Registration methods for pulmonary image Analysis: Integration of morphological and physiological knowledge

In this thesis, methods for the registration of pulmonary CT images are developed in which segmentations are used to model lung-specific morphological and physiological aspects of the lung. Various applications in the field of pulmonary image analysis require a registration of CT images of the lung. It is needed, for example, to align different scans of one patient for assessing local ventilation parameters or monitoring disease progression. A registration-based estimation of the breathing motion is employed to increase accuracy in radiotherapy. Moreover, the registration of scans of different patients is required for atlas-based segmentation and motion modeling. While several different registration techniques have been developed in the past, the image content is often regarded as a homogeneous structure and no specific characteristics of the lung are modeled. In this work, the hypothesis is followed that an explicit consideration of morphological and physiological aspects of the lung can considerably improve registration accuracy and plausibility. For this purpose, two approaches are developed in which segmentations of specific organs are employed to integrate such knowledge into the registration algorithm. In the first part, a method for integrated segmentation and registration is presented to explicitly align the pulmonary lobes. It is motivated by the observation that conventional intensity-based registration approaches often expose an insufficient lobe alignment because the interlobular fissures are very low-contrasted in CT images and therefore provide little information for the algorithm. This problem is addressed in two steps: First, an automatic approach for lobe segmentation is presented in which shape information is incorporated by a novel force term. The term relies on a supervised fissure detection and causes an attraction of the contour in direction of the lobe boundaries. Then, the segmentation component is integrated into the registration framework. In this way, an alignment of the fissures is explicitly promoted. In the second part, a segmentation is used to model physiological properties of the breathing motion at the lung boundaries. In this region, respiration causes inner and outer lung pleura to slide along each other, which entails discontinuities in the motion field. Such sliding motion contradicts the smoothing employed in most registration algorithms and entails severe errors in the estimation if not properly accounted for. To remedy this problem, a new segmentation-based regularization approach is presented in which normal- and tangential-directed motion are regarded separately. This allows a physiologically plausible modeling of the sliding motion at the lung boundaries. Both approaches are extensively evaluated using clinical CT images, including publicly available data to allow a comparison with other approaches. The evaluation shows that improvements over state-of-the-art techniques can be obtained using the developed methods. It is concluded that a segmentation-based consideration of lung morphology and physiology is beneficial for registration accuracy and plausibility.