More than 6,000 infants develop hydrocephalus in East Africa every year. Magnetic Resonance Imaging is the preferred technique to diagnose hydrocephalus. In countries such as Uganda, MRI is unaffordable at even major referral hospitals. In order to provide a sustainable diagnostic tool we are developing an inexpensive and easy-to-use MRI system that yields images of sufficient quality to diagnose hydrocephalus. This paper describes our first prototype of such a scanner. We explain the lessons that we have learned from this prototype and how we used these to come up with an improved design. We also describe a dataset that has been obtained with this scanner that will be made publically available.
Hydrocephalus is a build-up of cerebrospinal fluid. The increased pressure causes
the head to swell and damages brain tissue. If left untreated it may cause pain,
blindness, mental disorders, and ultimately death. In industrialized countries,
infant hydrocephalus is usually due to either a congenital anomaly or, in low
birthweight premature infants, due to brain hemorrhages from immature blood
vessels. Most of the time it is detected at an early stage, when it is relatively
easy to treat the condition. Hydrocephalus in children in Uganda and other
developing countries, on the other hand, is dominated by infectious causes, is
seasonal, and appears to be linked to environment and farm animals living in
close proximity to families [
Hydrocephalus can be well treated if properly diagnosed. MRI is a widely used diagnostic tool for intracranial disease. Conventional MRI scanners yield high-resolution images, but are often out of reach in developing countries due to their cost, high maintenance, and requirement for cryogenic cooling. Highresolution images are, however, not necessary to diagnose hydrocephalus. A simpler device that yields lower resolution images should be sufficient for diagnostic and treatment purposes.
The aim of our research is to develop a prototype of such an inexpensive MRI scanner that is easy to install, operate and maintain, and that can provide images of sufficient quality to diagnose hydrocephalus and manage its surgical treatment. We aim at a device that costs less than 50,000 EUROs and yields images with a resolution of 5 mm3, which is sufficient for treatment of hydrocephalus. The project is a collaboration between Leiden University Medical Center (LUMC, Netherlands), Pennsylvania State Unverisity (PSU, USA), Mbarara University of Science and Technology (MUST, Uganda), and the Delft University of Technology (TU Delft, Netherlands).
This paper describes our first prototype of a low-field MRI scanner that we have been developing over the past three years. The magnetic field inside the scanner is generated solely using inexpensive permanent magnets. We explain the design choices that we have made, and show some of the bottlenecks that we encountered. We also explain the lessons learned and how we have used these to make an improved design.
Low-field MRI has received quite some attention recently in literature. In [
The structure of this paper is as follows. Section 2 provides relevant background on low-field MR imaging. In particular it gives the signal model. Section 3 explains the configuration of the magnets. Section 4 describes the electronics. Section 5, explains how we created a phantom which allows for a plethora of different geometric configurations. We describe the dataset of 53 different configurations or images we use in our experiments and we give some reconstruction results. In the final section we describe the lessons learned and how we used these lessons in our second protoype scanner. 2
Low-field MRI Conventional MRI scanners use superconducting magnets to generate a strong homogeneous magnetic field, which makes the spins of the hydrogen nuclei line up along the magnetic field. An RF-pulse at resonance frequency is applied, which make the nuclei absorb energy. After the RF-pulse stops, the energy is released which induces a tissue specific signal in a receiver. The resonance frequency depends on the magnetic field strength. To control the area that is examined, linear variations are applied to the magentic field using so-called gradient coils. The data acquired in this way can be turned into an image by applying an inverse Fourier transform.
The most expensive part of a conventional MRI system is the superconducting magnets. In our design we therefore replaced these magnets by a configuration of inexpensive off-the-shelf permanent magnets. Moreover, variations in the magnetic field can be used for spatial encoding, thus removing the necessity for gradient coils.
The signal model for low-field MR imaging is described in detail in [
b = Ax + e aij = ωj2e−iΔωjti ΔxΔyΔz where ΔxΔyΔz is the voxel size, Δω = ω − ω0, j is the voxel number, ω denotes the resonance frequency, or Larmor frequency, and ω0 is the demodulation frequency. The angular frequency ω depends linearly on B: ω = γB, ω0 = γB0.
Here, γ is the gyromagnetic ratio.
Since frequencies, or equivalently magnetic field strength, are not uniquely
mapped onto a location, one measurement is insufficient to make an image. In
order to (partly) overcome this problem multiple measurements can be made by
rotating the magnet (or the sample to be imaged). This technique was used in
[
In this section, the configuration of rings and magnets in our MRI scanner is
described. A more extensive description can be found in [
The first part of the magnet consists of the magnet cylinder which is the core part
of the magnet. The cylinder consists of four rings with in each ring 24 1-inch cube
magnets NdFeB N52 magnets. The residual flux density, Br, of these magnets
is between 1.45 and 1.48 T. These magnets are put into a first order Halbach
configuration [
(a) Cylinder consisting (b) Configuration of (c) Simulated magnetic field reof four rings. the 24 1-inch cube sulting from the ring in 1b.
magnets.
To obtain a more homogeneous field, two extra magnet rings are inserted in the cylinder. These rings are called shimming rings. Both shimming rings are of a first order Halbach configuration, with magnets removed at 0◦ and 180◦ , as can be seen in Figure 2a. The outermost ring represents a ring from the cylinder as in Figure 2, the two rings within this ring are the shimming rings. The first shimming ring consists of 20 magnets and has a radius of 0.0830 m. The second shimming ring consists of 6 magnets and has a radius of 0.060 m. The magnets that are used in these shimming rings are 12 mm NdFeB N48 cubes that have a Br in the range of 1.37-1.42 T. Both rings are placed at a distance of 0.0827 m from the middle of the z-axis.
Besides the shimming ring, gradient rings are added in order to superimpose a gradient on top of the background field. The gradient rings are three identical rings combining 16 NdFeB N42 5 mm cube magnets. These magnets are weaker compared to the other magnets described earlier; Br is between 1.29-1.32 T. The radii of these rings are 0.0905 m and the rings are placed 0.050 m apart from each other. The magnets in these rings are arranged in an adapted version of a first order Halbach ring orientation. Some magnets have an orientation that is rotated over 180◦ , as can be seen in Figure 2b. The red line indicates the separation between the magnets in the normal Halbach configuration and the magnets that have a Halbach configuration rotated over 180◦ .
(a) A cylinder ring with (b) The gradient ring. (c) Magnetic field
meathe shimming rings. surement. The xy-plane is
shown for z = 0.
The magnetic field strength B of the final configuration has been measured using
a Gaussmeter Model GM 2 from Alphalab Inc. This meter can measure strong
and moderate magnetic fields using a probe with an accuracy of 10−5 T. The
measurement process is relatively easy. A COSI Measure moves the probe along
a preprogrammed route through the 3D space within the cylinder. This COSI
Measure is an open source multipurpose 3-axis robot with very high spacial
fidelity; the robot is reliable on a spacial scale of less than a millimeter [
The block diagram in Figure 4 shows the MRI system with each block representing an actual part of it. The PC controls the signals that are generated by the Universal Software Radio Peripheral (USRP) data acquisition. These signals, 2 RF pulses, influence the protons in the sample such that it will transmit a signal back some time after the transmitted RF pulses. This signal is received by the same coil that was transmitting just before. After amplification, filtering and limiting, the received MR signal is sampled by the USRP and transmitted to the PC for further processing. 4.1
The RF excitation pulses that result in 90◦ and 180◦ rotations of the spins in the sample require a lot of power. It takes up to 1 kW in the resistance of the coil’s wire for about 20 μs with a repetition time of say 200 ms. This power is produced by the RF power amplifier (PA). The gain of the PA is about 54 dB. 4.2
The same coil is used for sending and receiving the MRI signals. The transmit/receive switch directs the high power signal from the power amplifier to the RF coil. The very weak signal that is picked up by the RF coil is amplified by 70 dB before entering the USRP. Fig. 4: Block diagram representing the MRI system
The coil that transmits signals to the sample and receives them back from it, is a simple wire wound around a cylinder. It is accurately tuned to the frequency that corresponds to the static magnetic field. The RF frequency at which protons resonate depends on the strength of this magnetic field, see Section 2. The static magnetic field in the coil has a magnitude of about 60mT which means that the resonance frequency is approximately 2.6 MHz. 4.4
This filter is located directly at the input of the USRP to prevent aliasing in the digitized sampled signal. It also limits the amplitude to prevent damage to the analogue input. The filter is a 3rd order Butterworth filter at 20 MHz. 4.5
The data acquisition is done by the USRP1 of Ettus Research. It has two 12 bit ADC boards operating at 64 MS/s and two 14 bit DAC boards operating at 128 MS/s. Each board has two inputs or outputs. An FPGA takes care of the up- and down-conversion in frequency of the signals. This means that a fixed input value can produce an RF output signal from a DAC. Also the received MR signal is down converted to produce a low frequency complex signal. The USRP1 is combined with the GNU software defined radio for generation of the signals and processing at reception. 5
This section describes some of the data we have acquired with our system, and
also gives preliminary analysis results. A dataset has been created for public
use. The samples (images) consist of variations of four geometric shapes: circles,
ellipses, squares, and rectangles. In this section, we summarise how we created
and developed the phantom used to generate the dataset and how the dataset
was acquired. A complete description can be found in [
For the measurements an efficient phantom has been designed that can contain multiple geometric shapes in 12 different positions. The shapes consist of squares, circles, ellipses, and rectangles. A 2D illustration of the phantom and shapes is given in Figure 5. The phantom has a radius of 40 mm. Each hole in the phantom is made of two congruent squares with the same center at 45◦ angles. The dimensions of each square are 5 × 5 × 2.3 mm. A block with the same dimensions is placed behind each shape so that each shape can be rotated by 45◦ in each hole. Each shape is filled with sunflower oil until 3 mm from the top.
Fig. 6: The height h, length l, and width w illustrated for each shape. For each geometric shape, variations are made with the same height but with different length and width, see Table 1. This is done to make the dataset as diverse as possible. A suitable repetition time depends on the T1 value of the liquid placed in the phantom. We use sunflower oil for the measurements for its short T1 value of 90 ms. To avoid spilling, the total height of the shapes is 9 mm. Each shape is divided by a horizontal line at 3 mm from the top. This is illustrated in Figure 6.
SolidWorks, a computer-aided design software, is used to create 3D printable models. The phantom and shapes are printed using the 3D printer Formlabs Form 2, which uses clear resin (SLA). Using clear resin to print the shapes is very convenient as the material does not absorb sunflower oil.
The phantom is printed using the fused filament fabrication 3D printer Ultimaker 2+, which uses a continuous filament of polylactide (PLA). PLA is a compostable thermoplastic material made from renewable resources, such as sugarcane or corn starch.
To improve spatial encoding the RF coil is rotated in the magnet with an angular increment θ. At each increment, the field experienced by the sample changes due to the inhomogeneity of B0. This is illustrated in Figure 7. Starting from θ = 0◦ the RF coil is rotated by increments of ten degrees and after each rotation the signal is measured (0◦ − 350◦). The number of samples recorded after a spin echo sequence is set to 512. One spin echo sequence results in measured signal with a low SNR. Averaging can be applied to improve the SNR of the measurements. For each rotation, the measurements are repeated a hundred times. The hundred measurement are then summed up which results in the measured signal vector of size (1 × 512) for each angle θi. The measurements are done for 36 angles, so the signal matrix for one phantom is of size (36 × 512).
Of each variation of shapes in the phantom a digital image x is implemented. The image matrix x is represented by a square matrix of size (64 × 64) whose elements are pixel values corresponding to black or white. Black pixel values represent areas that do not give signal and white values are the areas filled with oil, which give signal. The process of creating one labelled sample is illustrated in Figure 8. Each sample i ∈ {1, . . . , 53} consists of a signal and its corresponding image (label).
The data of each measurement are stored in a netCDF file, which contains all
relevant acquisition parameters. All variable are in standard SI units. A total of
i = 53 signals has been measured.
Using the model described in Section 2 it is in principle possible to compute an
image from the measurements. For the description of the processing techniques
we refer to [
σj s0
= Pj5=31 yjsj x0 = Pj5=31 yjxj where, uj and vj are the columns of U and V , respectively. Now y can be used to reconstruct the signal s and the corresponding image x, where the reconstruction is denoted by s0 and x0: (5) (6) (7) Let s and x correspond to the first measured sample, Signal 1 and Image 1. The signal and the signal corresponding to the reconstruction are shown in Figure 9. We can see that the amplitude of the signal for each angle is reconstructed reasonably well.
The reconstruction of the measured signal looks promising and therefore it should also be possible to reconstruct a fairly reasonable image. The image and its reconstruction are visualized in Figure 10.
The samples in the measured dataset are varied. For each measurement the
shapes are placed inside the phantom in a different way. From Figure 11 we can
see that the images are built up of other images containing shapes in the same
location. We can conclude that the signal contains information about the position
of the shapes. These results give us confidence that a dataset as described above
can be at the basis of a deep-learning approach to reconstruct the image, without
the need to have accurate knowledge of the magnetic field. We refer to [
In this paper we have described our first prototype for a low-cost MRI scanner. All parts of a conventional scanner have been redesigned and are replaced by less expensive alternatives. Both the main magnet and the gradient coils have been replaced by configurations of inexpensive off-the-shelf permanent magnets. The (hardware) cost of our design is approximately 15,000 EUROs.
Images can in principle be obtained by exploiting variations in the magnetics field. However, this requires very accurate knowledge of the magnetic field. Until now we have not been able to obtain more than rudimentary images, of much lower quality than is needed for medical purposes. As an alternative we are investigating a data-driven approach. We have acquired a dataset for this purpose and have applied some basic processing techniques to it. This dataset will be made publically available through the data repository https://researchdata.4tu.nl/en/. We aim to extend the dataset to make it large enough to be used as a training set for deep learning.
One of the main lessons learned is that encoding spatial information using
field variations induced by the permanent magnets proved to be quite challenging
for image reconstruction. The second prototype, which has been designed by the
LUMC and is described in [
This research is partly funded by TU Delft Global Initiative, by NWO WOTRO under grant W07.303.101 and by STW Open Mind under grant 15549.
The authors thank all the members of the low-field MRI team, in Uganda, the USA and in the Netherlands, for the nice collaboration and many interesting discussions. In particular we thank Andrew Webb, Tom O’Reilly, and Kirsten Koolstra of the Leiden University Medical Center for their assitance with and advice on many aspects of the work. We thank the referees for their useful comments.