced but still present in msEPI. Signal pileup is minimized in turboPROP (with slight signal drop off) and SPLICE-PROPELLER. SNR of turboPROP is much better than SPLICE-PROPELLER, while slightly lower than but still close to msEPI. (b) shows images from a volunteer with a surgical patch that causes signal pile up in ssEPI and msEPI, which is not visible in turboPROP. T2 TSE is acquired as reference

Fig. 5. Capability of motion correction with turboPROP. Image quality of both msEPI (left column) and turboPROP (middle column) are degraded by rigid body motion. With motion correction (including both PROPELLER motion correction of each source image and image registration between diffusion directions), the motion artifacts can be minimized (right column, same raw data as the middle column).

Figure 3: Representative slices of the whole brain MPRAGE scans. Note that the same windowing is used for all images, which highlight the similar contrast between the conventional (top row) and quiet scan (bottom row). A slight misalignment between anatomy might be visible due to subject movement.

Figure 4: Part of the sound measurement used to estimate the sound levels. Here, three shots of the MPRAGE-scans are displayed. Top: the audio signal from the microphone after A-weighting. Bottom: the calculated sound pressure levels (dB(A) and fast time weighting) for each shot.

Figure 4: Top) Mean low-resolution images (1x5x1mm³) for each of the different F-states in the T₂-COMBINE acquisition (TR=8ms, TE=4ms, α=1°). Centre) Bicubic interpolation of the mean low-resolution images (top) to a nominal voxel size of 1mm isotropic. Bottom) Corresponding COMBINE reconstruction to a nominal voxel size of 1mm isotropic.

Figure 2: Schematic of the pulse sequence for the T₂-COMBINE experiment. Higher order F-states are reached by adding integer increments of the unbalanced gradient area (denoted A) either side of the readout.

Sequence diagram of the proposed sequence. It employs standard radial FSE readouts, interleaved with diffusion blocks. The diffusion blocks can be placed after arbitrary numbers of readout blocks. The figure shows one diffusion block after 2 readout blocks each. For simulation and measurements, we use a combination of 3 readout blocks and one diffusion block (not shown).

Results of the phantom measurement. The left column shows ADC values, the middle column shows T₂ values. The first two rows show the reconstruction result for different number of spokes per ΔTE-b-pair (30 and 11 spokes). The third row shows reference measurements using conventional DW-EPI and FSE sequences. Rows 4 and 5 show corresponding difference images. The two plots on the bottom right show a ROI based analysis for all cases. The position of the ROI is indicated in the top right plot.

T2 maps for the ETL = 96 (left) and ETL = 192 (right) variable flip angle trains computed from the two echo times of 6 and 132 ms. Noise is more visible in the 192 case. Note that the slice for the 192 case is slightly offset from the 96 case due to the difference in resolution (0.9 mm³ vs 1.3 mm³)

Images acquired using the proposed echo trains (left four images – 96, and right four – 192). The top row are all images at TE = 6 ms, and the bottom is at TE = 132 ms. Axial and sagittal views are shown for each combination of TE and train length.

The single-shot ES-GRASE sequence. The areas and polarities of prewinding and rewinding gradients (i.e. the green gradient) of the readout gradient are designed to remove high-order echo pathways. The alternating direction of the refocusing RF pules (i.e. 180x-180y-180x-180y) can generate 90° phase difference between the primary and residual high-order echo pathways. The phase difference is 90+2𝜃° when the background phase is involved.

Data preprcessing frameowrk: First, images of four echoes were reconstructed using parallel imaging techniques, and the N/2 ghost was removed using phase cycling method. Second, the background phase was estimated from the first-echo image (see the upper panel). Third, rotating the signal from the third and fourth echo with the estimated background phase . The real part of the rotated signal is only from the signal of the primary echo pathway, with which the primary echo signal can be calculated by basic Trigonometry (see the lower panel in which only the third echo is shown as an example).

Figure 1. EPG Simulation and scan parameters. (a) RF flip angles (90-degree excitation not shown). (b) Transverse and longitudinal magnetization for T1=1600 ms, T2=80 ms (after arriving at a stable equilibrium magnetization). (c) Scan parameters for simulation and phantom experiments. The first acquisition block uses an FSE train with non-selective variable refocusing flip angles and a fast recovery tip-up pulse. After an adiabatic inversion pulse, a series of SPGRE blocks are played in sequence.

Figure 4. Animated gif of the reconstructed time series of images at three different slices in the ISMRM-NIST phantom, corresponding to the three plates of contrast spheres. In total, a time series of 300 images are reconstructed.

Fig 2. (A,B) Coronal and (C,D) axial reformatted images of a painful left revision total hip replacement in a 61-year-old female demonstrating abnormal signal adjacent to the lateral aspect of the femoral component on isotropic images (arrow). However, this abnormality cannot be seen on conventional MAVRIC SL images, despite the overall image quality being graded higher. Signal abnormality corresponded to fluid on inversion recovery sequences, confirming loosening (not shown).

Fig 3. (A,B) Coronal and (C,D) axial reformatted images of a painful right total hip replacement in a 73-year-old male. A discrete trochanteric fluid collection is present (arrow), but is less well seen on conventional MAVRIC SL, particularly along the superior margin. Axial reformatted images confirm the presence of a trochanteric fluid collection, but the collection is less well seen on an axial reformatted conventional MAVRIC SL image (arrow).

Excitation (red curves) and refocusing (purple curves) profiles of regular pulses (a), which are used by default in 1.5 T Siemens Magnetom Avanto scanner for FSE in "Low SAR" mode, and of the constructed SLR pulses (b) for a sample with neglected relaxation. Refocusing profiles of four echo signals obtained in steady-state FSE modeling with regular pulses (c) and SLR pulses (d) (T1 = 329 ms). Black dashed vertical lines indicate the desired slice thickness (TH), black solid vertical lines indicate the entire refocused area of profiles (TH1). (c) Selectivity metric (S/S₁) in FSE modeling.

(a) Parameters of RF pulses used in modeling and experiment. (b) The evolution in time of RF amplitudes (black and blue curves for regular and SLR-based FSE, respectively) and slice selection gradient amplitudes with crushers (orange curve) that were generated. An additional crusher gradient was applied in the end of the cycle (not shown), in order to fully spoil the transverse magnetization. The parameters of the pulse sequence: repetition/echo time TR/TE = 550/13 ms, excitation angle α = 90°, refocusing angle β = 180°, inter-echo time IET = 13 ms, echo train length ETL = 4.

Figure 1. The diagram in one echo spacing of the TSE Dixon method (a) and proposed combined echo two-points Dixon sequence (b). The dead time is marked with diagonal stripes, where no data sampling happens. In the proposed method, the dead time is reduced. Note that E1 and E4 are partially opp-phase echoes, while E2 and E3 are partially in-phase echoes.

Figure 4. Water images of a volunteer acquired with product TSE Dixon (left) and combined echo two-points Dixon (right). SNR and image quality are comparable.

Figure 5: Double resonance effect at two different frequencies of the Spin locking (a) 120 Hz and (b) 240 Hz. For both images, the frequency of the applied current was set at 120 Hz and the amplitude was calculated by the Biot-Savart law to be approximately 10 nT. To facilitate the visualization of the results, the rest of the deconvolution between SL on and SL off is displayed.

Figure 3: Bloch simulation of the behaviour of the three SL approaches showed in Figure 1. a) Contrast as a function of the SL frequency for homogeneous fields. b) Contrast in resonance (120 Hz) for the three approaches as a function of the excitation angle, where the variation is induced by B1 inhomogeneity. c) Contrast in resonance (120 Hz) for the three approaches as a function of the off resonant frequency induced by B0 inhomogeneity.

Integrated signal area of the lactate doublet at 1.3 ppm as a function of the echo time (T_E = τ₁ + τ₂). The coloured lines represent different combinations of τ₁ and τ₂ to achieve the effective T_E with either τ₁ = τ₂ (red), holding τ₂ constant (blue) or τ₁ constant (green). The measurements were repeated with four different orientations of the RF coil (grey box) relative to the main magnetic field (0° to 45°) with estimated fibre orientations ranging from 42° to 16° (projections to the coronal plane hatched in red).

Averaged ¹H MR spectra (n = 30, T_R = 3 s) acquired using semi-LASER (left) and with the adiabatically refocused DQF acquisition scheme (right). Spectra were acquired from (a) lithium lactate solution in a spherical phantom (d = 160 mm, 96mM LiLac in H₂O) and (b) sodium lactate (60 %) injected into a pork meat specimen. Only lactate is visible (at 1.3 ppm) in spectra from both experiments with a peak splitting of 7.3 Hz for lactate in solution and 13 Hz for lactate injected in meat.

Figure 3 Coronal images of mid-tongue region in 28 frames (about 0.8s) of subject counting ‘Four’

Figure 1. A simplified pulse sequence diagram illustrating (k, t)-space sampling patterns. The navigator dataset is acquired using a spiral trajectory and is only acquired once every 4 imaging data acquisitions. The imaging dataset is acquired using a Cartesian trajectory with random phase and slice encoding to provide structural quality images.

Figure 4. Experimental readout-segmented 2D images in a healthy subject: Top row) single-transmit, Middle row) volumetric-B₁⁺ shimming, and Bottom row) slice-by-slice B₁⁺ shimming. The arrows highlight a few areas where slice-by-slice shimming mitigates the B₁⁺ inhomogeneity particularly well compared to single-transmit and volumetric-shimming. Meanwhile, there is a global improvement in the column 4 slice. Each column of slice images are windowed the same.

Figure 2. Magnitude RF waveforms in volts for 8 transmit channels, excitation pulse on the left and refocusing pulse on the right. Top row) single-transmit waveforms (all channels same magnitude with incremental 45° phase offset); Second row) volumetric-shim waveforms (in this case, the shimmed solution alternated between two magnitude values on all channels); Third row) slice 1 shim waveforms; Fourth row) slice 4 shim waveforms.

Fig. 1. Pulse sequence diagram of DW-SESTE with the flipangle α and amplitude G_D and duration δ of the diffusion gradients.

Fig. 2. (a, c) ROI (red dots and crosses) and (b, d) signal-b curves of UHb-DWI. Signal-b curves were fit to a single-exponential function for the phantom data in (b) and human aDWI data (red O, X)) in (d), and a double-exponential function for human rDWI data (black O, X) in (d).

Figure 3. Quantitative mapping results of one healthy volunteer (male, 30 years). (a) T1 maps estimated by modified Look-Locker inversion recovery (MOLLI) and the proposed sequence; (b) T2 maps estimated by multi-echo tubor spin echo (ME-TSE) and the proposed sequence; (c) T2* maps estimated by multi-echo turbo field echo (ME-TFE) and the proposed sequence; (d) vessel wall image reconstructed from the proposed sequence.

Figure 2. Quantitative mapping results of the phantom. (a) T1 maps estimated by inversion recovery spin echo (IR-SE) and the proposed sequence with quantitative comparison (mean and SD); (b) T2 maps estimated by multi-echo spin echo (ME-SE) and the proposed sequence with quantitative comparison (mean and SD); (c) T2* maps estimated by multi-echo gradient echo (ME-GRE) and the proposed sequence with quantitative comparison (mean and SD).

Figure 2. Representative images showing the effect of additional magnetization preparation pulses (f=1200 Hz, a=1.0) on EPI FLAIR contrast compared to the clinical reference TSE FLAIR scan (top left). Optimal EPI FLAIR contrast (i.e., most closely matching the TSE reference scan) was achieved with 7 preparation pulses (see Figure 4 for quantitative evaluation).

Figure 4. Contrast ratio for EPI FLAIR images (defined as average GM / WM signal) as a function of the number of MT preparation pulses (n), resonance offset (f), and pulse amplitude (a). Contrast ratio for the TSE reference scan is shown by the dashed red line. Optimal contrast (i.e., most clostly matching the TSE reference contrast) was achieved with n=7, f=1200Hz, and a=1.0.

Figure 3. Efficacy of HSn pulse (47-year-old male). (A) ZTE imaging without IR and (B) STAIR-ZTE. A regular ZTE does not show a dramatic difference between images with a hard pulse and an HSn pulse because the excitation bandwidths of both hard and HSn pulses are broad enough to include the targeted region. STAIR-ZTE shows dramatic improvement with an HSn pulse, where the low frequency signal bias (yellow arrows) affecting myelin contrast (red arrows) is well suppressed. With suppression of long T₂ signals, the remaining myelin signal becomes much lower and prone to imaging artifacts.

Figure 5. A representative MS patient (22-year-old female). (A) MP-RAGE, (B) FLAIR, and (C) STAIR-ZTE. STAIR-ZTE delineates a demyelinated lesion which corresponds well with the clinical MP-RAGE and FLAIR images (red arrows).

Figure 3. T2w-FSE (A, B, C), FSPGR (D, E, F), 3D IR-FS-UTE Cones (G, H, I), and 3D DIR-UTE-Cones (J, L, M) performed in knees of two healthy volunteers (the first row represent the first volunteer). On T2w-FSE and FSPGR, the signal in the OCJ region cannot be detected due to its fast signal decay. (G-I) 3D IR-FS-UTE Cones sequences highlighting the OCJ and suppressing SC and subchondral bone. (J-L) 3D DIR-UTE-Cones sequences also highlight the OCJ but with better subchondral bone fat suppression.

Figure 4. T2w-FSE (A, B), FSPGR (C, D), 3D IR-FS-UTE Cones (E, F), and 3D DIR-UTE-Cones (G, H) performed in the knees of two patients with osteoarthritis. (Arrows in A, C, E, G) show interruption of the bright-line representing OCJ on the weight-bearing surface of the medial femoral condyle and medial tibial plateau in the first patient and on the femoral trochlea in the second patient (open arrows in B, D, F, H). The 3D IR-UTE-FS Cones sequences (G and H) are more efficient in suppressing the subchondral bone fat than the 3D IR-FS-UTE Cones (E, F).

Figure 2. The off-resonance artifact improved with the deblurring algorithm in both the 2D Spiral (a.1-2) and the 2D Radial (a.3-4) acquisitions. The off-resonance artifact improved with the deblurring algorithm in both the 3D Spiral-cone (b.1-2) and 3D Radial (“Koosh-Ball”) (b.3-4) acquisitions. The red and green enlargements in figure b highlight the comparison before and after the correction in 3D. Beginning from the top and going clockwise, the vials contain rubber, water, agar and oil.

Figure 4. Comparison of 3D radial and spiral trajectories among the regular half-Sinc pulse (RHSP) excitation, VERSE-modified half-Sinc pulse (VHSP) excitation, and the rectangular pulse (RP) excitation with matching scan parameters (a.1-6). SNR of the rubber (short T2 material) was generally higher for the spiral trajectories as compared to the radial trajectories. SP-spiral; and RD-radial. Beginning from the top and going clockwise, the vials contain rubber, water, agar and oil.

Figure 3. the images including the resolution grids and 1D profiles of the ACR phantom acquired using conventional UTE (C-UTE) (a-d) and VS-UTE (e-h). The C-UTE image is difficult to show the internal structure due to the phantom signal outside the ROI when the number of projections was 5k, but VS-UTE provided a better internal structure image. Also, the overall CNR (contrast to noise ratio) of C-UTE was lower than that of VS-UTE(i-l).

Figure 1. Pulse sequence diagrams(a, b) and data acquisition method(c, d) of conventional UTE (C-UTE, a) and volume-selective UTE (VS-UTE, b). C-UTE, which uses a non-selective pulse for spin excitation(a), acquires projection data including spin information of the entire object(c). In contrast, The spin excitation of the unexcited object is not included in the signal(d) by using the slab selective pulse and the slab selective gradient simultaneously and setting the projection direction in the same direction as the slab selection direction(b).

Designed (left) and precompensated (right) waveforms for the Standard trajectory, from (a) the initial design, then adding (b) frequency limiting, (c) transition smoothing, and (d) precompensation conditioning. Frequency, slew, and gradient amplitude waveforms are colored red, blue, and green, while the gray waveforms show the positive part of Gx. The blue arrows illustrate the spikes in the slew rate of the precompensated waveforms, which are further mitigated with each layer. Waveforms are normalized by the design maximum for easier visualization.

Designed (left) and precompensated (right) waveforms for the (a) spiral-in, (b) variable density, and (c) single shot trajectories. Frequency, slew, and gradient amplitude waveforms are colored red, blue, and green, while the gray waveforms show the positive part of Gx. Waveforms are normalized by the design maximum for easier visualization.

Figure 3. Comparison between Cartesian TSE and Spiral TSE images at five representative slices. The total scan time were 4 min 30 sec, 3 min 48 sec and 1 min 54 sec respectively.

Figure 1. a) The spiral in-out trajectory used in TSE sequence; b) Illustration of reversed arm ordering for T2-decay compensation.

Fig. 3: Sequence comparison results in the brain of a healthy volunteer. A particularly sensitive area are the sinuses (depicted by arrows). B₀ inhomogeneities again lead to geometric distortions (EPI), blurring (Spirals) and also signal dropout, particularly in case of the single shot spiral.

Fig. 1: Sequence diagram and corresponding k-space of the 3 BURST sampling schemes. A train of 4 slice-selective RF pulses excites magnetisation, storing it in outer parts of k-space through gradients in (logical) xy (top and middle) or z direction (bottom). The readout is split into 4 segments, and each respective sub-RF pulse is refocused in centre of k-space (central line of EPI and beginning of Spiral trajectory). Hence, in the k-space depiction, the 4 separate regions in k-space all encode a similar area in k-space with EPI shifted by a quarter of a line and Spiral interleaved by 90°.

Fig. 3. Comparison of geometric distortion artifacts in 2 volunteers. The strong distortion artifacts in ssEPI (red arrows) are reduced but still present in msEPI (yellow arrows), while are eliminated in ssGRASE, as indicated by the contour extracted from FLAIR.

Fig. 5. Illustration of T and L spine images. As can be observed, strong distortions exist in ssEPI. These artifacts are significantly reduced in msEPI with some residuals. msEPI may exhibit some artifacts (cyan arrows) and degraded quality of the vertebral body (orange arrows), which are improved in ssGRASE. It is typical that the spinal nerve bundle is not well seen in L-spine, where instead the vertebral bodies are the primary focus in many neurological applications. The shading in the lower corner of ssGRASE T-spine image (blue arrow) will be investigated.

Eight consecutive frames from systole to diastole. Two slices were imaged simultaneously and shown in the first and second rows. The temporal resolution of each frame is 3.7 ms × 15 views per shot = 56 ms. The heart contraction and relaxation motion is well preserved in both slices.

Slice-dependent phase modulation can be achieved by adjusting the RF excitation pulse. The blue and red pulses are components of the multiband rf pulse for slice 1 and slice 2 respectively. The phase of slice 1 is not modulated, while the phase of slice 2 changes by π in between PE lines. In adjacent frames, a phase offset of π is applied on slice 2.

Fig. 3 Sagittal example slice of S(TR1) magnitude images and corresponding FA maps for 5mm isotropic AFI (A, top) and AFEPI protocols at 2.5mm (B,C) and 5mm (D,E) isotropic resolution without distortion correction. Regardless of TE and segmentation factor, major signal dropouts in the FA maps of 5mm AFEPI scans with anterior-posterior PE direction make those FA maps unreliable above the sphenoid sinus. The 2.5mm AFEPI maps match well with the AFI reference maps, again regardless of TE and segmentation. [Open figure to see both PE directions as animated GIF.]

Fig. 1 Schematic sequence diagram of the AFI sequence (top) with optional multi-echo extension (dotted outlines), and the proposed AFEPI sequence with y- and a z-blips (red) to achieve skipped-CAIPI sampling (bottom). Spoiler gradient pulses (shaded gray) correspond to 6π/2π/2π (RO/PE/3D) and n times increased voxel dephasing at the end of TR1 and TR2, respectively². For simplicity, an example EF=5 and n=TR2/TR1=3 is displayed.

Figure 4 T2*-weighted 3D EPI images acquired using the proposed implementation with 2 subjects. Imaging protocols are summarized shown in Figure 3. Sections of the images were enlarged for detailed examination.

Figure 1 Examples of k-Space sampling patterns when fully-sampled (a), and with GRAPPA 2×1 (b) and CAIPIRINHA 1×2 (c) accelerations. (a): The 3D EPI volume was separated into shots in the kx-ky plane (four shots in this case; blue, yellow, red and green) and partitions in the kz direction. (b): The GRAPPA acceleration was achieved by skipping the 2nd and 4th shots for all the partitions. (c): The CAIPIRINHA acceleration was achieved by alternatively skipping 2nd and 4th shots and then 1st and 3rd shots between successive partitions.

Figure 2 Experimental results of ACR phantom in inhomogeneous field. (a) Image acquired by interleaved single-shot EPI. (b) Image acquired by conventional EPI. The yellow boundaries were extracted from the standard 2D FSE image. The vertical stripes are induced by phase correction error.

Figure 3 Results on human brain. Left column: Images acquired by isEPI. Right column: Images acquired by conventional EPI. Both isEPI and conventional EPI images were acquired with 2-fold acceleration. The yellow boundaries were extracted from the standard 2D FSE image.

Figure 1. Schematic plot of Varied ESP EPI pulse sequence design.

Figure 3. a) The optimized Varied ESP values along the echo train; b) Noise spectrums of different sequences; c) Zoomed part of interest in Figure 3b.

Figure 1: (A): A schematic to illustrate 3D rFOVI by replacing slice selection with slab selection using a 2D RF pulse in (C), followed by through-slab phase-encoding as shown in (B). (B): A conceptual 3D rFOVI sequence with phase-encoding gradient along the slab direction. (C): Details of the 2D RF excitation pulse with a fly-back EPI excitation k-space trajectory to avoid the issues associated with Nyquist ghosts.

Figure 4: (A): Visual fMRI activation maps, selected from the 3D volume, overlaid onto the baseline images acquired with 3D-rFOVI GRE-EPI. (B): The averaged time course showing the activation. The signal change in (B) is approximately 2.5%. The green bar represents the time period for visual stimulus, which was 24s.

Figure 3: Rat 1, coronal slice, at the left side MP-RAGE with a simple inversion, at the right side MP-RAGE with a T2prep-IR, both optimized with an optimal control framework aiming at enhancing contrast between cortex and corpus callosum. Same acquisition parameters are used (segment duration: 3s, TR: 6.5ms, flip angle: 13°).

Figure 2 : Contrast-to-noise ratio over the two rats, the four segment duration and the different slices orientations

The bSSFP profiles for the mean value of a white matter ROI, shown in the reference on the bottom right, are displayed. The asymmetry for $$$\mathit{TR}=5\,\mathrm{ms}$$$ $$$(\mathit{AI}=0.14)$$$ and $$$\mathit{TR}=3\,\mathrm{ms}$$$ $$$(\mathit{AI}=0.08)$$$ is clearly visible, while the profile for $$$\mathit{TR}=1.5\,\mathrm{ms}$$$ $$$(\mathit{AI}=0.02)$$$ exhibits almost no asymmetry.

The calculated $$$\mathit{AI}$$$ maps for the different $$$\mathit{TR}$$$ are shown for three different axial slices of the in vivo brain scans. It is clearly visible that, overall, the values for $$$\mathit{TR} = 1.5\,\mathrm{ms}$$$ are smaller than for $$$\mathit{TR} = 3\,\mathrm{ms}$$$ and $$$\mathit{TR} = 5\,\mathrm{ms}$$$. The right pictures show anatomical references.

Figure 1. Refocusing radio frequency pulse scheme for TRAPS HASTE DW: Central k-space lines (30 lines, including preparing echoes shown by the red color) are sampled with nominal 180° flip angles, followed by a smooth ramp of 6 echoes (TRAPS). All other refocusing pulses have a low fixed nutation angle. Horizontal axis shows the number of echoes collected for the protocol used in this study.

Figure 3. Example coronal slices from TRAPS HASTE DW in a healthy female subject at 3T: (left) T₂-weighted (b0), (middle) diffusion-weighted (b = 1000 s/mm²), and (right) trace apparent diffusion coefficient (ADC) maps. Background noise has been masked from ADC maps.

An experimental demonstration of the distortion correction technique, showing a single plane along the second phase encoded dimension of the image to highlight the correction along the frequency encoded axis. Zoomed insets show the geometric and intensity distortion correction, and the B0 estimate is in units of voxels. The receiver bandwidth was 650 Hz/voxel, so the B0 estimate in this plane corresponds to approximately +/- 3250kHz.

A comparison of the normalized fully sampled reconstruction to the accelerated reconstruction for one gradient polarity in a single plane along the frequency encoding dimension. There are some small high spatial frequency artifacts along the edges of the plastic grid in the phantom, but the residual aliasing artifacts are extremely minor. The peak SNR in the fully sampled cross section shown is approximately 97.43, and approximately 36.83 in the accelerated scan, in good agreement with the g-factor map and the acceleration (R=3x2) used.

Figure 1: MTF (a-c) and PSF(d-f) at variant Mz(0) levels, and the linear relationship between Mz(0) and the PSF peak-amplitude (g-i) of a typical TFE (left), bSSFP (center) and TSE sequence (right) when T1=1000ms, T2=100ms and T2*=50ms. Optimal TFE/TSE factors and FAs are chosen for TFE and TSE sequence at TR/ESP=10ms that maximize CNR. The TFE factor and the flip angle of bSSFP are chosen with 95% maximum CNR. Grey lines in g-i demonstrate contrast loss in all sequences and increased intercept (bias) in TFE and bSSFP with TR or TFE/TSE Factors 8 times larger than the optimal choice.

Figure 2: MTF (a-c) and PSF(d-f) with different TFE/TSE factors of a typical TFE (a,d), bSSFP (b,e) and TSE sequence (c,f) when T1=1000ms, T2=100ms, T2*=50ms and Mz(0)=1. The FWHM of the main PSF lobe (g) and CNR (h) are plotted as curves over TFE/TSE factors, demonstrating the blurring effect and contrast preservation performance, respectively. The colored markers are from the respective TFE/TSE factors shown in (a-f). All methods are subject to peak-amplitude loss of the PSF peak with longer TFE/TSE factors.

Figure 1: MRzero schematic. Figure a) shows the general MRI pipeline from spin system to reconstructed image. A differentiable MR scanner simulation implements Bloch equations for signal generation. (b): The output of forward process is compared to the target image, analytical derivatives w.r.t. sequence parameters are computed using auto-differentiation, and gradient descent is performed in parameter space to update sequence parameters. (c) Final or intermediate sequences can then be applied at the real scanner using the pulseq framework⁷.

Figure 2: Learning RF and spatial encoding. Row a: k-space sampling at different iterations, Row b: flip angles over measurement repetitions. Row c: simulation-based reconstruction at different iterations 9, 99, 255, 355, and 1000. row d: phantom measurement, row e: in vivo brain scan. Row f: training error curve. An animated version can be found at: www.tinyurl.com/y4blmpe7. Target sequence: 2D transient gradient- and RF-spoiled GRE, matrix size 96, TR = 25 ms, TE = 3.2 ms, FA=5˚.

Figure 3. Optimization of free non-Cartesian k-space trajectory and RF flip angles for parallel imaging (R=4) at various intermediate iterations, including the final result at the last iteration. (A) Reconstructions, (B) difference maps to the target image (Figure 2A) and (C) current trajectory (colors corresponding to shots) and (D) current flip angles across the iterations. (E) Loss curve, i.e. normalized root mean squared error to target image over iterations. An animated version of the Figure can be found at https://tinyurl.com/ydgb9gf8

Figure 2. Optimization of free non-Cartesian k-space trajectory for parallel imaging (R=3) at various intermediate iterations, including the final result at iteration 2550. (A) Reconstructions, (B) difference maps to the target image (Figure 2A) and (C) current trajectory (colors corresponding to shots) across the iterations. (D) Loss curve, i.e. normalized root mean squared error to the target image over iterations.

Figure 4: DANTE-SPACE data acquired using (a) 300 DANTE-pulses of 10 degrees, (b) 200 DANTE-pulses of 9 degrees, and (c) without DANTE-preparation. (d-e) The mean of Figures (a) and (b), showing (d) conservatively drawn tissue masks for the vessel wall (red), CSF (orange), and lumen (yellow), and (e) an example of a vessel wall location (red line). (f) The corresponding vessel wall profiles using “standard” DANTE-preparation (a) and “optimized” DANTE-preparation (b). Dashed lines indicate Gaussian fits to the 7 pixels surrounding the peaks.

Figure 3: Simulations of DANTE-SPACE using different DANTE-parameters. Blue dots correspond to “standard” DANTE; green asterisks to “optimized” DANTE. Black and blue lines indicate iso-contour SAR-lines of the DANTE-preparation (which accounts for approximately 30% of the total SAR in DANTE-SPACE), relative to the SAR of “standard” DANTE; values for each line are shown in (a). Figures show (a) contrast between the vessel wall and blood, (b) contrast between the vessel wall and CSF, (c) the FWHM of the vessel wall signal (px.), and (d) intensity of the vessel wall signal.

(a) Schematic representation of the unsupervised neural network for better sampling uniformity and reduced acoustic noise autocorrelation. The two green arrows indicate gridding operators. The gridding operator for kspace coordinates grids kspace to 4D array (3D + time) for assessing sampling density. (b) In our design, the TR is variable with a small delay and learned delay prior to excitation. These delay values are learned though gradient descent of the acoustic loss function through the gridding operator.

Phantom comparison of bit-reverse, golden mean, and learned view orders demonstrate a high level of sampling flexibility in learned sampling. Using 10,000 angles, learned angles avoid streak artifacts which occur in both bit-reverse and golden means (arrows). Subsampling acquisitions to 3750 projections maintains a low level of coherent artifact.