Available via license: CC BY 4.0
Content may be subject to copyright.
1
Repeatability of image quality in very low field MRI
Pavan Poojar1, Kunal Aggarwal1,
Marina Manso Jimeno2, and Sairam Geethanath1,2*
1Accessible Magnetic Resonance Laboratory, Biomedical Imaging and Engineering Institute,
Department of Diagnostic, Molecular and Interventional Radiology, Icahn School of Medicine at
Mount Sinai, New York, NY, United States
2Columbia Magnetic Resonance Research Center, Columbia University, New York, NY, United
States
Correspondence to:
Sairam Geethanath, Ph.D.
Accessible Magnetic Resonance Laboratory, Biomedical Imaging, and Engineering Institute,
Department of Diagnostic, Molecular, and Interventional Radiology
Icahn School of Medicine at Mt. Sinai, New York, NY, United States
E-mail: sairam.geethanath@mssm.edu
Phone: 717-590-0997
2
Abstract
Background: Low-field magnetic resonance (MR) has emerged as a promising alternative to
high-field MRI scanners, offering several advantages. One of the key benefits is that low-field
scanners are generally more portable and affordable to purchase and maintain, making them an
attractive option for medical facilities looking to reduce costs. Low-field MRI systems also have
lower radiofrequency (RF) power deposition, making them safer and less likely to cause tissue
heating or other safety concerns. They are also simpler to maintain, as they do not require cooling
agents such as liquid helium. However, these portable MR scanners are impacted by
temperature, lower magnetic field strength, and inhomogeneity resulting in images with lower
signal-to-noise ratio and geometric distortions. It is essential to investigate and tabulate the
variations in these parameters to establish bounds so that subsequent in vivo studies and
deployment of these portable systems can be well-informed.
Purpose: To investigate the repeatability of image quality metrics such as SNR, image uniformity,
and geometrical distortion at 0.05T over ten days and three sessions per day
Methods: We acquired repeatability data over ten days with three sessions per day. The
measurements included temperature, humidity, transmit frequency, off-resonance maps, and 3D turbo
spin echo (TSE) images of an in vitro phantom. This resulted in a protocol with nine pulse sequences.
We also acquired a 3T data set for reference. The image quality metrics included computing SNR,
image non-uniformity, and eccentricity (to assess geometrical distortion) to investigate the repeatability
of 0.05T image quality. The image reconstruction included drift correction, k-space filtering, and off-
resonance correction. We computed the coefficient of variation (CV) of the experimental parameters
and the resulting image quality metrics to assess repeatability.
Results: The range of temperature measured during the study was within 1.50C. The off-resonance
maps acquired before and after the 3D TSE showed similar hotspots and changed mainly by a
global constant. The SNR measurements were highly repeatable across sessions and over the ten
days, quantified by a CV of 4.9%. The magnetic field inhomogeneity effects quantified by eccentricity
showed a CV of 13.7% but less than 5.1% in two of the three sessions over ten days. The use of
conjugate phase reconstruction mitigated geometrical distortion artifacts. The repeatability of image
uniformity was moderate at 10.6%, with two of three sessions resulting in a CV of less than 7.8%.
Temperature and humidity did not significantly affect SNR and mean frequency drift within the
ranges of these environmental factors investigated.
3
Conclusions: We found that humidity and temperature in the range investigated did not impact
SNR and frequency. Based on the coefficient of variation values computed session-wise and for
the overall study, our findings indicate high repeatability for SNR and magnetic field
homogeneity; and moderate repeatability for image uniformity.
Introduction
Magnetic Resonance Imaging (MRI) is a life-saving technology widely used to investigate the
human brain. However, two-thirds of the world's population, particularly those in low-resource
settings, lack access to MRI due to the complex requirements of high-field systems, such as
infrastructure, engineering, and electrical power 1.
Very low-field MR (<0.1T) has emerged as a promising alternative to high-field MRI scanners,
offering several advantages 2–4. One of the key benefits is that low-field scanners are generally
more portable and affordable to purchase and maintain, making them an attractive option for
medical facilities looking to reduce costs. Low-field MRI systems also have lower radiofrequency
(RF) power deposition, making them safer and less likely to cause tissue heating or other safety
concerns 5. They are also simpler to maintain, as they do not require cooling agents such as liquid
helium. Therefore, low-field MRI scanners offer a more portable, accessible, and cost-effective
means of accessing vital medical imaging services, particularly in low-resource settings, and have
the potential to revolutionize healthcare delivery.
Yip and Konova 6 showed that more frequent and focused neuroimaging scans could allow
monitoring brain changes more effectively over time, which can lead to the development of
personalized treatment plans for patients. Low-field MRI can address this need for a dense
temporal sampling of neuroimaging data. Therefore, continued research and development in this
area could significantly impact the accessibility and quality of healthcare globally. However,
4
challenges remain in integrating low-field MRI into clinical practice and ensuring their widespread
adoption in both high and low-resource settings1. Low-field MRI systems also have a lower signal-
to-noise ratio (SNR), which can affect image quality and accuracy. Low spatial resolution is
another disadvantage of low-field MRI systems limiting their ability to detect small anatomical
details such as blood vessels or small lesions. These limitations result in longer scan times,
increasing the risk of patient motion during the scan and reducing image quality. Importantly, low-
field MRI systems are susceptible to environmental factors such as temperature, humidity, and
electromagnetic interference (EMI) and system factors such as main magnetic field
inhomogeneity and gradient heating 3,7–9, which cause image artifacts and degrade image quality.
Repeatable and consistent imaging is critical for diagnosing diseases or monitoring therapy
progression 9–14. This is more relevant in the context of these low-field scanners as inconsistencies
caused in scanner operation and image quality are due to environmental and system factors
changes, especially in permanent magnet constructions. However, the repeatability of low-field
MR imaging has not yet been investigated.
In this work, we performed a repeatability study on an in vitro phantom (ProMRI, ProProject, USA)
at 0.05T using 3D turbo spin echo (TSE) imaging over ten days and three sessions daily (30 scan
sessions). We investigated the precision of very low field MRI (0.05T) measurement quantified by
SNR, image uniformity, and geometrical distortion and; investigated the effect of temperature and
humidity on image quality. This data would help determine the reliability of the 0.05T scanner and
establish bounds of performance that can be used to monitor stable operation. We did not
consider the repeatability of contrast (T1w, T2w) as that is more optimally investigated in healthy
human volunteers.
5
Materials and Methods
Overview of the study: We performed the repeatability study using the in-vitro phantom (Pro-
MRI by Pro Lab) on a 0.05T scanner (MultiWave Technologies, France). The scanner and
hardware specifications are described in detail in Ref. 8 Three scanning sessions were conducted
daily: Session 1 began at 11 AM, Session 2 at 2 PM, and Session 3 at 5 PM local time. Each
session lasted approximately forty-five minutes. We measured temperature (°C) and humidity
(%RH) before and after each Session at three locations. Figure 1 presents a schematic diagram
outlining the details of the study, including the Session and time information. Supplementary
Figure 1 shows a picture of the 0.05T scanner and the locations where temperature and humidity
were measured using a portable digital thermo-hygrometer. The three locations were: in front of
the phantom within the bore (location 1), in front of the bore (location 2), and in the center of the
room (location 3), shown with orange, yellow, and green circles, respectively. These
measurements assessed the potential effect of temperature and humidity on image quality. The
device was also compared with a fluoroptic probe (LumaSense, USA) to ensure that the magnetic
field did not affect the device's measurements.
Figure 1. Repeatability protocol. The study was performed over ten days. Three imaging
sessions were performed daily, each containing temperature and humidity measurements before
and after the scan. The imaging protocol measured transmit frequency, off-resonance mapping,
and 3D imaging. The acquisition details are listed in Table 1
Session
1
Session
2
Session
3
Session
1
Session
2
Session
3
11:00 14:00 17:00
Time
Day 1 Day 2 Day 10
Session
11:00 14:00 17:00
Protocol
Temperature
&
Humidity
Temperature
&
Humidity
External
measurements
f0 3D TSE
without shift f0 3D TSE with
shift f0 3D TSE f0 3D TSE
without shift f0 3D TSE with
shift f0
B0 map
External
measurements
B0 map
Repeatability
study . . . Session
1
Session
2
Session
3
11:00 14:00 17:00
6
MR acquisition: The repeatability protocol utilized in this study included nine sequences. Figure
1(a) shows the protocol details during each Session. The imaging pulse sequences included a
3D turbo spin echo (TSE) and two off-resonance (B0) maps (before and after 3D TSE). Two 3D
TSE sequences with an echo shift of 49μs between them were used to generate the B0 map,
followed by the conjugate phase reconstruction (CPR) method to compensate for off-resonance
effects. 15 To find the transmit frequency, a ‘findf0’ sequence (f0 refers to Larmor frequency, using
a pulse and acquire pulse sequence) was employed before and after each sequence, resulting in
six such scans. Before beginning the protocol, we measured the noise (no transmit RF pulse, only
recording noise) using a ‘Monitor noise.’ During the repeatability study, we turned off the scanner
after each Session (except on day 1) but did not move the scanner and the phantom.
Table 1. MRI acquisition parameters at 3T and 0.05T with a spatial resolution of 0.75 x 0.75 x
5 mm3 and 1.5 x 1.5 x 5mm3 respectively.
During each session, we measured f0 before and after each scan to evaluate the effect of
frequency drift on the image. We scanned the phantom on the 0.05T scanner for ten days and on
a Siemens 3T Skyra scanner once. Table 1 lists the acquisition parameters of the protocol at 3T
MRI acquisition parameters
0.05T
3T
Field strength
3D Turbo
-spin-echo
2D Turbo spin echo
Sequence
500
6000
TR (
ms)
20
103
TE (
ms)
4
18
Echo train length (ETL)
40
56
Bandwidth (kHz)
230 x 230 x 125
192 x 192 x 125
Field of view (mm
3)
155 x 155 x 25
256 x 256 x 25
Matrix
8:01
2:02
Scan time (
min:sec)
7
and 0.05T for the 2D multi-slice and the 3D TSE sequences, respectively. The acquisition
parameters for the 3D TSE sequence with and without echo shift were the same as the 3D TSE,
except for the bandwidth (BW), which was 50 kHz. We used a higher BW for B0 mapping to reduce
geometric distortion with shorter readouts at the cost of lower SNR. The total scan time for the
protocol was approximately 45 minutes. We acquired 25 slices with a resolution of 5 mm (125mm
slab thickness) along the third dimension for a phantom that had a height of 90 mm. Hence, we
discarded the first four and last five slices outside the phantom and contained noise.
Phantom and image quality metrics: The Pro-MRI phantom solution contained ten mmol of
Nickel Chloride and 75 mmol of Sodium Chloride. We evaluated the SNR, image uniformity (IU),
and geometric accuracy (GA). These parameters were selected to assess the repeatability of the
scans across different imaging sessions over ten days and involved measurements on different
slices. Supplementary Figure 2(a-c) displays the representative specific slices, without off-
resonance correction, chosen to evaluate SNR, image uniformity, and geometric accuracy in this
study. The Pro-MRI phantom's construction is similar to the American College of Radiology
phantom, a standard phantom used in MRI. We chose slice 10 for SNR measurement because it
lacked apparent fine features or structures. The SNR was computed as the ratio of the mean
signal intensity of the phantom ROI to the standard deviation of the background noise. To evaluate
IU, we chose slice 16 and computed the standard deviation with the phantom region of interest
(ROI, red ROI in Figure 2c; the lesser the standard variation, the more uniform the intensity). For
GA, we chose slice number 8 and calculated the eccentricity of the phantom ROI, which measures
the degree of deviation from a perfect circle. This calculation was performed before and after
applying B0 correction on the 3D TSE data to quantify the benefit of the off-resonance correction.
Image reconstruction: Figure 2(a-c) shows the reconstruction, post-processing, and image
analysis pipelines. The k-space data underwent preprocessing steps, including drift correction
8
and k-space filtering (squared sine-bell). Drift correction was necessary to correct any signal drift
over time, which could lead to image blurring and ghosting, and high-frequency noise was
suppressed using k-space filtering. Next, the image was reconstructed using a fast Fourier
transform (FFT). However, the off-resonance effects distorted the reconstructed 3D TSE images.
B0 correction included the CPR method implemented in our open-source Python toolbox Off-
resonance Correction OPen soUrce Software (OCTOPUS) 15. The results were benchmarked
against the results of the code used in ref. 16. The off-resonance correction process consisted of
four steps. In the first step, image reconstruction from two 3D TSE k-space datasets, with a
difference in echo time of 49us. We calculated the B0 map from these images, including phase
unwrapping, Gaussian smoothing (sigma=21, refer to Supplementary Figure 3) to reduce noise
in the corrected images, and background masking (manually) using the magnitude images.
Figure 2. Image reconstruction, processing, and analysis pipelines. (a) The data were put
through a series of steps to extract and reorder the k-space from the scanner and corrected for
thermal drift and denoising in k-space. (b) The post-processing involved off-resonance correction
and low-pass filtering to yield the images for analysis. (c) Different slices were chosen to test for
signal-to-noise ratio (SNR), image uniformity, and geometrical distortion and were manually
segmented to provide input for computing the relevant image quality metric.
Raw file
k-space
Drift correction
k-space filter
(sine bell square)
3D TSE
without shift
3D TSE with
shift
3D TSE
B0 map
B0 map filtering
3D TSE without
correction
Off resonance correction
Slice 10 for
SNR Slice 16 for
IU Slice 8 for
GA
ROI signal
(s) ROI noise
(n)
ROI inner
circle
Compute
SNR Calculate standard
deviation
3D TSE after
correction
Select slices for image analysis
k-space filtering
3D TSE
Compute
eccentricity Compute
eccentricity
Reconstruction Post-processing Image analysis
Image
9
Finally, we performed the B0 correction using both CPR implementations. We first masked the B0
map to remove the unwanted noise (background). This 3D mask was manually generated from
no-shift 3D TSE data based on thresholding. All 30 datasets of the repeatability study underwent
this process. We used OCTOPUS for off-resonance correction as it was built in-house and is open
source.
Transmit frequency (f0) and difference in off-resonance measurements: We obtained six f0
values for each session by measuring the frequency before and after each sequence. We
graphed a box plot to compare the frequency drift over ten days and across all sessions. We
calculated the mean difference in off-resonance before and after the 3D TSE sequence. We
plotted the mean f0 (average of two consecutive f0 values – before and after an imaging pulse
sequence), temperature, and humidity for all ten days and the three Sessions.
Image analysis: We observed that the slice selected for IU had a ghosting and ringing-like artifact
caused due to the stimulated echos and residual magnetization of the long T2 components that
we verified by varying the echo train length (ETL; refer to Supplementary Figure 4), which we
suppressed by applying k-space filtering. We computed the IU before and after the filtering. We
used the Gaussian filter for k-space filtering and optimized the sigma values, varying from 15 to
25. We chose the optimal sigma value by visually inspecting the difference image obtained by
subtracting the original image from the filtered image. To evaluate the accuracy of the distortion
correction, we calculated the eccentricity of one slice (slice 8) of the phantom before and after
correction using the scikit-image library9. In the case of no geometric distortion, the eccentricity of
a circular slice of the phantom should be equal to 0. We computed the eccentricity of the original
image and two correction methods.
Statistical analysis: To determine the repeatability of experimental and measured parameters,
we computed the Session-wise and overall (all thirty sessions) mean, standard deviation, and
10
coefficient of variation for (i) environmental variables such as the temperature and humidity; MR
acquisition parameters such as transmit frequency; and (iii) image quality parameters quantified
by SNR, image non-uniformity and eccentricity. A scatter plot was generated to investigate the
relationship between (i) temperature and SNR; (ii) humidity and SNR; (iii) temperature and mean
frequency; (iv) humidity and mean frequency. The correlation coefficient was computed for each
plot to quantify this relationship's strength.
Results
Temperature, humidity, and f0 measurements: Figure 3 shows the temperature and humidity
measurements. The temperature ranges for all three locations over the thirty sessions were
primarily within 1.50C. As expected, the temperature after the scan (dashed) was higher than
before (solid), primarily due to gradient heating. The Mean ± SD temperature for all three sessions
at the three locations was 23.34 ± 0.270C, 23.49 ± 1.240C, and 23.44 ± 0.290C, respectively. The
Mean ± SD of humidity for Locations 1, 2, and 3 were 45.48 ±4.4 %RH, 44 ±4.6 %RH, and
47.6±4.6 %RH respectively. The low values of SD for temperature and humidity suggest minimal
variation at each location and across all three locations (10.3% for humidity and 3.2% for
temperature, see Supplementary Table 1 for corresponding location-wise numbers). The
temperature and humidity measurements recorded a sudden decrease on day seven at all three
locations before Session 1, potentially due to the open room door. This allowed the outside
temperature and humidity to influence the measurements. However, the door was closed during
all the remaining sessions, and the measurements after scanning (see S1A) showed no decrease
in measured values. Figure 3(b) shows that Location 2 recorded a higher temperature value on
day 3 (see S3A), which could be attributed to an inaccurate reading at Location 2 (edge of the
bore), where the fringe magnetic field is the highest. The rest of the experimental conditions were
11
identical to those on other days. Generally, temperatures for the third Session > second Session
> first Session are consistent and may be attributed to the room's air conditioning (refer to Figure
4).
Figure 3. Temperature and humidity measurements. (a-c), the temperature (in
℃
) was plotted
for ten days at three locations: the center of the bore (a, Location 1), the periphery of the bore (b,
Location 2), and the center of the room (c, Location 3). (d-f) shows the humidity for the
corresponding three locations. The temperatures and humidity values were measured before the
scan (S1B, red filled circle) and after (S1A, red hollow circle). Similarly, the green-filled circles
represent Session 2 before and after scanning, and the blue-filled circles with a solid line and the
hollow circles with a dashed line represent Session 3 before and after scan measurements.
Figure 4b shows the box plot of frequency values (MHz) for three sessions over ten days. The
plot shows that the frequency was highest during Session 1 and lowest in Session 3, indicating a
drift in the central frequency (f0) over time within the same day. This trend is valid for all ten days
and validates the temperature trend across sessions seen in Figure 3(a-c). A systematic drift in
the central frequency of the MRI scanner can cause distortions and artifacts in the images,
especially if the pulse sequence uses the transmit frequency to compute spatial gradient-
dependent parameters (slab location, for example) and lead to a loss of image quality. The Mean
± SD of the f0 shift for all three sessions was 28.04 ±14.4 Hz. The high value of the SD compared
ab c
d e f
12
to the Mean captures the variation in the difference in thermal drift ranges across the three
sessions seen in Figure 4.
Figure 4. Temperature dependence of transmit frequency in phantoms. a) Plot shows that
the average transmit frequency per session decreases with an increase in temperature, and
session 1 had the least temperature and the highest transmit frequency > session 2 > session 3;
b) the box plot shows the range of transmit frequencies for the six measurements obtained in
each session.
B0 mapping: Figure 5(a,b) shows the off-resonance maps for one slice (slice number 5) for three
Sessions over ten days before and after the 3D TSE acquisition. This slice was chosen as it
contained a fine structure and inserts with different signal intensities (see Supplementary Figure
2). Figure 5(c) shows the difference in off-resonance maps obtained before and after the 3D TSE
acquisition. It can be seen that the maps are in a similar range except for day 1, Session 1 (shown
with black arrow), and day 10, Session 1 (shown with green array). The shape of the off-
resonance map for day 1, Session 1, is different from the rest of the B0 maps because the scanner
was active (turned on) before the start of Session 1 and could be the reason for obtaining a distinct
off-resonance compared to the rest of the maps. The off-resonance was saturated with only
positive values with the mean value of 19.971 kHz, as shown in Figure 5(b) (day ten, Session 1).
One of the possible reasons for this could be electromagnetic interference during the scanning
session. The difference in off-resonance values is close to zero, indicating no significant
difference between the two measurements, which were acquired before and after the 3D TSE
a) b) c)
22.5 23.0 23.5 24.0
2.078
2.079
2.080
2.081
2.082
Temperature (°C)
Frequency (MHz)
Session 1
Session 2
Session 3
0 5 10
0.03
0.04
0.05
0.06
0.07
Day
Image non-uniformity
Session 1
Session 2
Session 3
0 5 10
20
22
24
Day
Signal to noise ratio (a.u.)
Session 1
Session 2
Session 3
0 5 10
0.1
0.2
0.3
0.4
0.5
0.6
Day
Eccentricity
Uncorrected Corrected (OCTOPUS)
d)
b)
13
scan. The changes in off-resonance maps were essentially a global shift rather than a hopping of
the hotspots in off-resonance.
Figure 5. Consistency of magnetic field homogeneity at 0.05T a) Off-resonance maps for the
ten days and three sessions before the 3D turbo spin echo (TSE) scan; b) corresponding off-
resonance maps after the 3D TSE scan that do not show any changes in hotspots compared to
off-resonance maps in a); c) the difference between a) and b) show examples of changes in off-
resonance that do not show a constant difference.
This can be deduced by the presence of similar red and blue regions in Figure 5(a,b) and the
difference depicting significantly either red (positive shift) or blue (negative shift). However,
examples like those shown by the black arrows in Figure 5c do not have a constant global shift.
These examples do not indicate a shift of the off-resonance hotspots (Figure 5b is similar to 5a in
Session 1
Session 2
Session 3
Session 1
Session 2
Session 3
Session 1
Session 2
Session 3
Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 Day 7 Day 8 Day 9 Day 10
-2000
2000
0
Hz
Hz
a
b
c
14
terms of hotspots) but that the difference between the two maps is not a constant. Therefore,
acquiring a B0 map post-acquisition may be relevant for downstream off-resonance correction if
the imaging sequence lasts beyond several minutes. Exploration of the combination of the two
off-resonance maps for optimal CPR correction is beyond the scope of this work. Figure 6(b)
shows the plot for the mean difference in off-resonance maps (Hz) over ten days and all three
sessions. These values were generally lesser than 400Hz compared to the dynamic range of
4000Hz (~10%).
Figure 6. Intra-session variation in frequency shift and off-resonance a) the mean frequency
shift before and after the 3D turbo spin echo (TSE) shows session 1 has the least frequency shift
and b) the mean difference in off-resonance before and after 3D TSE does not show a session-
wise trend. Day 10 Session 1 shows a significantly higher value, as visualized in Figure 5
The B0 maps shown here were masked (removed background) and followed by filtering (Gaussian
filter). Supplementary Figure 5 (a-c) shows the unfiltered off-resonance maps for one slice over
ten days and 3 Sessions. The obtained maps were noisy due to the high BW (50 kHz) acquisition.
Supplementary Figure 6(a,b) show the off-resonance maps (16 slices) for unfiltered and filtered
16 slices from day one and Session 1. The first and last slices show high noise as they are away
from the isocenter coupled with the low SNR acquisition at 50kHz.
a b
15
Off resonance correction: All datasets in the repeatability study showed visibly lesser geometric
distortion using both CPR implementations by qualitative inspection of the images before and
after correction (Supplementary Figure 7). There was no appreciable difference between the two
CPR implementations. The eccentricity calculation showed that the phantom's circularity
increases after correction, indicating a reduction in geometric distortion. We chose to continue
with CPR-OCTOPUS as it is open-source and was home-built.
Figure 7. Image quality comparison a) 3T images from a 2D turbo-spin echo (TSE) sequence
of the Pro-MRI phantom at 0.75 x 0.75 x 5mm3 resolution; b) 3D TSE images at 0.05T with no off-
resonance correction at 1.5 x 1.5 x 5mm3 resolution, note the offset in the slice direction (orange
arrows) and geometrical distortion due to off-resonance; c) off-resonance corrected images
reduce geometrical distortion and correct for the slice shift but suffer from a signal loss in the first
and last slices (red arrows) due to low signal-to-noise ration off-resonance maps (also see
Supplementary Figure 6).
Qualitative images: Figure 7(a) shows the 2D multi-slice TSE images obtained from Siemens
3T Skyra. These images were considered the gold standard and acquired only once. Figure 7(b,c)
shows the 3D TSE images obtained from 0.05T scanner, and off-resonance corrected images
using the CPR-OCTOPUS method. The differences in spatial resolution and geometric distortion
between the gold standard 3T and 0.05T can be observed by visually comparing Figures 7 (a,c).
These image panels show 16 slices from day 1 of Session 1. Slice 4 in all three image panels
with an orange arrow indicates partial volume. The first and last slices in corrected images show
low SNR (shown in red arrow) in corrected images as well as a shift in the slice direction. This
a b c
3T 0.05T –Raw data 0.05T –CPR corrected
16
may be attributed to a phase shift applied by the off-resonance correction in the slice direction.
The first row images from the 0.05T scanner are more distorted than other slices as these slices
are away from the isocenter (see Supplementary Figure 4b).
Image analysis: Figure 8(a) shows the SNR over ten days for three Sessions. The SNR is lowest
for Session 3 compared to Sessions 1 and 2 in 6 out of 10 days. This is in line with the observation
that Session 3 witnessed the highest temperature (Figures 3(a-c)) and lowest Larmor frequency
(Figure 4). From Figure 3, it can be observed that the mean temperature increases as the day
progress. Therefore, we can deduce that as the temperature increases, SNR decreases. SNR is
reduced by thermal noise resulting from the electrical resistance of the imaging coil, which
increases with temperature and consequently reduces the SNR. This effect was also reflected in
the Mean ± SD of SNR for Sessions 1, 2, and 3 over the ten days, which were 22.4±1.2, 22.05±0.1
and 21.6±1 respectively. However, these similar values (within the range of 1 a.u. of SNR)
indicate the repeatability of SNR over the three sessions across ten days. Figure 8(b) shows the
plot for IU of the 3D TSE on one slice over ten days and three Sessions. The plot shows that the
SD of the intensity values of the ROI were within 0.05 for all ten days. This indicates that the
intensity values of the images acquired over ten days using the 0.05T scanner were similar and
did not vary significantly, except for day 1. This may be attributed to the scanner not being turned
off before operation, as we also see a difference in off-resonance maps (eccentricity) and mean
f0 shift for the data from day 1. The IU was computed on the filtered image (Gaussian filter).
Supplementary Figure 3 shows the effect of different sigma values on the image. The difference
image (last column) was generated by subtracting the filtered image from the original image. As
the sigma decreases, the artifact reduces, but the blurring increases. Hence, we chose the sigma
value to be 21 based on visually inspecting the different images.
Figure 8 (c-e) shows the eccentricity plot for all three Sessions for the original (without
correction, red circle) along with the corrected image using two methods (OCTOPUS - blue circle
17
and CPR - green circle). In all cases, the eccentricity of the corrected image was lesser than the
uncorrected, which indicated that the corrected images were closer to a circular shape. The
eccentricity decreased by 15.74 ± 3.24% and 16.62 ± 6.99% after the correction (CPR and CPR-
OCTOPUS) methods were used, respectively. This discrepancy may be attributed to the code-
optimization process during linear algebra operations. However, the ranges of correction
significantly overlap.
Figure 8. Image quality repeatability at 0.05T a) signal-to-noise ratio for the three sessions and
ten days measured using slice 10; b) image uniformity quantified by the standard deviation (ideal
case = 0) in a uniform intensity slice (#16); c-e) geometric distortion quantified by eccentricity
(ideal case = 0) for each session over ten days without (original) and with off-resonance correction
using two implementations
Statistical analysis: Table 2 lists the session-wise and the overall mean, standard deviation, and
coefficient of variation (CV) for the image quality metrics investigated in this study. The session-
wise CV for SNR varied between 4.7% to 5.4%, with a CV of 4.9% across the three sessions over
ten days. This indicates that the SNR of the images produced at 0.05T is highly repeatable.
The image uniformity quantified by the standard deviation in slice 16 had a wider range of ~8%
for session-wise CV, with the overall CV at 10.6%. This moderate repeatability value can be
a b
cde
Session 1 Session 2 Session 3
18
attributed to this metric's fractional dynamic range, significantly impacting percentage changes.
Secondly, Gaussian filtering also affects image uniformity, which can be improved using
acquisition methods (see Discussion section). The repeatability measures of eccentricity for the
uncorrected data were 11%, 1.4%, and 1.1% for the three sessions. The CPR-corrected data had
CVs of 18.8%, 5.1%, and 3.2%. The difference in CVs between the first and the other two sessions
was due to the significant difference on day 1 when the scanner was not turned off before
operation (Figure 8(c), 6(a)). These measurements indicate that the eccentricity values were
highly repeatable (highest values of 1.4% and 5.1% for uncorrected and CPR-corrected) if the
first session (day 1) was excluded.
Table 2. 0.05T repeatability statistics Session-wise and overall mean, standard deviation, and
coefficient of variation for the transmit frequency and image quality metrics indicate a variation of
less than 1% (for frequency) to 19% for off-resonance-related artifacts.
Figure 9(a) shows the scatter plot for SNR against temperature with a correlation
coefficient (r) of 0.48, indicating a weak correlation between temperature and SNR within the
investigated range. However, the change in temperature between the Sessions and between the
days was within 1.5 ℃, which is very small. This also validates the non-significant difference in
Mean SNR compared across the three sessions, which shows a decreasing trend but is within
one a.u. change in SNR. Therefore, the SNR at 0.05T will not vary significantly for a temperature
range of approximately 1.50C. Figure 9(b) shows the correlation plot for SNR and humidity. The
correlation coefficient is close to zero and positive (r=0.05), indicating a very weak tendency for
19
SNR to change within the range of humidity observed in this study. The correlation is so weak
that humidity is unlikely to impact SNR meaningfully. Figure 9(c) shows the mean frequency drift
as a function of temperature. The r value of 0.27 indicates a weak positive correlation indicating
that the amount of drift is modestly dependent on the absolute temperature value within the ranges
investigated. Figure 9(d) shows the correlation plot for humidity versus mean frequency drift with
an R-value of 0.05, indicating a weak correlation. These plots suggest that the 0.05T scanner
produces images with SNR that are not impacted by temperature or humidity within the ranges
observed, enabling repeatability.
Figure 9. Image quality stability at 0.05T The dependence of signal-to-noise ratio on a)
temperature and b) humidity; and, correspondingly, the dependence of mean frequency drift on:
c) temperature and d) humidity show poor correlations indicating the stability of the 0.05T system,
to the ranges of temperature and humidity measured
a b
c d
20
Discussion and conclusions
In this study, we evaluated the effect of temperature, humidity, and off-resonance on image quality
on an in vitro phantom at 0.05T. The study provided bounds of temperature, humidity, SNR, image
non-uniformity, and eccentricity to assess scanner operation and image quality. These bounds
can enable better control of the scanner and the resulting images using look-up table methods to
deliver highly automated scanner operation 10,17–19. This repeatability dataset, including k-space
and image intensity data, will enable further studies, such as very low-field noise modeling for
subsequent denoising algorithms. The frequency–temperature data reproduced similar effects
seen in Figure 6 in ref. 8 for phantoms. These bounds provide a benchmark to achieve while
performing in vivo acquisitions. The supplementary material also shows repeatability data in the
coronal orientation. Supplementary Figure 8 (a, b) shows that the off-resonance was out of range
for day six and Session 1, which may have been due to electromagnetic interference. A similar
effect was observed for the axial orientation in 10-day Session 1 (see green arrow in Figure 5).
One of the limitations of this study is that we did not scan the phantom near the periphery
of the bore, which might be more relevant to brain imaging in such Halbach arrays with smaller
bore sizes. However, iso-center imaging relates well to musculoskeletal imaging. We did not also
examine image contrast in this study as such an investigation will benefit from imaging healthy
volunteers and relevant anatomy. Additionally, we observed ghosting and ringing-like artifacts in
some slices caused residual magnetization due to the four ETL of the 3D TSE (Supplementary
Figure 4) but did not change our acquisition as we do not expect high proportions of long T2
components in vivo as seen in the phantom. We optimized the sigma value to 21, and the artifact
was significantly removed, as seen in Supplementary Figure 3. We did not further investigate k-
space filtering as it is ideal for removing it during acquisition. Our sequences already included RF
21
spoiling. This issue can be further addressed by optimizing the TE or applying crushers to reduce
the undesired coherences caused by increased ETL. Finally, the total scan time for B0 mapping
took 16 minutes per B0 map (2 x 3D TSE scans every 8 minutes), which is not practical for in vivo
studies. Therefore, the B0 acquisition must be further time-optimized by acquiring low-resolution
off-resonance maps. Our choice was directed by the goal of measuring changes in these maps
before and after the image sequence and determining any time-dependent changes in off-
resonance hotspots.
In summary, we found that humidity and temperature in the range investigated did not impact
SNR and frequency. Our findings indicate a high level of repeatability for SNR and magnetic field
homogeneity (quantified by eccentricity, except for day 1, session 1); and moderate repeatability
for image uniformity.
Acknowledgment
The authors would like to thank grant support from the Friedman Brain Institute Research
Scholars program at Mt. Sinai, the faculty idea innovation prize at Mt. Sinai, and the Center for
Precision Medicine joint collaborative research award – a joint program between Mt. Sinai and
Rensselaer Polytechnic Institute (PI: Geethanath).
22
Reference:
1. Geethanath S, Vaughan JT. Accessible magnetic resonance imaging: A review. J Magn
Reson Imaging. 2019. doi:10.1002/jmri.26638
2. Arnold TC, Freeman CW, Litt B, Stein JM. Low-field MRI: Clinical promise and
challenges. J Magn Reson Imaging. 2023;57(1):25-44. doi:10.1002/jmri.28408
3. Liu Y, Leong ATL, Zhao Y, et al. A low-cost and shielding-free ultra-low-field brain MRI
scanner. Nat Commun. 2021;12(1). doi:10.1038/s41467-021-27317-1
4. Bhat SS, Fernandes TT, Poojar P, et al. Low-Field MRI of Stroke: Challenges and
Opportunities. J Magn Reson Imaging. 2020. doi:10.1002/jmri.27324
5. Parsa J, Webb A. Specific absorption rate (SAR) simulations for low-field (< 0.1 T) MRI
systems. Magn Reson Mater Physics, Biol Med. 2023. doi:10.1007/s10334-023-01073-3
6. Yip SW, Konova AB. Densely sampled neuroimaging for maximizing clinical insight in
psychiatric and addiction disorders. Neuropsychopharmacology. 2022;47(1).
doi:10.1038/s41386-021-01124-0
7. Marques JP, Simonis FFJ, Webb AG. Low-field MRI: An MR physics perspective. J Magn
Reson Imaging. 2019. doi:10.1002/jmri.26637
8. O’Reilly T, Teeuwisse WM, de Gans D, Koolstra K, Webb AG. In vivo 3D brain and
extremity MRI at 50 mT using a permanent magnet Halbach array. Magn Reson Med.
2021. doi:10.1002/mrm.28396
9. Geethanath S, Vaughan JT. Accessible magnetic resonance imaging: A review. J Magn
Reson Imaging. 2019;49(7). doi:10.1002/jmri.26638
10. Ravi KS, Geethanath S. Autonomous magnetic resonance imaging. Magn Reson
Imaging. 2020;73:177-185. doi:https://doi.org/10.1016/j.mri.2020.08.010
11. Shridhar Konar A, Qian E, Geethanath S, et al. Quantitative imaging metrics derived from
magnetic resonance fingerprinting using ISMRM/NIST MRI system phantom: An
international multicenter repeatability and reproducibility study. Med Phys. 2021.
doi:10.1002/mp.14833
12. Qian E, Poojar P, Vaughan JT, Jin Z, Geethanath S. Tailored magnetic resonance
fingerprinting for simultaneous non-synthetic and quantitative imaging: A repeatability
study. Med Phys. 2022;49(3). doi:10.1002/mp.15465
13. Shukla-Dave A, Obuchowski NA, Chenevert TL, et al. Quantitative imaging biomarkers
alliance (QIBA) recommendations for improved precision of DWI and DCE-MRI derived
biomarkers in multicenter oncology trials. J Magn Reson Imaging. November 2018.
doi:10.1002/jmri.26518
14. Jackson EF, Gupta SN, Rosen MA, et al. QIBA DCE-MRI Technical Committee Update :
Phantom Studies and First DCE-MRI Profile. Cancer. 2011:9-9.
http://qibawiki.rsna.org/images/2/28/QIBA_Phantom_Poster_RSNA2010_Final.pdf.
Accessed February 3, 2019.
15. Manso Jimeno M, Vaughan J, Geethanath S. Off-resonance CorrecTion OPen soUrce
Software (OCTOPUS). J Open Source Softw. 2021. doi:10.21105/joss.02578
16. Koolstra K, O’Reilly T, Börnert P, Webb A. Image distortion correction for MRI in low field
permanent magnet systems with strong B0 inhomogeneity and gradient field
nonlinearities. Magn Reson Mater Physics, Biol Med. 2021;34(4). doi:10.1007/s10334-
021-00907-2
17. Ravi KS, Geethanath S, Vaughan Jr. JT. Self-administered exam using Autonomous
Magnetic Resonance Imaging (AMRI). In: International Society of Magnetic Resonance in
Medicine. Montreal; 2019.
18. Ravi KS, Geethanath S, Quarterman P, Fung M, Vaughan JTJ. Intelligent protocolling for
23
autonomous MRI. In: International Society for Magnetic Resonance in Medicine2. ;
2020:4154.
19. Ravi KS, Nandakumar G, Thomas N, et al. Accelerated MRI using intelligent protocolling
and subject-specific denoising applied to Alzheimer{\textquoteright}s disease imaging.
medRxiv. 2022. doi:10.1101/2022.10.24.22281473
24
Supplementary material
Supplementary Figure 1. The portable 0.05T scanner and the three locations used to measure
temperature and humidity 1) inside the bore; 2) near the mouth of the bore; 3) in the room
1
2
3
25
Supplementary Figure 2. Raw data (not corrected for off-resonance) slices considered for image
quality metrics: a) signal-to-noise ratio – slice 10; b) image non-uniformity quantified by the
standard deviation inside the inner circle (ideal = 0)– slice 16; c) geometric accuracy quantified
by eccentricity (ideal = 0) – slice 8
SNR (slice 10) Image non-uniformity (slice 16) Geometric accuracy (slice 8)
a b c
26
Supplementary Figure 3. The choice of the standard deviation for the Gaussian filter was based
on the balance between smoothening the ringing artifact and the loss of details, by visual
inspection.
Original image K- space Filter Filtered k-space Filtered image Difference image
Sigma = 15
Sigma = 17
Sigma = 19
Sigma = 21
Sigma = 25 Sigma = 23
27
Supplementary Figure 4. The effect of echo train length (ETL) on image quality a) ETL =1;
b) ETL =2; c) ETL = 4. Increasing ETL increasing the ringing (red arrow) in the image due to lack
of suppression of residual magnetizations (stimulated echoes) from previous refocusing pulses.
ETL 2ETL 1 ETL 4
a) b) c)
28
Supplementary Table 1. The mean, standard deviation and coefficient of variation in temperature
and humidity measurements reported location-wise and for all locations.
0.05T repeatability statistics
Location 1
Location 2
Location 3
All locations
Mean
± SD
CV
(%)
Mean
± SD
CV(%)
Mean
± SD
CV(%)
Mean
± SD
CV(%)
Temperature
(0C)
23.34
± 0.27
1.2
23.49
± 1.24
5.3
23.44
± 0.29
1.2
23.42
± 0.75
3.2
Humidity
(%RH)
45.48
± 4.35
9.5
43.99
± 4.55
10.3
47.60
± 4.78
9.5
45.69
± 4.70
10.29
29
Supplementary Figure 5. Unfiltered off-resonance maps for the three sessions and ten days a)
before and, b) after the 3D turbo-spin echo sequence; c) difference between a) and b). The noise
in these maps is reduced by filtering as shown in Figure 5
Session 1
Session 2
Session 3
Session 1
Session 2
Session 3 Session 1
Session 2
Session 3
Session 1
Session 2
Session 3
Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 Day 7 Day 8 Day 9 Day 10
-2000
2000
0
Hz
Hz
a
b
c
30
Supplementary Figure 6. The effect of filtering on off-resonance maps a) unfiltered and b) filtered
off-resonance maps. The low signal-to-noise ratio of the first and last slices can be noted and the
corresponding filtered outputs show different off-resonance quality compared to the rest of the
slices.
-2000
2000
ab
0
Hz
Hz
B0 map Filtered B0 map
31
Supplementary Figure 7. The comparison of off-resonance correction methods applied on the
a) reconstructed image using the available b) conjugate phase reconstruction (CPR) and c) CPR
using our custom implementation OCTOOPUS
Slice 5
Slice 9
Slice 15
Original CPR CPR - OCTOPUS
32
Supplementary Figure 8. Off-resonance maps before and after 3D turbo spin echo acquisition
acquired in the coronal orientation also shows a single instance of increased off-resonance range
similar to the axial measurements in Figure 5.
a
b
Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 Day 7 Day 8 Day 9 Day 10
Session 1
Session 2
Session 3
Session 1
Session 2
Session 3
Before TSE
After TSE
