Atlas-based segmentation and classification of magnetic resonance brain images

A wide range of different image modalities can be found today in medical imaging. These modalities allow the physician to obtain a non-invasive view of the internal organs of the human body, such as the brain. All these three dimensional images are of extreme importance in several domains of medicine, for example, to detect pathologies, follow the evolution of these pathologies, prepare and realize surgical planning with, or without, the help of robot systems or for statistical studies. Among all the medical image modalities, Magnetic Resonance (MR) imaging has become of great interest in many research areas due to its great spatial and contrast image resolution. It is therefore perfectly suited for anatomic visualization of the human body such as deep structures and tissues of the brain. Medical image analysis is a complex task because medical images usually involve a large amount of data and they sometimes present some undesirable artifacts, as for instance the noise. However, the use of a priori knowledge in the analysis of these images can greatly simplify this task. This prior information is usually represented by the reference images or atlases. Modern brain atlases are derived from high resolution cryosections or in vivo images, single subject-based or population-based, and they provide detailed images that may be interactively and easily examined in their digital format in computer assisted diagnosis or intervention. Then, in order to efficiently combine all this information, a battery of registration techniques is emerging based on transformations that bring two medical images into voxel-to-voxel correspondence. One of the main aims of this thesis is to outline the importance of including prior knowledge in the medical image analysis framework and the indispensable role of registration techniques in this task. In order to do that, several applications using atlas information are presented. First, the atlas-based segmentation in normal anatomy is shown as it is a key application of medical image analysis using prior knowledge. It consists of registering the brain images derived from different subjects and modalities within the atlas coordinate system to improve the localization and delineation of the structures of interest. However, the use of an atlas can be problematic in some particular cases where some structures, for instance a tumor or a sulcus, exists in the subject and not in the atlas. In order to solve this limitation of the atlases, a new atlas-based segmentation method for pathological brains is proposed in this thesis as well as a validation method to assess this new approach. Results show that deep structures of the brain can still be efficiently segmented using an anatomic atlas even if they are largely deformed because of a lesion. The importance of including a priori knowledge is also presented in the application of brain tissue classification. The prior information represented by the tissue templates can be included in a brain tissue segmentation approach thanks to the registration techniques. This is another important issue presented in this thesis and it is analyzed through a comparative study of several non-supervised classification techniques. These methods are selected to represent the whole range of prior information that can be used in the classification process: the image intensity, the local spatial model, and the anatomical priors. Results show that the registration between the subject and the tissue templates allows the use of prior information but the accuracy of both the prior information and the registration highly influence the performance of the classification techniques. Another aim of this thesis is to present the concept of dynamic medical image analysis, in which the prior knowledge and the registration techniques are also of main importance. Actually, many medical image applications have the objective of statically analyzing one single image, as for instance in the case of atlas-based segmentation or brain tissue classification. But in other cases the implicit idea of changes detection is present. Intuitively, since the human body is changing continuously, we would like to do the image analysis from a dynamic point of view by detecting these changes, and by comparing them afterwards with templates to know if they are normal. The need of such approaches is even more evident in the case of many brain pathologies such as tumors, multiple sclerosis or degenerative diseases. In these cases, the key point is not only to detect but also to quantify and even characterize the evolving pathology. The evaluation of lesion variations over time can be very useful, for instance in the pharmaceutical research and clinical follow up. Of course, a sequence of images is needed in order to do such an analysis. Two approaches dealing with the idea of change detection are proposed as the last (but not least) issue presented in this work. The first one consists of performing a static analysis of each image forming the data set and, then, of comparing them. The second one consists of analyzing the non-rigid transformation between the sequence images instead of the images itself. Finally, both static and dynamic approaches are illustrated with a potential application: the cortical degeneration study is done using brain tissue segmentation, and the study of multiple sclerosis lesion evolution is performed by non-rigid deformation analysis. In conclusion, the importance of including a priori information encoded in the brain atlases in medical image analysis has been put in evidence with a wide range of possible applications. In the same way, the key role of registration techniques is shown not only as an efficient way to combine all the medical image modalities but also as a main element in the dynamic medical image analysis.