This thesis research focuses on MRI technology and more specifically the evaluation of the generalized signal model (full-wave signal). The research aims to investigate how dielectric properties influence the magnetic fields in MRI scans, by using a more complex signal model that takes into account scattered fields caused by dielectric properties of human tissue. Optimizing the Signal-to-Noise ratio (SNR) has been a main research focus since the invention of MRI machines. Higher-quality images, result in better diagnosis and therefore help the investigations of patient conditions. One way to improve the SNR is to operate MRIs in higher frequencies. Operating an MRI in higher frequencies, affects the measured signal at the receiving coil, as dielectric properties of biological tissue, induce currents due to magnetization. By having a full wave signal model, one can model the effect of the dielectric properties and thus take them into account when reconstructing an MRI image. The first step was to investigate whether a magnetic dipole can be simulated as a single emitting loop. The results from the simulation and the analytical solution have been compared. There’s a high correlation be- tween the two and thus it is concluded that it can be used. Using a single emitting loop, the effect of varying conductivity was simulated for a dielectric sphere placed in free space, with the single emitting loop submerged into the solution. The simulations showed that by increasing the conductivity, the measured voltage at the receiver coil reduces. Note that this simulation was performed twice, for relative permittivity values εr = 1 and εr = 80. The reason for choosing these two values was to first eliminate the effect of relative permittivity εr = 1 and to simulate a water-based solution εr = 80. Simulations verified that for εr = 1, the measured voltage at the receiver location is lower than εr = 80. After verifying the full-wave model and observing the effect of varying the electrical conductivity, two more simulations were performed to prepare for the experiment. In the first simulation, instead of a single loop, a series of loops was used to simulate the beam excitation of an MRI. As it is shown in the results, there is not a linear correlation between increasing the conductivity and the induced voltage at the receiver coil. Secondly, the experimental setup was simulated. In addition to the previous simulation, a plastic layer was added to the outside of the ball, to model the plastic sphere used during the experiment. In chapter 4, the results of the experiment are introduced. The results from the experiment could not justify the simulations due to reasons that are explained in the conclusions.

The aim of this research is to build a machine learning model in order to predict the success of invasive surgical treatment on a degenerated cervical spine, based on the baseline X-ray images. The purpose of the results of this research is an application in computer-aided diagnostics and treatment planning in the field of neurosurgery. Spinal degeneration can be described as the gradual loss of spinal structure and a decreased functioning of the spine over time. Before the diagnosis of spine degeneration can be made, a sagittal X-ray analysis of the spine is very important. With the current ageing population and the relatively high prevalence of neck pain and spinal complaints of approximately 30%, there is a great demand on MRI and X-ray analysis in healthcare. Machine learning techniques, and especially convolutional neural networks, seem promising for the application of X-ray image analysis. This research was preceded by a literature study about cervical degeneration and the implementation of ML in spine research. Of the available machine learning techniques, artificial neural networks show the best classification accuracy when it comes to image classification. Convolutional Neural Networks (CNNs) in particular are applicable for the computer vision classification task of this study. That is why, in consultation with the neurosurgery department of the Leiden University Medical Center (LUMC), it was decided to focus on the development of a Convolutional Neural Network (CNN) to perform the binary classification task. The final configuration of the convolutional neural network (CNN) consists of four convolutional layers with ReLU activation function and a maximal pooling function. Batch normalization was applied after the first convolutional layer of the model in order to create a more stable training environment. The false positive rate (FPR) is 19% on average and 15% during the best performing run. The ROC curve shows an AUC during the best performing run of 0.91 and 0.86 on average, whereby 1.0 would be a perfect classifier. It can be concluded that this classification task could be performed by a convolutional neural network (CNN), adapted to the specific classification task. The lowest achieved false positive rate of the model was 15%, which shows a major improvement compared to the current clinical situation at the department of neurosurgery in the Leiden University Medical Center, where about 25% of invasive spinal surgical operations, involving cervical degeneration, are of no benefit to the patient.

Magnetic Resonance Imaging (MRI) is a non-invasive, non-ionizing imaging modality that is commonly used in the clinic today. However, it is an expensive technique. The high purchase, operational and maintenance costs, as well as the need for trained staff with technical expertise, put this technique out of reach for a large part of the world population. To combat this, low-cost MRI systems are being developed. The use of lower magnetic fields allows for reductions in cost and size, increasing the accessibility and portability of the device in developing countries. Nevertheless, the signal-to-noise ratio is proportional to the magnetic field, and so low-field MR images are of a significantly lower quality. As such, a reconstruction algorithm is necessary in order to denoise the image, while preserving the details. Besides the cost, MRI is a relatively slow modality, leading to decreased patient comfort and increased chance of motion artifacts. The data acquisitions can be sped up through Parallel Imaging, which requires the use of a receiver coil array rather than a single RF coil. This leads to incomplete data as well as spatial variations in the coil sensitivity profiles that must be accounted for. Furthermore, while the Field-of-View of the image is generally larger than the spatial support of the object, information on this support is not used in MR image reconstructions. In this work, a reconstruction algorithm is developed to reconstruct MR images using incomplete multicoil data and spatial support information in a low-field setting. This algorithm is based on the single coil algorithm of De Leeuw Den Bouter et al. (2021), which will be extended to the multicoil case using the Contrast Source Inversion method. Three adaptations of the algorithm have been considered. Their performance is characterized in terms of denoising ability, edge-preservation and execution time, using both simulated and real low-field data. A mask S is introduced to include information on the spatial support. The use of this mask is shown to have a positive effect on the quality of the reconstruction, improving the preservation of edges and details. Furthermore, a mask R is introduced to deal with incomplete data in case of accelerated acquisitions. In conjuction with a Parallel Imaging technique, the reconstruction algorithm can yield de-aliased and denoised reconstructions, combatting the inherent drop in SNR caused by the acceleration. Chances for improving the algorithms remain, for instance by the exploring the possibilities for a multiplicative implementation of the Total Generalized Variation regularization term, as well as the automatic detection of the spatial support and the inclusion of compressed sensing.

Magnetic resonance imaging (MRI) is a non-invasive tool to image the body’s anatomy and physiology, but suffers from long scan times. Compressed Sensing (CS) is used to accelerate MRI scans by incoherently taking fewer measurements and using a nonlinear optimization algorithm to image the undersampled data. Convex optimization techniques are generally used for image reconstruction. Minimizing a data fidelity term along with two regularization terms, which are a total variation (TV) based and a wavelet transform based function, is a standard procedure in CS-MRI. Regularization parameters are needed to balance the different terms, but it is impossible to know upfront what the optimal regularization parameters are to get the desired output. A consequence is that the algorithm of choice needs to be executed many times for many different values of the regularization parameters, which is a time-consuming process requiring knowledge of the algorithm. In this work we rewrite and implement the regularization functions in a multiplicative manner by multiplying the data fidelity term with the regularization terms, thereby eliminating the need to tune the regularization parameters. Moreover, we include a region of support (ROS) mask to further accelerate reconstruction. The performance of different combinations of regularization functions and reconstruction algorithms are validated on a simulation study and various experiments on a low-field MRI scanner. This also shows the capability of CS applied to low-field MRI, which has lower signal-to-noise ratio compared to conventional MRI. Of all proposed methods, a nonlinear conjugate gradient method applied to the fully multiplicatively regularized objective function shows the most robust performance.

In order to make offshore working conditions safer, Ampelmann B.V. builds platforms that stay stable by compensating for a ship’s motion. These platforms are controlled via motion reference units (MRUs). In order to test the performances of these MRUs, a linear motion MRU test setup was developed. The system was divided into three parts: hardware, software and MRU assessment. This thesis focuses on the design and implementation of the MRU assessment part and on testing MRUs with the complete prototype setup. Ampelmann supplied the MRU with which the test were conducted. Validation tests have been done with a simulator and with the setup itself. Included in the assessment of the MRUs was the latency, signal-tonoise ratio and low-frequency behaviour, compared to a ground truth. Concluded from the tests was that the designed test setup has potential to accurately assess the performance of MRUs. There are some minor improvements to be made, but the overall system works as intended.

The relation between the input of an ideal optical single-mode fiber and the corresponding fiber output constitutes a nonlinear system that can be described using the nonlinear Schrödinger equation. This nonlinear system has the interesting property that it can be solved analytically using nonlinear Fourier transforms. To utilize this property, new methods of optical communication are being developed by embedding information in the nonlinear Fourier domain and employing fast nonlinear Fourier transforms. Many of the recent works use a specialized form of inverse nonlinear Fourier transform to generate information-bearing fiber inputs in the form of so-called multi-soliton pulses. Recently, multiple fast inverse nonlinear Fourier transform algorithms that can generate multi-solitons have been proposed. The goal of this thesis is to study and improve these algorithms, in particular, with respect to their computational complexity. Based on the literature survey, discrete Darboux transform combined with other discrete techniques is studied and a new algorithm is proposed. The algorithm employs a single-start approach in which discrete Darboux matrix is computed at only one sample point and rest of the samples are computed by evolution of the Darboux matrix. The algorithm is hence named as discrete Darboux evolution algorithm (DDE). The errors in the generated signal and run-time are studied by comparison with the classical Darboux transform (CDT). The DDE algorithm is shown to have floating point operations complexity of O(KN) for K eigenvalues and N samples. However, in a limited precision environment the number of samples that can be generated is found to be limited. To better understand the effects of machine precision, both the CDT and DDE algorithms are studied in a multi-precision environment. Certain insights from the study are used to develop two modifications to overcome the limitations. The first modification computes the signal using multiple single-start runs while the second one uses a multi-start approach. The second modification is shown to have errors comparable with other fast algorithms in literature. Additionally, in a qualitative comparison it is shown to be potentially faster than existing algorithms in a certain regime.

Magnetic resonance imaging (MRI) scanners are a crucial diagnostic tool for radiologists. They are able to render two­ and three­dimensional images of the body without exposure to harmful radiation. MRI systems are, however, costly to build and maintain. This adversely impacts access to these scanners in developing regions. In an effort to combat this problem, a low­field MRI scanner is being developed. Conventional MRI scanners utilize a superconducting solenoid to generate the main magnetic field. The low­field scanner, on the other hand, induces the main magnetic field through a Hallbach array of permanent neodymium magnets. While beneficial for production and maintenance costs, as well as portability, the Hallbach array is not able to generate a perfectly homogeneous magnetic field. The inhomogeneities present in the main magnetic field result in distortion of the images when reconstructed using conventional fast Fourier transform (FFT) methods. To counteract this, a reconstruction method that utilizes field information needs to be employed. In this thesis, existing methods to determine and utilize the field information to correct image distortion are explored. From this analysis, it is evident that model­based (MB) methods are most suitable for reconstruction of data from the low­field scanner. Current MB methods are only implemented for two­dimensional reconstruction. The goal of this thesis is to expand these methods to three­dimensional reconstruction. A novel MB method for three­dimensional reconstruction is presented. This new method is able to circumvent memory constraints that arise from reconstruction of large data sets. Though the new method requires several hours to reconstruct a 128 × 128 × 30 data set, visual inspection indicates that an accurate result is achieved.

Current high-end medical X-ray intervention devices provide a tremendous amount of high-definition images per second. Combined with the additional inputs from the numerous auxiliary devices, processing and compositing the data in real-time quickly becomes an arduous engineering challenge. Furthermore, the critical nature of this application domain allows no room for error, whereas the process has to be simultaneously flexible enough to permit the unhindered work of the corresponding medical professional. Such medical imaging applications presently rely on complete hardware implementations of their processing pipelines and compositor engines, as opposed to earlier adopted paradigms of hardware-accelerated solutions. Said engines are usually realised in FPGA platforms due to their massively parallel computational capabilities and exceptional connectivity. However, given the vast amount of input streams arriving at the compositor's pipeline, switching and routing to available processing resources has remained a persistent issue. Present commercial solutions prove to be either too costly, too slow or too inflexible for developers to consider. On the other hand, academic literature centred around the subject is severely lacking. This thesis proposes a unique hardware architecture solution to mitigate the aforementioned problem. The proposed architecture, developed in collaboration with Philips Healthcare and TU Delft, exhibits an average end-to-end latency of 100 ns and a throughput of 7.2 Gb/s. Additionally, these results were realised under marginal resource utilisation, proving that such a switch can fit in the FPGA device and, subsequently, be developed as an integrated part of the compositor engine. Finally, the limited amount of reserved memory resources allowed for near-linear scalability of the proposed design should an increased amount of I/O devices be added; as well as dynamic portability should migration to smaller devices be a concern.

Magnetic resonance electrical properties tomography is a type of quantitative magnetic resonance imaging that aims to reconstruct the conductivity and permittivity of biological tissue. These electrical properties of the tissue can be used to compute the specific absorption rate, to differentiate tumours from healthy tissue and for hyperthermia treatment planning. Several methods to reconstruct these electrical properties exist with different degrees of success. Combining analytical reconstruction methods with deep learning methods is left relatively unexplored in the field of magnetic resonance electrical properties tomography. Hence, this work explores such hybrid methods in which deep learning is embedded in an analytical reconstruction method. A recurrent inference machine is integrated into the iterative reconstruction scheme called Contrast Source Inversion, in an attempt to decrease its high computational load. Additionally, a U-net is trained to correct reconstructed conductivity maps using discrepancies in measured- and reconstructed phase data, which is based on the relation between conductivity and phase in the Helmholtz equation. The recurrent inference machine embedded version of contrast source inversion failed to achieve a desirable reconstruction quality with its current implementation. However, the large amount of potential improvements to its implementation motivates further research into its application before discarding it. The conductivity correction U-net is able to correct conductivity errors as small as 0.13 S/m when used iteratively or 0.05 S/m when used a single time when noiseless data is used. Further research in its capabilities of handling noisy data is required to assess practical usage.

Commissioned by Ampelmann Operations B.V., a testing system for Motion Reference Units has been developed. This testing system provides a platform to compare different motion reference units to be able to verify their capabilities. The design of this system includes validation of a mechanical setup, as well as the development of an electrical controller to drive the mechanical setup.

The objective of the bachelor graduation project detailed in this thesis is to develop a wireless sensor patch that monitors the health of newborns. The sensor patch replaces the monitoring tasks of nurses during maternity care. The patch measures temperature, heart rate, oxygen saturation and bilirubin. Many babies are born with high bilirubin, causing a condition called newborn jaundice. All measurements will be done once every two hours and the data will be sent to the hospital. This thesis focuses on everything related to energy and power within the wireless sensor patch. Energy storage, power and energy management and visualising data with the aid of a MATLAB model are topics that will be designed and implemented in this thesis.

Magnetic Resonance Imaging is a widely used technique to obtain images of the interior of the human body for diagnosis and treatment. MRI machines capture the raw signal in spatial frequency domain i.e. k-space and the image is obtained via Fourier transform. The Cartesian acquisition is one of the most commonly used acquisition patterns in MRI and is most susceptible to the patient's motion. Due to long scanning times, the possibility of the patient's movement is higher which introduces bulk motion artifacts reducing the quality of the image. Motion artifacts can affect the diagnosis and the necessity of re-scanning can cause significant financial costs as well as delays in diagnostics. Current methods for correcting motion artifacts work in image domain which need completely sampled k-space for reconstruction and hence are not useful for real-time artifacts correction. In this thesis, machine learning methods that can detect, estimate and correct motion artifacts in k-space were investigated making it possible to correct artifacts in real-time without the necessity of reconstruction. For each of these methods, we analyze the performance and discuss the merits and demerits.

The TU Delft and the LUMC are creating a low-field MRI scanner to use in third world countries. This type of MRI scanner has many advantages, but a downside is the amount of noise in the obtained images. A way to reduce noise is diffusion filtering. This thesis discusses the theory of some linear and nonlinear diffusion filtering methods and tests them on several test problems. The methods range from the basic linear method, the heat equation, to the more advanced Perona-Malik method. The results indeed show that the Perona-Malik method generates better results, sometimes when combined with a Gaussian kernel to decrease its ill-posedness. Two numerical methods have been compared for these anisotropic diffusion filtering methods: the Forward Time, Central Space method and the Additive Operator Splitting method. The advantage of the AOS method is the unconditional stability for al positive time step sizes, while FTCS implementation is only stable for time steps smaller than 0.25. The results for the AOS method were also slightly better than for the FTCS method and almost never lead to instability. The FTCS method showed instability more often. The parameter choice is of significant importance for the outcome. Several options for determining the gradient threshold parameter K, time step size and stopping time S have been investigated. A new estimation for the time step size and stopping time S have been proposed and compared. The results are images with good visual quality for both introduced methods. Also, an adaptive time step has been tested to overcome the instability for unstable methods and this seems to be succesful. Lastly, region growing segmentation has been applied to images obtained with the proposed stopping time S. Segmentation is an additional way to investigate the quality of these outcomes. For most test problems the partitioning resembles the partitioning of the original image. However, choosing certain parameters stays a challenge and can be of interest for further examination. It also remains a goal to investigate the discussed methods on data obtained with the actual TUD/LUMC MRI scanner.

Magnetic Resonance Imaging is one of the most widely used imaging modalities nowadays and it performs especially well imaging human organs such as the brain and liver. One of its main limitations is the relatively long imaging times, to overcome this issue and speed up the data acquisition, several techniques such as Parallel Imaging or PI have been developed. These techniques require advanced hardware and software to be able to decrease the acquisition time. On the hardware side, a highly efficient insert gradient coil has been designed and built at the University Medical Center Utrecht. Specialized software has to be implemented to optimally make use of this hardware. One of the recently proposed PI methods called Wave-CAIPI has been proved to achieve a ninth fold acceleration factor without compromising image quality. This project aims to investigate the time gain that can be achieved when combing the insert gradient coil with a Wave-CAIPI strategy. Two main aspects are reviewed. The first one is the maximum achievable under-sampling factor that does not compromise image quality. The second one is the decrease in acquisition time that can be obtained when using the insert gradient coil compared to conventional gradient systems while maintaining image quality. To do so, the strategy has been implemented and extensive simulations have been performed to optimize the MR acquisition parameters. To prove the results from the simulations, the Wave-CAIPI sequence was implemented in a 7T scanner at the UMCU, where the acquired data was retrospectively under-sampled, obtaining the wave image to be further reconstructed. Limitations of previous works on Wave-CAIPI have been the gradient specifications, which can be overcome by the high-efficiency coil. It has been concluded that shorter acquisition times without compromising image quality are possible when using the insert coil compared with conventional systems. The time gain can be up to a factor of five, and sixteen fold under-sampling factors could be possible. The time gain can be especially useful for Echo Planar Imaging sequences, where switching faster gradients allows to acquire more signals in less time. The next steps for this research are to prospectively under-sample the data in a Wave-CAIPI fashion with a sixteen under-sampling factor and corroborate if sequences such as Echo Planar Imaging can be benefited with the time gain.

The development of neural prostheses, especially those directly targeting the brain, requires extensive research and modelling before clinical trials can be performed. Currently, the resolution of artificial vision is not sufficient for everyday tasks. By studying the expected spatial extent of stimulation, we aim to provide insights to researchers that can be used to improve the resolution of artificial vision. The goal of this MSc thesis was to visualise the shape and intensity of electric fields in the cortex as a response to intracortical microelectrode stimulation to observe the expected regions of neuronal activation considering electrode design parameters. To do this, parameters that can influence the generated electric field and the regions of activated tissue have been defined. Implementing these parameters in a Finite Element Model (FEM) allows the computation of the generated electric field in 3D of a stimulating electrode to observe the spatial extent of activated tissue. The spatial extent of activated tissue can be estimated using simplified methods such as the activating function (AF) or current density threshold. The result is a parameterised framework that creates a Visualisation of Neuronal Activation (VoNA) that can be used to assess activated tissue regions for varying scenarios by defining material properties and dimensions of the model, and allows for the adjustment of the stimulation configuration, electrode contact spacing, customisation of electrode size and modulation of stimulation current. It enables the user to tune the model settings towards their specific needs and explore the possibilities by visualising the results from different angles by defining subsets of the entire solution. In line with expectations, the presented models show that the model parameters can influence factors such as the generated electric field, the current density and electric potential, which are indicators of neuronal activation. The findings support the hypothesis that these parameters should be considered during electrode design to achieve accurate stimulation.

A major cause for voiding dysfunction is the inability to relax the urethral sphincter. KiloHertz Frequency Alternating Current (KHFAC) stimulation can block signals that are travelling through the body; applying this type of stimulation at the pudendal nerve could inhibit the pulses that lead to contraction of the urethral sphincter, and restore voiding ability. In order to design a successful KHFAC block therapy for the pudendal nerve, it is necessary to understand what impact different stimulation parameters have on efficacy, safety and power-efficiency. This thesis will therefore test earlier researched KHFAC stimulation parameters against a new quality measure, study the impact of new waveform alterations, and study how bipolar electrode design can improve KHFAC therapy. By utilizing the theory behind the mechanism of the KHFAC nerve block, a new block-determination model was developed that is over thirty times faster than the classic model. The McIntyre-Richardson-Grill model was chosen as the implementation of the axon model, and the bipolar electrode was modelled as an electric dipole. The simulation experiments revealed that the charge per phase of the KHFAC signal at block threshold could be reduced, without increasing the amplitude of the signal, by introducing interphase delays to the waveforms and by creating asymmetric charge-balanced waveforms. Triangular waveforms were shown to also require less charge per phase than a regular square wave to block, albeit with a higher amplitude. A correctly aligned bipolar electrode set-up with an interpolar distance that was about the same as the electrode-to-axon distance was shown to result in reduced block thresholds. Overall, this thesis has shown how stimulation parameters can be chosen to develop an effective KHFAC block therapy for the pudendal nerve.

Artificial Intelligence has become a dominant part of our lives, however, complex artificial intelligence models tend to use a lot of energy, computationally complex operations, and a lot of memory resources. Therefore, it excluded a whole class of hardware in its applicability. Namely, relatively resource-constrained low-cost hardware. This paper investigates learning methods that are potentially better suited for these types of devices: the forward-forward algorithm and Hebbian learning rules. The results are compared to backpropagation with equivalent network configurations, training hyperparameters and internal data types on different types of low-cost hardware. Backpropagation has consistently outperformed other algorithms in various tests. It exhibits higher accuracy, faster training, and faster inference compared to forward-forward models. However, forward-forward models can come close to matching backpropagation's performance, but they suffer from longer training times and decreased performance with multi-layer networks. Additionally, a poorly trained forward-forward model is sensitive to quantization, resulting in a significant drop in accuracy. On the other hand, forward-forward models offer the benefit of independently training each layer, allowing for more flexibility in optimizing the training process. Hebbian models were not found to be competitive, displaying performance below the required threshold. Smaller models for MCU and FPGA would likely perform even worse.

Quasi-Normal Modes (QNMs) are a key concept in reduced-order models. In this thesis, we use Finite-Difference approach to create a discretized model of an open electromagnetic system in order to identify its QNMs and the Perfectly Matched Layer (PML) modes. We develop a structure of the QNMs and the PML modes and identify different regions of the eigenvalue distribution in our system. We validate the idea of dominant QNMs by using the identified QNMs to get the solution that was obtained using the Finite-Difference method.

For a proper three-dimensional reconstruction of histology serial sections, adjustment of the slices is necessary for combining serial sections. Deformations occur due to the sectioning and acquisition pro-cess of the microscopic analysis of histology. By reconstructing the deformations with a transformation of the sections, a mathematical correction on the images can be applied. By using image registration, a transformation function is searched for to minimize differences between the histology slices. Due to the non-linearity of the distortions, prior knowledge is required in order to have a solvable problem. Additional information in the form of elasticity regularisation is considered. Implementing the elastic regularisation with a finite element method, provides a continuous transformation function With the continuous function, in a natural way, alignment can be monitored for folding transformations. In this work, the (bi-)linear and (bi-)quadratic elements for the finite element method are implemented and compared with the finite difference method. It is observed that for the different kinds of elements, the (bi-)linear elements yield best results with the validity of the transformation. Moreover, the computa- tional costs for the bi-linear elements are the cheapest. Compared with the finite difference method, the differences in accuracy are not noteworthy but the computational time of the finite element method is longer. Furthermore, to steer the matching in an accurate direction, improvements are proposed by applying local stiffness of the elements or adding soft constraints on the volume of the elements. This results in significant improvements in the transformation. For these two approaches it is observed that local stiffness is more restrictive than volume-preserving. Solving the optimization problem, a Gauss-Newton method to search for descent directions is applied. A matrix-based and a matrix-free approach of elasticity regularisation is considered in the linear system of finding a descent direction. While the matrix-free approach decreases the memory usage, the computational costs are significantly increased.

Ultrasound imaging is an important modality in ocular oncology, allowing for fast examination of the eye by the ophthalmologist themselves. It is clinically used to measure tumour sizes for treatment planning. However, ocular ultrasound is limited to two-dimensional imaging, and suffers from poor contrast between tumour and sclera, which negatively impacts the accuracy of tumour measurements. In this work, low field MRI is investigated as a possible alternative for ultrasound imaging. Design requirements are a scan time of less than 4 minutes; resolution of 1.0 mm isotropic; Field of View (FOV) large enough to contain the eye and the orbit; contrast sufficient to distinguish the sclera, vitreous, tumour, lens and lipid. The experimental setup consists of a 46 mT Halbach-array based scanner, a volume coil as transmit coil and a custom-built surface coil as receive coil. Images are made of a water phantom to characterise the FOV, and a porcine eye to characterise the contrast. The FOV is found to meet the requirements, and the contrast is sufficient to distinguish the sclera, vitreous, lens and lipid in porcine eyes. The resolution is too low and the scans take too long (about 5 minutes at a resolution of 1.0 × 1.0 × 7.5 mm). Increasing the resolution and decreasing the scan time will result in a low Contrast to Noise Ratio (CNR), causing the contrast requirement to be violated. Fast, high-resolution three-dimensional imaging is therefore not feasible on the current system. The CNR can be improved by using a higher field strength, which requires the development of new hardware. Furthermore, in order to develop a clinically useable system, it is necessary to determine tumour contrast, design optimised pulse sequences, and test the method on human subjects.