Author:

Keywords:

PSI_MIC

Abstract:

Many neurodegenerative diseases such as Alzheimer’s disease can be characterized by their gradual modification of the cellular environment resulting in macroscopic brain changes. Groupwise analysis of large sets of magnetic resonance (MR) images, visualizing the morphology of the human brain, can provide reliable measurable image features indicative for a specific disease and/or disease stage. Such features, called biomarkers, can contribute to early diagnosis, to the assessment of therapy response as well as to the study of the underlying causes of the neurodegenerative disease. In this PhD, we developed a Bayesian framework for both cross-sectional and longitudinal groupwise analysis of a collection of brain MR images. Contrary to most state-of-the-art strategies for cross-sectional analysis, which compare homogeneous subgroups of images selected based on prior clinical knowledge, our framework can handle heterogeneous collections of images, for instance of both normal controls as well as patients in different disease states. The proposed framework simultaneously performs segmentation of each image into the major tissue classes (white matter, gray matter, CSF) and automatically clusters all images of the set in homogeneous subgroups based on the image morphology extracted from the image segmentations. A probabilistic brain atlas is iteratively constructed for each cluster within the framework. These atlases visualize the mean morphology of the subgroup and support the clustering and segmentation processes. The constructed morphological subgroups can be correlated with the clinical diagnosis of each subject, while the clustering process reveals the distinctive cluster-specific image features globally and for each individual image. The framework is adapted to take additional clinical knowledge into account in the clustering process if desired. The longitudinal framework is a direct extension of the cross-sectional framework, analyzing multi-temporal brain MR image sequences of different subjects suffering from the same neurodegenerative disease, aiming to assess both the patient-specific and the population-wise disease evolution based on the structural changes over time. The clustering process reveals the subject-specific disease evolution, while the constructed atlases form a longitudinal atlas representing the temporal changes in brain morphology caused by the disorder. A proof-of-concept of the framework is demonstrated on synthetic images and its feasibility is further shown on brain MR images from different publicly available data sets, including the simulated BrainWeb data set as well as the large international initiatives ADNI and OASIS. The experiments show that our framework provides accurate, more unbiased (and temporally consistent) image segmentations and that it can detect the clinically relevant morphological subgroups (cross-sectionally) or the individual disease progression (longitudinally) in a set of images or image sequences. The constructed atlases indicate the disease (stage) morphology better than traditional tools for cross-sectional and longitudinal analysis based on clinical prior knowledge. Finally, it is illustrated that our combinational approach for segmentation and clustering can find distinctive image features that would not have been picked up by handling both processes separately and subsequently. In conclusion, the presented method combines segmentation, clustering and atlas construction in a unified probabilistic framework such that all techniques can cooperate and such that the method becomes more data-driven. We provide initial evidence that our method facilitates a better representation of the disease evolution, the detection of novel subgroups and the detection of novel disease-specific features. Therefore, our method can become an important tool for the development of novel and refined imaging biomarkers, providing new insights in structural change and development.