Background: Surgical removal of liver tumors necessitates a thorough preoperative assessment to ensure adequate future liver remnant function, which is crucial for hepatic regeneration. Imaging techniques like hepatobiliary scintigraphy (HBS) and dynamic contrast-enhanced (DCE) magnetic resonance imaging (MRI) assess liver function by measuring the uptake of liver-specific contrast agents. Intravoxel incoherent motion (IVIM)-MRI measures both molecular diffusion and perfusion-related motion of water molecules in the liver. This provides valuable insights into tissue microenvironment changes that can indicate liver dysfunction. However, the potential of IVIM-MRI in this context remains unexplored. This study aims to evaluate the feasibility of IVIM-MRI for liver function assessment and its relationship with DCE MRI. Methods: Twenty-one patients scheduled for major hepatectomy underwent preoperative assessment involving HBS, a 20-minute DCE-MRI series, and IVIM-MRI with 15 b-values. DCE-MRI parameters (hepatocyte uptake Ki(min−1), arterial plasma flow Fa (mL/min/100 mL), and venous plasma flow Fv (mL/min/100 mL)), were analyzed using the Sourbron model. IVIM-MRI parameters (diffusion D (mm2/s), pseudo-diffusion Dp (mm2/s), and perfusion fraction f (%)) were extracted using a UNET model developed at Amsterdam University Medical Centers. Correlation between parameters was assessed using Pearson correlation analysis. Furthermore, Blant-Altman was employed to assess the inter-observer variability and the reproducibility of the DCE-MRI parameters. Results: In 19 patients, weak correlations were observed between DCE- and IVIM-MRI parameters, with correlation coefficients ranging from r = −0.326 to r = 0.443. Despite the lack of significant correlations between these parameters, strong correlations were observed between DCE-MRI Ki and HBS (r = 0.80, p < 0.001). Moreover, DCE-MRI parameters demonstrated high reproducibility, with Bland-Altman mean biases ranging from -1.79 to -0.08. Conclusion: The weak correlation observed between DCE- and IVIM-MRI parameters suggests that IVIM-MRI may have limited utility in preoperative liver function assessment. Nevertheless, DCE-MRI may serve as an alternative to HBS, potentially providing a one-stop shop for preoperative liver assessment with MRI. Further research is necessary to explore its potential in diverse populations with varying liver function.

Flow visualization is an important topic in many scientific domains and has been an active field of research for many years. Many different methods of analysis can be used in order to analyze flow, however recently big progress have been reported on the manifold learning algorithms for high-dimensional data. This thesis investigates the use of Stochastic Neighbor Embedding (SNE) methods, t-distributed stochastic Neighbour embedding(t-SNE) and hierarchical stochastic Neighbor embedding (HSNE) for flow analysis. In this thesis the Manivault ,Veith et al., 2024, software platform has been used in order to create an interactive analysis tool for SNE methods used on flow data. This tool consists of a 3D viewer plugin that visualizes the full path lines and an existing scatterplot plugin that is used in order to interact with the created SNE maps. The experiments and comparisons reported in this thesis aimed to compare the use of t-SNE and HSNE for analysis of flow structures in 4D Flow MRI data. From this, it can be concluded both t-SNE and HSNE are useful for interaction with and analysis of 4D Flow MRI data. t-SNE can best be used in order to explore and analyze flow data in search for flow structures, and comparing flow patterns between subjects. HSNE on the other hand gives a better separation between different flow components, however at the expense of a less accurate preservation of vortices and other flow structures.

Magnetic Resonance Imaging (MRI) is an important imaging modality, since it can create high-resolution cross-sectional images of the human body. In MRI scanners, the nuclear spin magnetization is excited using radio-frequency pulses. Images are created based on the time-evolution of this magnetization, which is characterized by relaxation times (T₁,T₂,T_1ρ,…). These relaxation times change from tissue to tissue, and between healthy and diseased tissue.

Rotating frame (T_1ρ) relaxation measurements are a promising technique for assessing slow molecular interactions in tissue. This has applications in articular cartilage imaging, and cardiac imaging without contrast agent injection. T_1ρ measurements require continuous application of an electromagnetic excitation field. Variations of both the main magnetic field and the excitation field strength cause this excitation to be off-resonant. This in turn leads to contrast loss in the final images.

Adiabatic pulses, whose orientation changes slowly in time, are resistant to these off-resonance effects. Their effectiveness is dependent on their parameters, such as the peak sharpness β or the frequency modulation amplitude A. Conventional optimization techniques for these parameters neglect off-resonance effects.

In this project Redfield theory was used to create a pulse optimization algorithm that can take this off-resonance behaviour into account.

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.

Long acquisition times impede the routine clinical use of quantitative magnetic resonance imaging (qMRI). qMRI quantifies meaningful tissue parameters in T1-, T2-, and PD-maps, as opposed to conventional (qualitative) weighted MRI (wMRI), which only visualises contrast between tissues. Although methods exist that generate synthetic wMRI from qMRI, the inverse problem has not been thoroughly studied yet. A method to generate qMRI from wMRI would be beneficial as it does not change current clinical workflows and enables retrospective quantitative analysis. This thesis investigates to what extent fully convolutional networks are successful in generating qMRI from T1-weighted, T2-weighted, PD-weighted and T2-weighted-FLAIR scans. A set of synthetic wMRI scans from 97 healthy volunteers was split into training, validation and test sets for development of our models. We varied model architectures, loss functions and learning rates during training, in order to find the best performing models. These were able to predict qMRI with median errors of approximately 5% on the test set. Additionally, we determined the amount of information contained in the input scans by training models using different combinations of the input. These results showed that T1-weighted, T2-weighted and PD-weighted scans were the most important. Models trained on synthetic wMRI were tested on an additional dataset of real wMRI. This resulted in higher median errors of 27.4%, 12.0% and 8.7% for T1-, T2- and PD-maps respectively. Furthermore, the same models were tested on a third dataset of synthetic tumour scans and mainly showed errors around the tumour core. These results show that more research is necessary in order to improve the performances of models generating qMRI to a clinical standard.

The Recurrent Inference Machine (RIM) has been developed as an alternative to the clinically used Compressed Sensing (CS) algorithm, using Deep Learning (DL). A common issue with DL networks is the generalization of the network to features that have not been trained for. In this study we evaluate the robustness of the RIM to white matter lesions in FLuid Attenuated Inversion Recovery (FLAIR) brain MRI data. We are evaluating two pre-trained RIM networks, one trained on T1 brain data and another trained on T2 knee data. This evaluation was done by comparing the two networks to CS in terms of the average relative Signal-to-Noise Ratio (SNR) and Contrast Resolution (CR) that is achieved on 15 datasets acquired from Multiple Sclerosis patients. From these comparisons it shows that the network trained on T2 knee data performs similar to CS in terms of the relative SNR, while having a higher CR. The network trained on T1 brain data has both a lower relative SNR and CR, compared to CS. The data suggest that the RIM trained on T2 knee data is robust to the inclusion of lesions in an area that the network was not trained on.

In this report, the conversion from spin-echo signals, obtained with a low-field hand-held MRI scanner that was designed and built at the Leiden University, to images of the proton density within the sample is considered. This scanner does not make use of switchable gradient coils, but instead relies solely on the natural inhomogeneity of the field and on translations of the sample over this field for its spatial encoding. Specifically, an attempt is made to answer the question of how we can reconstruct a phantom using this kind of scanner. This is done by deriving a signal model, discretising it and writing it as a linear least squares problem. Then, we can make use of the techniques of Cojugate Gradient for Least Squares (CGLS) wih `2-regularization and Generalized Conjugate Gradient Minimal Error (GCGME) with `1-regularization for the difference between neighbouring pixels in order to solve this inverse problem. Firstly, we theoretically consider combinations for magnetic field geometry and measurement strategy for their usability for image reconstruction. After this, the obtained strategies are tested in three experiments, with two magnets and two samples. We start by doing this numerically, using a simulated phantom in combination with a measured magnetic field and the translation strategy. By doing this, we can determine if reconstruction is possible using that combination of field and strategy. Finally, the strategy is tested on real samples. Using numerical phantoms in combination with the magnetic field and translation strategy used in the measurements, we were able to correctly reconstruct the phantoms. However, the reconstruction broke down when data from real samples was considered. A variety of possible improvements is discussed. The improvement that would have the most impact would be to design a magnet that has a less uniform gradient in the z-direction, and instead has some locations in the xy-plane where the field falls slowly as function of z, but quickly in other locations.

Early detection of Alzheimer's Disease (AD), i.e. before symptom onset, would provide the opportunity for development and testing of interventions at earlier stages, when the disease process may still be altered or interrupted. Computer algorithms combining machine learning with non-invasive imaging and other biomarkers for AD have been developed in an effort to improve early detection methods. However, so far, none of the individual algorithms perform at a level that qualifies for clinical use. In this study, we investigated whether combining several existing AD prediction algorithms improves performance and generalisability.
State-of-the-art AD progression prediction algorithms were collected from the TADPOLE-SHARE project. Algorithms were trained on data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) study and made forecasts of the clinical diagnosis (CN, MCI, or AD). These algorithms were combined using i) simple, unlearned fuser methods and ii) learned fuser methods. In total, seven experiments were conducted, exploring different combination strategies with increasing complexity of fusers. Finally, we implemented and added our own individual algorithm, a residual neural network (ResNet). All individual algorithms and ensembles were evaluated with the multiclass area under the curve (mAUC) and the balanced classification accuracy (BCA) performance metrics. Statistical significance was evaluated with the McNemar test.
Results. TADPOLE-SHARE resulted in the collection of eight algorithms, from which five were reused for combination. Overall, combining algorithms slightly improves performance (i.e. increased BCA and mAUC), although improvements were not statistically significant (McNemar test). Both BCA and mAUC showed a trend of improved performance with increasing fuser complexity i.e. data learned fusers and re-entering original data features. DoubleResNet was the best performing ensemble (BCA = 0.809 [±0.026], mAUC = 0.902 [±0.020]) and performed slightly better than the best scoring fused algorithm EMCEB (BCA = 0.761 [±0.029]; mAUC = 0.866 [±0.020]).
These preliminary results suggest that combining pre-existing AD progression prediction algorithms might provide the increase in performance and generalisability needed to enable clinical translation. To do so, future work should be focused on increasing the interoperability of currently existing and newly developed algorithms.

Fat fraction (FF) and apparent diffusion coefficient (ADC) values estimated by Dixon MRI and diffusion weighted MRI (DWI) techniques respectively, are relatively new quantitative imaging parameters and increasingly accepted as imaging biomakers for all sorts of purposes. The aim of the BOCASEcA study is to research whether these techniques can be used as biomarkers for patient-reported xerostomia and dysphagia post-radiotherapy. In this project, steps have been taken to validate the use of certain fat quantification and ADC mapping protocols in the BOCASEcA study. mDIXON Quant is a Philips product designed for MR fat quantification. We performed phantom studies and a healthy volunteer study to evaluate the accuracy and repeatability of a standard mDIXON Quant protocol with default parameters and an mDIXON Quant protocol that is used in LUMC on muscles throughout the whole body. Another phantom study was done to evaluate the (geometrical) accuracy of DWI-SPLICE, a technique that can be used for ADC mapping. This DWI technique is known to have less susceptibility issues than conventional EPI-DWI. We tested both the accuracy and deforming artifacts for an EPI sequence and a clinically used DWI-SPLICE protocol from LUMC. Adequate accuracy and robustness were observed for the standard Philips mDIXON Quant protocol. The LUMC muscle protocol, however, yielded incorrect measurements that were underestimations of the real FF values. The DWI-SPLICE protocol showed better geometrical accuracy than the EPI protocol. Accuracy of the ADC measurements was sufficient for ADC values higher than 0.6 x 10−3 mm2/s which is a clinically relevant range. Since the accuracy of both the standard Philips mDIXON Quant protocol and DWI-SPLICE protocol was validated, it is recommended that they are used and further optimized for the BOCASEcA study.

A craniotomy is a procedure were a neurosurgeon has to open the skull to gain direct access to the brain. When a brain tumor has to be removed from a patient, the craniotomy position is of great importance. This mostly defines the access path from the skull surface to the tumor and thus also what healthy brain tissue will be removed to access to the tumor itself. To minimize the amount of important brain structures that are removed, the craniotomy has to be carefully planned. This is a complex procedure, where a neurosurgeon is required to mentally reconstruct spatial relations of important brain structures to avoid these as much as possible.

We propose a visualization using augmented reality which may assist in the planning of a craniotomy. In this visualization the goal is to show important brain structures aligned with the physical position of the patient. This should allow better perception of the spatial relations of these structures and thus assist the neurosurgeon. Additionally to the visualization of the structures, we created a heat map that is projected on top of the skull. This should give a quick overview of in which areas there are many important structures between the tumor and the skull surface, and should therefore be avoided.

User studies were conducted amongst neurosurgeons and surgeons from other fields to evaluate the proposed visualization. We found that many of the participants indeed thought that the visualization can assist in surgery. For the specific case of craniotomy planning, several improvements have to be made on the heat map before it can be useful. Nevertheless, the visualization of the structures in itself can assist neurosurgeons in the planning of a craniotomy. Although more work has be performed at practical aspects of the visualization to make it ready for clinical experiments.

Biological vision adopts a coarse-to-fine information processing pathway, from initial visual detection and binding of salient features of a visual scene, to the enhanced and preferential processing given relevant stimuli. On the contrary, CNNs employ a fine-to-coarse processing, moving from local, edge-detecting filters to more global ones extracting abstract representations of the input. In the current paper we propose the extraction of top-down networks, by reversing the feature extraction part of the baseline, bottom-up architecture. This coarse-to-fine pathway, by blurring out higher frequency information and restoring it only at later stages, offers a line of defence against attacks introducing high frequency noise. High resolution of the final convolutional layer's feature map can contribute to the transparency of the network's decision making process, as well as favor more object-driven decisions over context driven ones and thus provide better localized class activation maps. The paper offers empirical evidence for the applicability of the method to various existing architectures, but also on multiple visual recognition tasks.

The operating room is one of the most complex and expensive environments in the hospital. Research has been focusing on improving the efficient use of the OR time, for instance by using intraoperative data to update the planning of the OR during the day. This thesis used a deep learning network to automatically recognize surgical tools and pre-defined surgical phases present in the recordings, to ultimately track the progress of the procedure. The aim of this thesis is to assess the performance and applicability of this deep learning method for the use of image recognition in a medical environment. To ultimately predict the remaining surgery duration and improve the efficiency of the OR planning. Two datasets of laparoscopic recordings were used, one containing laparoscopic cholecystectomies and one containing total laparoscopic hysterectomies. The surgical tools and the pre-defined phases of the procedure were annotated in every recording, after which the deep learning network was trained with this data. The performance of the network was tested in multiple experiments. The results showed that the performance of the deep learning network was promising and in line with published literature, but that the results varied between recordings. An experiment using three different sized datasets showed that a larger dataset corresponded with the best results and results that varied the least between recordings. Testing the generalizability of the network showed that a network trained on one type of surgery can also be used to recognize similar tools in a different type of surgery. Important is that the tools have the same design. It was found that the most important resources for a project like this are a dedicated hardware with image recognition software and time. This thesis showed the applicability of a deep learning network to automatically recognize the progress of a surgery and provided insight into the steps that need to be taken to use it on a larger scale.

The gap between predicted brain age and chronological age could serve as biomarker for early-stage neurodegeneration and as potentially as a risk indicator for dementia. We assess the utility of this age gap as a risk biomarker for incident dementia in a general elderly population. The brain age is estimated from longitudinal brain magnetic resonance imaging (MRI) data using deep learning models. From the population-based Rotterdam Study, 5656 dementia-free and stroke-free participants (mean age 64.67±9.82, 54.73% women) underwent brain MRI at 1.5T, including three-dimensional (3D) T1-weighted sequence, between 2006 and 2015. All participants were followed for incident dementia until 2016. During 6.66±2.46 years of follow-up, 159 subjects developed dementia. The entire dataset was split into control (N=5497) and incident dementia (N=159) groups. We then built a convolutional neural network (CNN) model trained on the control group to predict brain age based on brain MRI. Model prediction performance was measured in mean absolute error MAE=4.45±3.59 years of brain age prediction. Reproducibility of prediction was tested using the intra-class correlation coefficient ICC=0.97 (95% confidence interval CI=0.96-0.98), computed on a subset of 80 subjects. Hereafter, we investigated the gap between model predicted age and chronological age of the incident dementia group data, compared to control group. Logistic regressions and Cox proportional hazards models were used to assess the association of the age gap with incident dementia, adjusted for years of education, ApoE4 allele carriership, GM and intracranial volume. These models showed that the age gap was significantly related to incident dementia (odds ratio OR=1.11 and 95% confidence intervals CI=1.05-1.16; hazard ratio HR=1.11 and 95% CI=1.06-1.15, respectively). Additionally, we computed the attention maps of CNN, which shows the importance of brain regions for age prediction. These were particularly focused on the amygdalae and hippocampi. We show that the gap between predicted and chronological brain age is a biomarker, associated with a risk of dementia development. This suggests that it can potentially be used as a complimentary biomarker for early-stage dementia risk screening.