In clinical practice, as a first approximation, the severity of an abnormality on an image is often determined by measuring its volume. Researchers often first segment this abnormality with a neural network trained by voxel-wise labels and thereafter extract the volume. Instead of this indirect two steps approach, we propose to train neural networks directly using the volumes as image-level label and predict the volume directly. Using image-level labels to train automatic abnormality prediction could decrease the labeling burden for clinical experts, which is both expensive and time consuming. In this report, a neural network that consisted of a segmentation part and an appended regression part was compared with the indirect segmentation approach. It was investigated if networks trained with image-level labels have the same performance of image-level prediction as networks trained with voxel-wise labels. The neural networks were trained on a large local dataset to quantify white matter hyperintensity (WMH) burden from brain MRI, and their performance was evaluated on a held-out test set. Furthermore, generalization properties were compared by applying the trained networks on four independent public datasets. The networks trained with image-level labels achieved volume quantification that was slightly better than their counterpart on the held-out test set. The attention maps of these networks showed that the networks were able to focus on the surroundings of the WMH, and hence learned meaningful image features. Nevertheless, the attention maps were not suitable to achieve a compatible segmentation. In terms of generalization towards external datasets, the advantage of weak labels for volume quantification did not hold as there was no significant difference between the performance of the label types. The results suggest that neural networks optimized with image-level labels were able to directly predict WMH volume as well as neural networks trained with voxel-wise labels. Subsequently, we also studied networks that were optimized on both image-level and voxel-wise labels. Those networks reached a lower performance, which suggested that the tasks and their image features learned were not similar enough.

Magnetic Resonance Imaging is a popular modality for brain imaging in present times. The quality of the images depends on the strength of the magnetic field. An MRI scanner with a magnetic field strength of 3 Tesla(T) is pre-dominantly used for clinical purposes. However, with the advancement in technology, and the need to image finer image finer structures, we are gradually shifting to higher magnetic field strengths like 7T and above. One of the major bottlenecks in these systems is the bias induced in the images due to field inhomogeneities of higher field strengths. Yet another drawback of these systems is the increase in power dissipation in the tissues. This is measured by a quantity called Specific Absorption Rate(SAR). The goal of this thesis is to accurately predict SAR values for various volunteers as they may differ from subject to subject. Many homogeneous models have been created earlier to estimate the value of SAR, however, these estimates are often over-conservative and safe which compromises with image quality. A personalised numerical body model is created for all volunteers using relevant information derived from simulations performed on generalized models. Various tissues have to be segmented to create these 3D numerical body models. However, there has to be a trade-off between ease of segmentation and SAR accuracy. Keeping that in mind, it was found that the optimum number of tissues to get reliable SAR estimates is six. A deep learning method was then used for segmentation. A numerical body model was derived for all the volunteers using the deep learning segmentation. An adapted ForkNet, which is similar to U-net in architecture is used to segment these images. The SAR values derived from the predicted numerical body model and the original body model are similar, hence speeding up the process for SAR prediction. However, there are certain limitations of the thesis that can be addressed in the future. Inadequate data remains a major bottleneck for the project, increasing the data should result in improved segmentation, this can be addressed by acquiring more data. Another major drawback of the thesis is the segmentation accuracy, the ground truth segmentation is performed to the best of our knowledge, however, some errors are still present in the ground truth segmentation. The next steps for this project would be to acquire more data, train the data on multiple input sequences, use a 3D network and using localizer images that are acquired at the start of the MRI scan. Nonetheless, the principles established in this thesis confirm that a deep learning approach can be used to create numerical body models for SAR estimates. It also establishes the fact that these SAR estimates are comparable to the SAR estimates generated from the ground truth numerical body models.

Image registration is a fundamental requirement for many medical applications. In recent years, deep learning approaches for registration have shown to be a promising alternative to conventional methods. However, most learning based methods do not consider the different physical properties of various tissues, which can result in unrealistic deformation in anatomical regions where both deformable tissue and rigid bone is present. In this work, we develop and evaluate deep learning methods for intrapatient CT-MR registration while maintaining rigidity of the bones. Unconstrained and locally constrained registration methods are compared in an unsupervised and weakly-supervised setting. The results show that qualitatively and quantitatively accurate registrations can be obtained.

Perivascular spaces (PVS) visible on MRI are currently emerging as an important potential neuroimaging marker for several pathologies in the brain like Alzheimer’s disease and cerebral small vessel disease. PVS are fluid-filled spaces surrounding vessels as they enter the brain. Although PVS are normally not noticeable on MRI scans acquired at clinical field strengths, when these spaces increase in size they become increasingly visible and quantifiable. To study these spaces it is important to have a robust method for quantifying PVS. Manual quantification of PVS is challenging, time-consuming and subject to observer bias due to the difficulty of distinguishing PVS from mimics and the large number of PVS that can occur in MRI scans. Many promising (semi-)automated methods have been proposed recently to decrease annotation time and intra- and inter-observer variability while providing more information about EPVS. However there are still various limitations in the current methods that need to be overcome. An important limitation is that most of the methods are based on elaborate preprocessing steps, feature extraction and heuristic fine-tuning of parameters, making the use of these methods on new datasets cumbersome. Furthermore the majority of the currently proposed methods have been evaluated on small sets of barely 30 images, as most of these methods aim to segment PVS and require voxel-wise annotations for evaluation. In this thesis we propose a method for automated detection of perivascular spaces that combines a convolutional neural network and geodesic distance transform (GDT). We propose to use dot annotations instead of voxel-wise segmentations as this is less time-consuming than fully segmenting PVS while still providing the location of PVS. This enables us to use a considerably larger dataset with ground truth locations than is used in all previously proposed (semi-)automatic methods that provide the location of PVS. We investigated two approaches of using geodesic distance transform to optimize the CNN to detect PVS. The first approach focuses on optimizing the CNN for voxel-wise regression of the geodesic distance map (GDM) computed from the dots and the intensity image. The second approach aims to predict segmentations of the PVS using a CNN that is trained on approximated segmentations obtained by thresholding GDMs. We use 1202 proton density-weighted (PDw) MRI scans to develop our methods and 1000 other scans are used to evaluate the performance of the methods. We show that our methods match human intra-rater performance on detecting PVS without the need for any user interaction. Additionally we show that GDMs are extremely useful for capturing complex morphologies when computed from dot annotations. Our experiments indicate that GDMs can be used to provide valuable additional information to CNNs during training.

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.

The amount of personal imagery kept on (mobile) devices is increasing by the day. Analysis and organization of these large collections of data are becoming increasingly important in the field of digital forensics, as they can aid in the search for legal evidence. The grouping of faces based on their identity is an important aspect as it provides an overview of the person in question and their connection with scenes, objects and other people. In this work, we propose a fuzzy approach to the hard partitioning problem of face clustering for the specific field of forensic investigations. We constructed a pipeline consisting of deep models for face detection and feature extraction, a method for transforming the resulting feature vectors to a graph representation and a graph-based clustering algorithm for the final partitioning. Focusing on the clustering step, we propose to assign face images to identity clusters using confidence values (rather than a hard cutoff) based on the average similarity with images present in the cluster relative to other clusters. Compared to existing methods, the approach is not only fuzzy but also embraces na¨ıve linking, and instead of transitively merging the links it uses a graph-based algorithm to produce the clusters. Furthermore, we propose an adapted version of the MaxMax algorithm because the original method only returned fuzzy results if weights were exactly equal. However, similarities between images are continuous, making it unsuitable for the case of face clustering. Evaluation of the performance on the Labeled Face in the Wild (LFW) dataset and the challenging IARPA JANUS Benchmark B (IJB-B) shows promising results comparable with state-ofthe-art face clustering algorithms.