Glioma progression is monitored by routine MR scanning, enabling tumor growth evaluation with respect to earlier time-points. This growth may present both as a mass effect and as an extension of abnormalities into previously healthy tissue. To accurately quantify tumor growth and tumor-induced deformations, longitudinal intrasubject image registration is often used. However, such registration in cases with large deformations and tissue change is highly challenging. Longitudinal image registration may benefit from groupwise strategies in which multiple images are concurrently aligned. This avoids introducing bias towards an a priori-selected reference image. However, existing learning-based methods for image registration mostly concern pair-wise approaches. Moreover, the few proposed learning-based methods for groupwise registration are designed for the analysis of images without pathologies and are prone to fail to register glioma images. To bridge this gap, we present a learning-based method for the non-linear registration of longitudinal glioma images. We adapt an existing learning-based groupwise method to handle tumor infiltration by means of cost-function masking. The proposed method is able to register glioma images despite the presence of non-correspondences across the time-points by focusing on the normal-appearing tissue similarity. We train the framework both in one resolution and with a multi-stage strategy exploring multiple resolutions. We evaluate on a dataset from the Glioma Longitudinal AnalySiS consortium and compare it to conventional groupwise registration methods. We achieve comparable Dice coefficients, with higher SSIM and more detailed registrations. These evaluation metrics are further improved when trained as a multi-stage method. The proposed framework preserves the diffeomorphic conditions and the geometric centrality of the deformation fields, while significantly reducing the runtime to under a minute. The proposed methods may serve as an alternative to conventional toolboxes to provide further insight into glioma growth.

Many machine learning models have been developed to aid in the diagnosis of dementia, to predict dementia risk and to determine cognitive performance. While it is well known that vascular pathology is a critical contributing factor to dementia, cerebral small vessel disease is often not addressed by these methods, possibly causing impeded generalizability of the models from research setting to clinical practice. The aim of this thesis is to evaluate whether vascular pathology, as represented by white matter hyperintensities (WMHs), influences the outcome of machine learning prediction models for dementia and as such hampers generalizability of these models from the research setting to clinical practice. To answer this research question, we developed a convolutional neural network (CNN) for prediction of cognitive performance measured by MMSE and ADAS13. We used ADNI data and included all subject timepoints that adhered to our inclusion criteria (N = 4846). Next, we derived gray matter density maps (GMDMs) and WMH density maps (WMHDMs) from T1 and FLAIR MRI scans using voxel-based morphometry. Six CNN models were implemented, differing in input: model 1 (GMDM), model 2 (GMDM, age, sex), model 3 (WMHDM), model 4 (WMHDM, age, sex), model 5 (GMDM, WMHDM, age, sex) and model 6 (GMDM, log(WMH ratio), age, sex). WMH ratio was defined as WMH volume corrected for intracranial volume. We split the data into a low, middle and high WMH load group based on WMH ratio tertiles and performed two sets of experiments. First, we trained all models on the low WMH load group with the intention of testing on the high WMH load group as a measure of generalizability. Second, we trained model 2 and 5 on data from all WMH load groups. For these sets of experiments, we split both the low WMH load group and all WMH load groups into a train, validation and test set. The models trained on the low WMH load group obtained a similar performance to prediction of a constant value and prediction by a model developed on randomized cognitive scores. These models likely suffered from the underrepresentation of lower cognitive performance data (i.e. high ADAS13 scores and low MMSE scores). This is supported by the fact that model 2 and 5 outperformed constant value prediction when trained on data from all WMH load groups. The current models did in general not benefit from adding WMHDM input to the model, while they did benefit slightly from log(WMH ratio) information. However, we were unable to answer our research question due to the suboptimal performance of the models. Our future research will focus on revisiting our research question through brain age prediction in order to improve generalizability of ML methods for AD to clinical populations with prominent vascular pathology.

To optimize treatment in patients with craniosynostosis, better understanding of the disease process is essential, for example using brain MRI analysis to study brain volume, brain perfusion, and brain micro-architecture. However, such analyses require image registration, which is challenging because of disease-related brain deformations. Therefore, the aim of this project is to optimize image registration for children with syndromic craniosynostosis, aged 0 to 6 years old. We compared conventional and deep learning registration methods in a quantitative evaluation using synthetic data (i.e. deformed atlases) and in a qualitative experiment using registration of atlas scans to craniosynostosis scans. In addition to comparing registration methods, we evaluate the in uence of using both T1-weighted and T2-weighted scans and using an infant or adult atlas. Our qualitative results showed that head shape was registered well by both the conventional and the deep learning registration method, while the deep learning method performed better regarding registration of the ventricles. Quantitatively, our results showed that white matter structures were registered well (Dice: 0.70-0.81). However, regarding registration of the cortical brain regions, both methods resulted in a sub-optimal accuracy (Dice: 0.45-0.63). In general, the approaches of using T2-weighted infant atlases or T1-weighted adult atlases outperformed the alternative approaches. In conclusion, we obtained the best registration result using the deep learning approach, probably as prior spatial information is incorporated in the training process. In addition, we showed that infant atlases based on T2-weighted scans lead to the best results in registration of infant scans.