Radial compressed sensing imaging improves the velocity detection limit of single cell tracking time‐lapse MRI

Time‐lapse MRI enables tracking of single iron‐labeled cells. Yet, due to temporal blurring, only slowly moving cells can be resolved. To study faster cells for example during inflammatory processes, accelerated acquisition is needed.


INTRODUCTION
With increasing knowledge on cell-based therapies and the interplay of immune cells in inflammation and cancer, the need for reliable methods to track cells is growing.Various MR-, optical-, and radioisotope-based imaging methodologies exist to track cell motion in vivo, but all suffer from different disadvantages like invasiveness, low spatial or temporal resolution, limited penetration depth, or low sensitivity. 1Tracking of single cells moving at high velocity inside blood vessels is only possible with intravital confocal microscopy using cranial or skinfold windows thus accompanied by an invasive procedure and limited field-of-view. 1 For immune cells, different motion patterns have been observed previously: While cells travel passively within the blood stream at velocities >12 mm/min, 2 also active migration of leukocytes along the endothelium has been described.Monocytes start rolling with velocities of around 2.4 mm/min upon an inflammatory stimulus, 3 while non-classical monocytes intermittently exhibit a slow movement along the vascular endothelium under healthy conditions. 3This "patrolling" with a mean velocity of 12 μm/min occurs for varying durations, and is either limited to small areas (<40 μm, i.e., within one MRI voxel) or may also exceed distances greater than 100 μm. 3 MRI offers a non-invasive alternative imaging approach that allows unlimited tissue penetration and whole-organ coverage at high spatial resolution, simultaneous anatomical information and high soft tissue contrast. 4,5[8][9][10] Consequently, single cells can be resolved as hypointense spots in T2*-weighted images both, in vitro and in vivo.2][13] However, when cells move too fast, temporal blurring occurs and, for current frame rates of 8 min, cells moving with velocities beyond a few tens of μm/min cannot be distinguished from the background anymore. 13Although changes in the behavior of patrolling monocytes as part of an immune response can be examined, 12,13 rolling or flowing monocytes cannot be resolved yet.Hence, accelerated data acquisition is necessary to gain a deeper comprehension on inflammatory processes or cancer progression through single cell tracking with time-lapse MRI.
One way to achieve a higher temporal resolution in MRI is undersampling k-space, thus acquiring fewer data in shorter scan times, 14 which has recently been shown to improve Cartesian sampled time-lapse MRI when combined with compressed sensing (CS). 15While for Cartesian sampling, due to the fixed phase-encoding direction, regular undersampling results in coherent aliasing artifacts observed as wrap-arounds, for radial k-space trajectories aliasing artifacts are distributed along varying directions.These noise-like, incoherent artifacts, as well as streaking artifacts resulting from regridding algorithms can be tolerated to a certain extent since the main image content is preserved superseding the need for additional reconstruction techniques. 16Moreover, since the center of k-space is sampled several times and with higher density, radial sampling is robust against motion, 17 which benefits imaging dynamic processes.In the case of irregular radial sampling patterns, additionally CS can be applied to remove noise-like artifacts allowing even higher undersampling ratios without a loss in image quality. 18uch irregular sampling can be achieved by an interleaved acquisition mode, where subsets of all spokes, each covering k-space uniformly, are acquired sequentially.Fully sampled (FS) and accelerated frames are obtained simultaneously, and both can be reconstructed retrospectively, similar to a recent approach for Cartesian sampling. 15o assess the currently achievable maximum detectable speed of cells in time-lapse MRI, contrast simulations and in vivo measurements were performed previously. 12,13Although they provide a good estimate, a more quantitative measure of the velocity detection limit is advantageous to evaluate possible acceleration techniques in dynamic single cell tracking.Thus, phantoms imitating slow motion of the order of 10 0 -10 2 μm/min need to be developed.Also in other MRI research areas, like flow measurements using phase contrast MRI, 19,20 such phantoms can be of interest to validate simulations or in vivo measurements.
In this work, first, a rotating phantom system was built that mimics moving cells with a known and adjustable velocity around the estimated detection limit of time-lapse MRI of 60 μm/min. 13The velocity detection limit was determined by comparing the signal contrast single micron-sized iron particles (MPIOs) generate, which is similar to the contrast of in vivo labeled cells. 11,13econd, radial acquisition was applied to single cell tracking with time-lapse MRI.An interleaved acquisition scheme permitted retrospective reconstruction of both FS and accelerated images.Undersampled (US) data were reconstructed with and without CS to improve SNR and detection of single cells.In phantom measurements and in in vivo time-lapse MRI of mouse brain, temporal resolution was improved, and the velocity detection limit increased.
Last, Cartesian and radial sampling in dynamic single cell tracking was compared regarding the maximum detectable velocity of cells.

F I G U R E 1
Rotating phantom system.(A) Design sketch and (B) photo of the whole system installed at the small animal MR scanner.The system can be divided into two parts as indicated by the broken lined rectangles: (C-E) an MRI insert and (F) the drive frame that is mounted on the AutoPac table.(C,D) The MRI insert consists of the acrylic glass construction, where ball bearings ensure a smooth rotation, and (E) a glass fiber stick holding the phantom.(F) The aluminum drive frame accommodates the drive mechanics like motor, toothed belts, and gear wheels.All distances are indicated in mm.

METHODS
All MRI experiments were performed at 9.4 T on a 94/20 Bruker Biospec equipped with cryoprobe and an AutoPac table, and operated under Paravision 6 (Bruker Biospin, Ettlingen, Germany).

Rotating phantom system
A partly MR-safe system (Figure 1) was built in house, which enabled the rotation of 2 mL Eppendorf phantoms at very low rotational velocity.The system consisted of two modules linked through a carbon drive shaft: an MRI insert holding the phantom and a drive frame that was mounted on the AutoPac table.
The MRI insert was manufactured from acrylic glass and was placed inside the support mounting of the cryoprobe.The Eppendorf tube phantom was fixed on a glass fiber shaft.Glass ball bearings assured a smooth rotation and a centered position of the phantom inside the scanner.
The drive frame was made of aluminum and accommodated the stepper motor with mounted precision planetary gear and optical encoder.The rotation was driven via toothed belts and gearwheels with an additional gear ratio of 24:18.
Motor power supply and controller were placed outside the scanner room.The motor was operated using Plug&Drive Studio (Nanotec, Feldkirchen, Germany).
A list of all components can be found in Table S1.

Measuring the velocity detection limit
The velocity detection limit of time-lapse MRI was defined as the velocity at which single iron particles were not discernible from the background anymore, and was determined using an agarose phantom containing embedded MPIOs as follows: First, T2* weighted images were acquired for the static phantom as a reference.Individual particles were identified as hypointense spots and the signal loss SL stat,i was calculated as SL= [mean(signal intensity enclosing area ) -min(signal intensity hypointense spot )]/ mean(signal intensity enclosing area ) for all detected particles i (Figure S1A-C).
Subsequently, the phantom was scanned while rotating.Rotational speeds ranging from 1.46 × 10 −3 rpm to 4.69 × 10 −2 rpm resulted in particle velocities of up to 1.3 mm/min, which were calculated based on the radial distance of the hypointense spots from the axis of rotation in the stationary image.The signal loss during rotation of the phantom SL rot,i of all detected particles i was determined.Due to temporal blurring, some particles were not detectable anymore in the rotating phantom, that is, not visible to the observer's eye.Then, SL rot,i was set to zero.These and particles with a signal loss SL rot,i below 0.05 are referred to as "non-visible." For each hypointensity, the change in signal loss due to rotation was calculated as ΔSL i = SL stat,i − SL rot,i .For visible and non-visible particles respectively, linear regression of the velocity-dependent change in signal loss was performed independently using MS Excel (Microsoft Corporation, Redmond, WA, USA; LINEST-function).For visible MPIOs the line fit was forced to pass through the origin.The intersection of the two fits, resulting from visible and non-visible particles, was defined as the velocity detection limit v max (SL stat ).This calculation was repeated for initial signal losses SL stat of [0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5] where only respective particles with an initial loss in the range of [SL stat − 0.05; SL stat + 0.05] were considered.
Following this procedure, the velocity detection limit was determined for the full-brain Cartesian time-lapse protocol, for radial FS, US and CS images, and for the adjusted Cartesian sequence, respectively.See Section 2.3 for details on the respective sequence parameters.

In vivo measurements
Animal experiments were carried out according to local animal welfare guidelines and were approved by local authorities (ID: 81-02.04.2020.A194).Female BALB/c mice (n = 3) were obtained from Charles River Laboratories (Sulzfeld, Germany) and housed under a 12 h light-dark cycle and provided with food and water ad libitum.Cells were labeled in vivo by i.v.injection of 3 mL per kg body weight Ferucarbotran (Resovist, Bayer AG, Leverkusen, Germany; 0.5 mmol Fe/mL) via the tail vein.
In vivo time-lapse MRI of the brain was performed 24 h after labeling to ensure clearing of free iron particles from the blood stream. 8Mice were anesthetized with 1%-1.5% isoflurane in 1 L per minute of oxygen and compressed air (20:80) under continuous respiratory and temperature monitoring.Mice were kept at physiologic temperature by a custom-designed animal heating device.Pronounced reduction of body temperature or breathing frequency despite lowering of anesthetic dose were stop criteria for the measurements.Individual cells were identified as hypointense spots and manually tagged using FIJI 22 by two independent observers to reduce observer-dependent bias.Coordinates were linked to trajectories with an adapted cell tracking tool that uses the corresponding part of the MATLAB (The Mathworks, Inc., Natick, MA, USA) implementation by D. Blair and E. Dufresne 23 of the IDL particle tracking code of J. Crocker, D. Grier, E. Weeks. 24,25To differentiate detected cells based on their patrolling behavior, they were categorized into short-term (observed in one or two consecutive timeframes), long-term short-range (observed in three or more consecutive timeframes) and long-term long-range cells (observed in three or more consecutive timeframes with an in-plane motion of more than 1.5 pixels). 12,13To avoid mistaking small vessels for cells, stationary features visible in all timeframes were not counted.
In both, phantoms and in vivo, an RF-spoiled radial FLASH sequence 26 with interleaved ordering scheme was used for radial sampling with scan parameters of TE/TR: 11/400 ms, FA: 15 • , averages: 4, acquisition bandwidth: 50 000 Hz, 195 spokes, interleaved factor: 5, 256 sampling points along each spoke, in-plane resolution: 59 × 59 μm 2 , 15 contiguous slices, slice thickness: 300 μm, scan time per timeframe: 5 min 12 s, 10 repetitions.TR was reduced to shorten acquisition time, limiting number of slices and consequently full-brain coverage.Yet, the acquired 15 slices comprised most of the brain.
Images acquired with radial sampling were reconstructed using inverse nonuniform fast Fourier transform operation.Although with 195 spokes the Nyquist criterium 16 of 256*π/2 ≈ 402 was not met, reconstruction using all of the acquired spokes is nevertheless referred to as FS in the following.Alternatively, due to the interleaved acquisition mode, all spokes can be divided into five sequential subsets of 39 spokes each covering k-space uniformly.Hence, five subframes of 1 min temporal resolution were reconstructed, referred to as US.For additional CS, the corresponding part of the MATLAB implemented iGRASP reconstruction by Feng et al. 27 was used with a weighting parameter λ of 0.1 for in vivo and 0.05 for phantom experiments.
To compare radial and Cartesian sampling, the Cartesian gradient echo sequence was adapted to match the scan parameters of the radial sequence regarding the number of slices, TE and TR (adjusted Cartesian protocol).

Image analysis and statistics
SNR was measured as mean signal intensity in regions of interest (ROIs) divided by the standard deviation of noise.
For n = 3 measurements each in phantoms or in vivo, representative slices were chosen and ROIs were placed in the mouse brain cortex or the upper part of the phantom and a mean SNR was calculated.Contrast-to-noise ratio (CNR) was measured as the difference between the minimum signal intensity of a hypointense spot caused by iron particles and the mean signal intensity in the enclosing area divided by the standard deviation of the signal intensity in the enclosing area.To account for the Rayleigh distribution of the background noise in magnitude images, noise was multiplied by the correction factor √ 2∕(4 − ) ≈ 1.53. 28,29o further compare FS with the accelerated acquisition, that is, US and CS images, all particles were divided into three groups based on their visibility in the rotating phantom: those that were visible in FS images, those that were only visible in US and/or CS images, and those that were not detectable anymore during rotation.For all particles, initial signal loss and velocity were determined.Using the MATLAB-implemented "classify" function with quadratic discriminant analysis, boundaries regarding their velocity and initial signal loss were extracted to separate the three classes.
To assess differences in performance of the three reconstruction modes, only particles that were visible in all three reconstructions were analyzed.First, temporal blurring was evaluated by quantifying signal loss and void size of each particle in the rotating phantom for FS, US, and CS reconstructions, respectively.The signal loss was calculated as described above (Figure S1A-C).To quantify the void size, briefly, an image threshold was set as the average of minimum value and background intensity, and the number of pixels below that threshold was counted (Figure S1D,E).Second, the reconstructed images were compared pairwise by calculating a relative change in signal loss for all particles i defined as the difference in signal loss between the two reconstruction modes divided by the average of the signal loss: relative change SL i (M1,M2) = (SL i (M1)-SL i (M2))/((SL i (M1)+SL i (M2)/2), where M1 and M2 were FS and US, FS and CS, and US and CS, respectively.
Results are shown as means and standard error of the mean.

Phantom experiments
To study motion at very low velocities, a rotating phantom system was built that enabled rotation of phantoms inside a small animal scanner at constant angular velocities of as little as 1.46 × 10 −3 rpm.Such slow motion was achieved by using a stepper motor with attached planetary gear.In agar gel phantoms with embedded MPIOs, the low angular velocity translated to particle speeds close to the estimated detection limit of time-lapse MRI around 60 μm/min.The motor was controlled using the supplied software providing simple handling like starting, stopping and setting the desired speed.Constant rotational speed was surveilled by the encoder and was not affected by the stray field of the MRI magnet (Figure S2).
Using the full-brain Cartesian time-lapse MRI protocol, measurements of a rotating phantom (Figure 2B) were in good agreement with previously performed simulations. 12,13Slowly moving particles close to the axis of rotation remained visible, unrelated to the initially Rotating phantom measurements with the full-brain Cartesian time-lapse MRI sequence.(A) An agarose phantom with embedded micron-sized iron particles (MPIOs) was scanned in a static position as a reference, and then (B) rotated (here at 4.39 × 10 −3 rpm) during the acquisition using the rotating phantom system.Exemplary slices are shown.Particles close to the axis of rotation (blue circle) remained visible independent of the generated contrast in the static phantom.For faster-moving particles toward the edge of the phantom, the visibility of single particles was dependent on the initial signal loss and the velocity: Low contrast features disappeared (red circles), while high contrast features became increasingly blurred (yellow circles).The shape of elongated particles was furthermore dependent on the relative direction of movement and phase (PE) and frequency encoding (FE).(C) For a given initial signal loss SL stat (here: [0.25; 0.35]), a maximum detectable velocity v max (SL stat ) of moving particles (green dashed line) was derived as the intersection of linear fits of the velocity-dependent change in signal loss for visible (black) and non-visible (red) MPIOs.Individual data points represent single particles.(D) This calculation was performed for varying initial signal losses, each resulting in one point.Here, v max increased linearly with the initial signal loss.generated contrast.Contrarily, for faster-moving particles toward the edge of the phantom, visibility varied not only with velocity but also depending on the signal loss in the static phantom, which is affected by the particle size and its position within the voxel.While particles with low contrast in the reference image (Figure 2A) disappeared, high-contrast features became increasingly blurred, resembling simulated time-lapse contrast in shape and size.Dependent on the relative direction of phase encoding and particle movement, elongations along these axes occurred resulting, in the extreme cases, in a narrow line if they matched, or in a square-like shape if they were perpendicular (Figure S3).Moreover, for all detected particles with a given initial signal loss, two classes were distinguished: those that vanished (non-visible), and those that remained visible for which a linear dependence of the difference in signal loss and the velocity was observed (Figure 2C).The intersection of the respective linear fits for the two classes revealed maximum detectable velocities between 119 and 172 μm/min dependent on the initial signal loss (Figure 2D).
Next, the capability of a radial sampling scheme to detect single iron particles was assessed in phantoms using FS, US, and CS reconstruction.In stationary Exemplary images of rotating phantom measurements using an interleaved radial FLASH sequence.phantoms, a mean SNR of 21.9 ± 0.7 was achieved in FS images (Figure 3A).Single MPIOs generated signal voids of a few pixels comparable to those of iron-labeled cells using the Cartesian time-lapse MRI protocol.A total of 567 detected particles had a mean signal loss of 0.303 ± 0.004 and a mean CNR of 5.89 ± 0.13.In US reconstructions (Figure 3B), the mean SNR decreased to 9.8 ± 0.3.Observed streaking artifacts could be tolerated and single particles were resolved with a mean signal drop of 0.316 ± 0.004 and a mean CNR of 4.19 ± 0.07.CS (Figure 3C) reduced streaking artifacts visibly and improved SNR to 14.2 ± 3.5.Mean signal loss and mean CNR of single iron particles were 0.244 ± 0.004 and 4.36 ± 0.12, respectively.
Under rotating conditions, similar observations regarding temporal blurring and the dependence of detectability on initial signal loss and velocity were made as for Cartesian sampling (Figure 3D-F, Video S1).However, the shape of blurred particles was independent of the movement direction since in radial sampling no distinct phase and frequency encoding directions exist.In US images, motion distortion of fast-moving particles decreased visibly (blue arrowheads in Figure 3).Compared to FS, additional fast-moving particles were recovered due to the higher temporal resolution (green rectangle in Figure 3E).Although in CS-reconstructed images temporal blurring was more pronounced than for US reconstruction, single particles appeared sharper, thus smaller and with a more pronounced signal loss, than in FS images (blue arrowheads in Figure 3F).
Quantification of particle sharpness by calculating signal loss and void size confirmed this observation (Figure 4).To reduce effects of variable velocities, 39 particles with similar speed (∼60 μm/min) were assessed.While in the FS images, hypointensities were relatively big ([13 ± 1] pixel) with a low signal loss (0.11 ± 0.01), in US the size decreased ([4 ± 0] pixel) and signal loss increased (0.28 ± 0.01) following conservation of signal.In CS reconstructions, both size and signal loss, were between those of FS and US images ([7 ± 0] pixel, signal loss = 0.16 ± 0.01).
Moreover, out of 567 detected particles in the static phantom, 238 (42%) particles were not visible in the rotating phantom regardless of the reconstruction method.249 (44%) particles were detected only in US and/or CS, but not Comparison of void size and signal loss for fully sampled (FS), undersampled (US), and compressed sensing (CS) in rotating phantom measurements using an interleaved radial FLASH sequence.For the three image reconstruction methods, FS, US, and CS, the void size of 39 particles with similar velocity (∼60 μm/min) was calculated.In each group, individual data points represent single particles with the intensity indicating the signal loss.Horizontal bars are group means.
in FS images.80 (14%) particles were observed in the FS images during rotation, thus accelerated acquisition was not needed to detect these particles.Figure 5 shows these three groups based on their velocity and their initial signal loss in the static phantom.Classification revealed their properties: While the non-visible particles (blue dots) were mostly of low contrast (initial signal loss <0.25) and high velocity (>200 μm/min), the additionally detected particles using US and/or CS (green dots) were of high velocity as well but had a higher initial signal loss (>0.25).Particles visible in FS images (red dots) were mostly slow-moving particles (v < 200 μm/min).
Nevertheless, especially fast-moving particles still benefitted from the higher temporal resolution of US and/or CS.This was demonstrated by comparing the relative change in contrast (i.e., signal loss) between FS and US, FS and CS, and US and CS, respectively (Figure 6).While some slow-moving cells with low initial contrast exhibited reduced signal loss in US images, particles with v > 60 μm/min showed a positive relative change in signal loss between US and FS, indicating that the signal loss was higher in US images regardless of the initial contrast.Similar observations were made when comparing the signal loss between CS and FS images.However, here the relative change was of lower magnitude than for US.Comparing US with CS images revealed that fast-moving particles did not profit from additional CS regarding signal loss.Yet, slowly moving low initial contrast particles showed higher signal loss in CS than in purely US images.

F I G U R E 5
Classification of detected micron-sized iron particles (MPIOs) in a rotating phantom using an interleaved radial FLASH sequence.All detected particles were plotted dependent on their initial signal loss in the static phantom and their velocity, and grouped based on their appearance in the rotating phantom.In red, particles are shown that were visible in the fully sampled (FS) reconstruction irrespective of whether they were detectable in the accelerated reconstructions.Blue represents particles that disappeared in the rotating phantom regardless of the reconstruction.Marked in green are particles that were not visible in FS images, but were recaptured in the accelerated images (undersampled [US] and/or compressed sensing [CS]).Using the classify-function implemented in MATLAB, boundaries of the three groups were derived (solid lines).Finally, the velocity detection limit as function of initial signal loss was calculated for radial (FS, US, and CS) and Cartesian sampling (Figure 7).For FS radial sampling, the detection limit was 165-287 μm/min, which increased for CS to 307 to 611 μm/mins, and to 625-1076 μm/min for US.The fastest particle detectable in FS images had a speed of 287 μm/min, the fastest in US of 1.17 mm/min, and 641 μm/min in CS images.Cartesian sampling with adjusted scan parameters achieved a v max of 290-425 μm/min.

In vivo experiments
Since radial sampling was successful in detecting single static and moving MPIOs in phantoms, the scheme was applied to time-lapse MRI in mouse brains.In all three reconstructions, FS, US, and CS, individual iron-labeled cells were resolved and different motion behavior of detected cells was observed (Figure 8, Video S2).FS reconstruction yielded images with a mean SNR of 17.5 ± 1.5.Although minor streaking artifacts occurred, a total of 41 ± 4 cells per mouse brain were detected.

F I G U R E 6
Velocity-dependent comparison of signal loss for fully sampled (FS), undersampled (US), and compressed sensing (CS) in rotating phantom measurements using an interleaved radial FLASH sequence.FS images were compared to US (left) and to CS reconstruction (middle).In addition, a comparison of CS and US images was performed (right).For each comparison, the relative change in signal loss was calculated as the difference in signal loss between the two images, divided by the average signal loss of the two methods.Individual data points represent single particles, and their intensity is given by the signal loss in the reference static phantom.Especially high-velocity particles (v > 60 μm/min) profited from the acceleration in US and CS images indicated by the positive relative change.While the signal loss was more pronounced in US than in CS images for these particles, signal loss of slowly moving particles with low initial contrast was improved using CS.

F I G U R E 7
Comparison of the velocity detection limit of Cartesian and radial time-lapse MRI.Using the rotating phantom system, the maximum detectable velocity of the different time-lapse sequences and reconstruction methods was derived for different given initial values of signal loss.While the full-brain Cartesian time-lapse sequence (purple) had a slightly lower detection limit than the fully sampled (FS) radial sequence (blue), the adjusted Cartesian time-lapse MRI protocol with matching scan parameters (black) showed an increased detection limit compared to radial sampling.However, undersampled (US) (green) and compressed sensing (CS) reconstructions (red) yielded velocity detection limits of up to 1.1 mm/min, indicating the advantage of simultaneous acquisition of FS and accelerated images possible in radial sampling.
In US reconstructions without CS, SNR decreased to 8.0 ± 0.3, and fewer cells were counted (6 ± 2 in total, 5 ± 2 short-term (3 ± 2 in a single subframe), 1 ± 0 long-term short-range, 0 ± 0 long-term long-range) compared to the FS images.Severe streaking artifacts of varying intensity and with changing direction over time were formed from high intensity features like the skull edges (Video S2).These as well as a higher level of noise hindered detection of especially small low-contrast cells (Figure S4A,C,  Video S2).

F I G U R E 8
In vivo time-lapse MRI using an interleaved radial acquisition scheme.(A) Image details (indicated by the red rectangle) show consecutive timeframes from fully sampled (FS) images with an example of a single iron-labeled cell (red arrowheads) moving across several voxels.(B) Different motion behavior of detected cells was observed.Quantification revealed that in undersampled (US) images fewer cells than in FS images were detected, while in compressed sensing (CS) images most cells were recaptured.Individual data points represent individual animals, and horizontal lines group means.Moreover, 7 ± 1 additional cells were detected (Figure 9), which were clearly discernible in at least one out of the five subframes after accelerated CS reconstruction, but showed no obvious cell features in the FS images.While some were also visible in US images (Figure S4B), others were only detected in CS (Figure S4C).Furthermore, single cell tracking at a higher temporal resolution was enabled.Consequently, for some cells that seemed static in the FS images, distinct movement in the corresponding subframes of the CS reconstruction was observed (Video S3).

DISCUSSION
In this work, a radial FLASH sequence was successfully implemented for single cell time-lapse MRI.Owing to the interleaved acquisition, a retrospective reconstruction of FS, US, and CS images was possible.Velocity detection limits of the various reconstructions and Cartesian sampling time-lapse MRI were measured using a rotating phantom system that enables rotation of phantoms at very low angular velocities.
Although it was the original purpose, the systems applications are not limited to single cell tracking experiments.It also offers the possibility to place and rotate other phantoms and can be used wherever it is necessary to imitate slow movement in other MRI fields.
The main challenge to mimic motion of circulating immune cells was, in addition to an MR-compatible construction, to achieve particle velocities around the estimated detection limit of time-lapse MRI of 60 μm/min.To deal with the required ultra-slow rotation, a stepper motor with planetary gear was chosen.While the lowest continuous rotational speed is 1.46 × 10 −3 rpm, even slower movement is feasible by alternating single steps of the motor and waiting periods.In case of long acquisition times, this non-continuous motion would approximate constant angular speed.However, commercially available motor supplies that avoid interference with the magnetic field of the scanner are limited.Hence, the motor was placed away from the bore center where an acceptable small remaining stray field is present, preventing sustained disturbance or damage.Using an aluminum structure, the distance of the motor to the bore center was set to be as large as necessary, but as short as possible to keep the system relatively small and practical.To address the compatibility issues for the rest of the system, MR-safe components, mostly made from plastic, like the toothed belts or the sample holder, and (acrylic) glass, like the MRI insert, the ball bearings and the drive shaft, were selected.A drawback of the system is that, due to the torsion backlash, only rotation in one direction is possible.Thus, for example, oscillating movement of single cells inside one voxel cannot be imitated and its effects on time-lapse contrast or other imaging-derived parameters cannot be studied.Moreover, unwanted translational movement of the phantom cannot be ruled out completely, and the axis of rotation might not lie exactly in the center axis of the phantom, which might lead to through-plane movement.This non-perfect rotation might result in changes in the visibility of single MPIOs.Despite these possible obstacles, using the system to rotate MPIO-containing agar gel phantoms, motion-dependent blurring of single iron particles was successfully mimicked.Shape and size resembled previously performed simulations on time-lapse contrast.These, as well as in vivo measurements, suggested that patrolling monocytes (∼12 μm/min) could be Additionally detected cells in in vivo single cell tracking by time-lapse MRI using accelerated compressed sensing (CS) reconstruction.Indicated by the red rectangle in exemplary slices at different positions in the mouse brain (1st column) image details of fully sampled (FS) images (second column) and the corresponding accelerated CS subframes (third-seventh column) are compared.They demonstrate examples of cells that were only resolved in at least two subframes of the accelerated CS reconstruction, but not (line 1, 2, 4, 5) or hardly (line 3, 6) in the FS images where temporal blurring was too strong.Line plots, marked by lines, through the respective hypointense spots show the loss in signal (eighth column).resolved, while rolling cells (∼2.4 mm/min) are beyond the detectable velocity range and, hence, not visible in time-lapse MRI experiments. 12Using the rotating phantom system, this assumption was corroborated: A velocity detection limit of up to 172 μm/min was measured for the full-brain Cartesian time-lapse protocol.This limit was dependent on the initial signal loss generated when particles did not move, emphasizing the need for highly efficient cell labeling for successful single cell tracking.
With the goal of enlarging the temporal window, an interleaved radial FLASH sequence was applied to time-lapse MRI.Although for radial sampling the k-space periphery is sampled with lower density, both, in phantoms and in vivo, single MPIOs and iron-labeled intravascular patrolling cells were observed as hypointensities.13]30 Various motion behavior, in line with cells patrolling for variable durations before being washed away by the blood stream, was detected: short-term patrolling, long-term patrolling in confined regions (short-range), and patrolling over several voxels in consecutive timeframes (long-range).While it is possible that noise may have been incorrectly identified as a cell, the observation of short-term cells at a single time point suggests that even cells that only moved slowly for the duration of one timeframe were detected.The mean velocity of moving cells matched the velocity of patrolling monocytes as determined in previous time-lapse MRI studies and by intravital microscopy. 3,12imilar as found for Cartesian sampling, 15 the interleaved acquisition scheme allowed for retrospective reconstruction of FS and US images with and without CS.The accelerated image acquisition improved the detection of moving particles and cells notably compared to FS images.Phantom measurements revealed that effects of US and CS reconstruction were different depending on the particles' motion and contrast characteristics.Slow-moving particles were in general detectable in FS images and thus accelerated acquisition did not improve the particles' detectability.In fact, especially for those with low initial contrast, the signal loss even decreased in US and CS images compared to FS.This can be attributed to increased noise and to the fact that low contrast features are likely to be removed by CS.High-velocity particles did profit from US and CS and were more clearly visible.Temporal blurring was reduced and signal loss increased.While elongated shapes due to blurring do not change the particles' detectability in phantoms, in mouse brains, these distortions may be problematic, since cells can be easily mistaken for other brain structures such as vessels.Furthermore, especially in US images, even additional particles, previously hidden in FS images, were recaptured.In phantom experiments, US reconstruction yielded the best results regarding improvement in contrast of moving particles and the velocity detection limit.Streaking artifacts occurred but could be tolerated and did not significantly hinder particle detection.However, in in vivo time-lapse MRI streaking artifacts were too severe to detect most cells in US images.Only very high-contrast cells remained discernible.Contrarily, in CS reconstructed images even low-contrast cells were recovered due to the better image quality and again additional cells, which were not visible in FS images, were captured owing to the higher temporal resolution.Cells not detected in FS images most likely changed their velocity during their in-plane movement by either starting rolling or being washed away by the blood stream.Consequently, they did not have sufficiently low average speed during the full timeframe, but rather only during few subframes.Thus, to best capture low-contrast slow-moving and fast-moving cells, analysis of FS and CS reconstructions should be combined.
Comparing in vitro and in vivo results, single iron particles and iron-labeled cells had a similar appearance and, in both cases, showed varying contrast due to their different positions within the slice.However, noise levels were notably different, and owing to missing anatomical structures, streaking artifacts were less pronounced in phantoms than in vivo.As a consequence, US performed best in phantoms regarding temporal blurring and the maximum detectable velocity, and there were no particles only visible in FS.The need for CS became obvious in vivo where only high contrast cells were visible in pure US images.Additionally, while elongated shapes were observed in phantom images, this form of temporal blurring could not be seen in vivo.Thus, the derived velocity detection limits in vitro may not necessarily be equal, but are most likely to be higher than in vivo.Overall, the velocity detection limit of time-lapse MRI was improved using the radial acquisition scheme, especially in combination with US and CS.However, to potentially capture rolling monocytes as well, even shorter scan times are needed.Thus, optimal k-space sampling will be investigated further for example by exploiting a golden angle scheme. 27Although the applied interleaved scheme allows already for simultaneous acquisition of FS and US data, golden angle sampling provides even higher flexibility regarding undersampling ratios. 27Further, the minimal number of spokes needed will be assessed in future work.

CONCLUSIONS
Radial sampling in time-lapse MRI was successfully applied to single cell tracking experiments.Both in phantoms and in vivo, MPIOs and iron-labeled cells can be followed dynamically.Without additional effort, both FS and accelerated data can be acquired simultaneously expanding the temporal window, reducing motion-dependent distortions and improving cell detection.The constructed rotating phantom system allowed to measure the velocity detection limit of time-lapse MRI, which was pushed to 1.1 mm/min by radial sampling.This approach is a valid method to investigate detection boundaries of single cell tracking and possible acceleration techniques, but not limited to that purpose.First, to avoid wrong void size estimates resulting from the inhomogeneous intensity profile of the cryo coil, the background was removed using the "subtract background" function in ImageJ (rolling ball radius = 60, sliding paraboloid, no smoothing) (D).As indicated by the red rectangle, image details (E, first row) of the single particles were extracted.Then, the minimum signal intensity of the hypointense spot and the mean intensity of the enclosing area was determined as described in (A-C), respectively for the three reconstruction methods.Finally, the number of pixels below the threshold of (mean(signal intensity enclosing area ) + min(signal intensity hypointense spot ))/2 was counted and defined as the void size (red area in E, second row).Figure S2.Time-dependency of the motor rotation in-and outside the scanner room.In order to assess a possible impact of the MRI stray field on the motor rotation, the actual velocity value of the motor was extracted from the optical encoder every 1000 ms over a total time of 8 min 28 s, thus covering at least one FS timeframe of time-lapse MRI.This measurement was performed for target velocities of the motor without gear (red dotted lines) of 3 and 12 rpm once outside the scanner room (black solid lines) and once with the system installed at the scanner (green solid lines), respectively.As expected fluctuations of the motor velocity around the demand value could be observed.While these were of greater extent for the higher demand value, in both cases, for a set velocity of 3 and of 12 rpm, these variations were negligible.Moreover, there was no obvious effect of the stray field on the rotational velocity observable.Figure S3.Comparison of time-lapse contrast in simulations and phantom measurements for Cartesian sampling.(A-D) Temporal blurring of time-lapse MRI contrast was simulated based on artificial k-space as previously described. 12,13Briefly, a synthetic phantom (intensity = 1) was created and artificial cells were added as signal voids 4 pixel in size (intensity in the central voxel = 0.5, neighboring voxels = 0.7) comparable to signal loss of iron-labeled cells observed in vivo (A).Cell motion was simulated by stepwise increasing the signal voids position and Rician noise with a level of 0.02 was added to the individual images.Synthetic k-space was then assembled from fractions of position-specific k-space data, and image reconstruction was performed by Fourier transformation as previously described. 12,13(E, F) In phantom measurements, agarose phantoms containing MPIOs were rotated using the rotating phantom system.Image details, acquired with the full-brain Cartesian time-lapse MRI sequence, showing an exemplary particle, static and moving at different time-points, are shown.Overall, shape and size of simulated cells and particles in the phantom were in good agreement.In the static case (A, E), the artificial cell and the iron particle was clearly visible as an hypointense spot.When the cell/particle was in motion with a velocity of ∼60 μm/min, temporal blurring occurred and different shapes were observed depending on the relative direction of movement (indicated by the red arrows) and phase encoding (PE, black arrow), which was either parallel (B, F), in a 45 • angle (C, G) or perpendicular (D, H).

ACKNOWLEDGMENTS
(A-C) In static agarose phantoms, single micron-sized iron particles (MPIOs) could be resolved as hypointense spots using a radial sampling scheme in time-lapse MRI for (A) fully sampled (FS), (B) undersampled (US), and (C) compressed sensing (CS) reconstruction.(D-F) Once the phantom rotated during the acquisition (here with a rotational speed of 8.79 × 10 −3 rpm), particles became increasingly blurred (blue arrowheads) or faded (green rectangles, red circles in A, B, C, E) in the FS images (F).(E) In US reconstructions, some particles were recovered (green rectangle), while for others motion distortion decreased (blue arrowheads) due to the higher temporal resolution.(C, F) Using CS, streaking artifacts were reduced.(F) Although temporal blurring of particles in the rotating phantom was more pronounced than in US images, hypointense spots resulting from MPIOs appeared nevertheless sharper than in FS images and had a higher signal loss (blue arrowheads).Red arrows indicate how far the phantom rotated during the respective timeframe.

Figure S4 .
Recovery of cells in in vivo single cell tracking by time-lapse MRI by accelerated CS reconstruction in comparison to US images.Indicated by the red rectangle in exemplary slices at different positions in the mouse brain (1st column), image details of FS images (2nd column) and the corresponding accelerated US (1st row of 3rd-7th column, respectively) and CS (2nd row of 3rd-7th column, respectively) subframes are compared.(A) shows an example of a cell that can be seen in the FS image, but not in the US reconstruction.Using CS, the feature was recovered indicating the need for CS in in vivo radial time-lapse imaging.(B) depicts an additionally detected cell in the accelerated reconstruction.While the high contrast cell was not visible in the FS image, it was discernible in US and CS reconstruction.(C) demonstrates a low-contrast cell that was not visible in FS images, nor in the US reconstruction.However, it was detectable using CS.Video S1.Rotating phantom measurements using an interleaved radial acquisition scheme with time-lapse MRI.Five consecutive full timeframes and the corresponding 25 subframes of the accelerated reconstructions are shown.The rotational speed was 8.79 × 10 −3 rpm.In the FS images (left), temporal blurring of moving particles was observed.In the US reconstruction (middle), motion distortion decreased, however streaking artifacts occurred.CS (right) removed these, increasing the SNR.While particles still had an elongated shape due to blurring, signal loss was stronger than in FS images, improving single particle detection especially for the fast-moving ones.Video S2.Example of in vivo time-lapse MRI using an interleaved radial acquisition scheme.Time-lapse MRI video of 15 slices with 10 FS timeframes each are shown.An acquisition time of 5 min 12 s per timeframe resulted in a total scan time of 52 min.In all three reconstructions, FS, US and CS, individual single cells were detected as hypointense spots and could be followed dynamically.Cells were further classified into short-term (red arrowhead), long-term short-range (blue arrowhead) and long-term long-range (green arrowhead) motion patterns.Video S3.Improvement of the temporal resolution in in vivo time-lapse MRI using an interleaved radial acquisition scheme and CS reconstruction.Image details, indicated by the red rectangle, show an example of a cell that seemed static in the FS but moves in the accelerated CS reconstruction owing to the higher temporal resolution.Three consecutive full timeframes and their corresponding 15 subframes are shown in which motion of the hypointense spot can be observed.It had a velocity of 32 μm/min.Note that these images were acquired with different scan parameters (TE/TR: 10/300 ms, 265 spokes).How to cite this article: Wilken E, HavlasA, Masthoff M, Moussavi A, Boretius S, Faber C. Radial compressed sensing imaging improves the velocity detection limit of single cell tracking time-lapse MRI.Magn Reson Med.2024;91:1449-1463.doi: 10.1002/mrm.29946 The signal loss of the void was calculated as SL = [mean(signal intensity enclosing area )-min(signal intensity hypointense spot )]/mean(signal intensity enclosing area ) with the ROI of the hypointensity being the yellow circle and the enclosing area the shaded area bordered with the yellow square.(D-E) Size calculation of hypointensities resulting from individual iron particles.An exemplary slice of the agarose phantom acquired with interleaved radial time-lapse sequence is shown.