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.

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.

Cardiovascular diseases and stroke are currently the leading causes of death worldwide. Atherosclerotic plaque is a mostly asymptotic vascular disease, but rupture of an atherosclerotic plaque in the carotid artery could lead to stroke. Automated segmentation of plaque components could help improve risk assessment by producing fast and reliable results while saving costs.

In this thesis two extensive comparisons have been made. First supervised classifiers are compared in the pixel-wise segmentation task of plaque components. In this comparison five conventional machine learning techniques and one deep learning architecture have been evaluated: linear and quadratic Bayes normal classifiers, linear logistic classifier, random forest and a U-net architecture. In the second comparison classifiers are evaluated in a detection task for their ability to learn with weakly labelled data. This is done within the multiple instance learning (MIL) framework. In addition to conventional multiple instance learning algorithms, a new MIL adaptation of the deep learning architecture, MIL U-net, is proposed and evaluated.

In the pixel-wise segmentation tasks the U-net architecture was the best overall classifier after the addition of 93 extra training patients to the original 20 training patients. A good inter-rater agreement was found for the haemorrhage class (ICC = 0.684) and the calcification class (ICC = 0.627). In the detection task the supervised methods, trained with one-sided noise, outperformed multiple instance classifiers such as MIL-Boost and the proposed MIL U-net. In this task both random forest and the linear logistic classifier obtained a fair Cohen's kappa (0.419 and 0.445 respectively) for detection of calcification per slice. The same classifiers obtained good correlation (Cohen's kappa 0.717 and 0.666 respectively) for haemorrhage detection per slice.