This thesis aims at developping new registration and segmentation methods for liver perfusion images. These methods are developed in a unified and efficient framework based on large deformation registration. The new registration method incorporates an incompressibility constraint and captures non-rigid deformations of large amplitude. The incompressibility constraint preserves the volume of intensity-enhanced structures due to the contrast agent injection. Modeling the liver as an incompressible organ (Yin et al., 2004) helps the motion correction step, where we verify that the volume of contrast-enhanced structures does not change after the registration process. The new segmentation method is based on large deformation registration. A binary template, that represents a model of the liver, is aligned to an image. A new region-based similarity measure is derived. This allows for a global regularization on the liver shape and the preservation of its topology. It is more robust to extract the liver boundary and it is possible to segment irregular shapes without leaks of the contour in other organs. The constrained registration and the automatic segmentation methods are implemented to be compatible with the physician's time constraints. They use low-complexity algorithms to accelerate data processing and maximize the use of computer resources.;To quantify the accuracy of the segmentation method, we use five standard metrics. We compare the results of our segmentation to manual reference segmentations. We evaluated 30 computed tomography images (images composed up to 512 × 512 × 500 voxels), provided by the workshop "3D Segmentation in the Clinic: A Grand Challenge". Through this workshop, we are able to compare the results of our segmentation with the results of segmentation methods developed by other recent methods. The results of our segmentations fall into third position in relation to methods of the state of the art (van Ginneken et al., 2007). Unlike contour-based segmentation techniques, this framework uses a global regularization of the template, and allows us to segment irregular shapes while avoiding leaks. Comparing with shape based methods, we have segmentations that are more accurate. We also propose a, new multi-label segmentation, that maximizes the likelihood of intensity distributions of different regions to segment. Preliminary results are very satisfactory.;This thesis develops new techniques for motion correction and boundary extraction of liver perfusion images. The registration with the incompressibility constraint and the automatic segmentation are robust and help doctors to make a better diagnosis and to improve the therapeutic planning of liver cancers. These new methods are efficient and allow us to use the registration and segmentation methods in a clinical environment where the time frame is compatible with physicians needs. (Abstract shortened by UMI.).;Properties of the registration with the incompressibility constraint are studied using synthetic data and four liver perfusion studies. These studies represent large amounts of data (images composed up to 512 × 512 × 220 voxels). A visual inspection of substracted aligned images and an inspection of the recovered transformation are used for validation. We also compute the size of pathological regions before and after the constrained registration to quantify relative volume differences in absolute value. The registration results prove that our approach is robust and improves the capture range of large deformations. Relative volume differences in absolute value of tumors before and after the contrained registration do not exceed 2.4%. This method prevents the shrinkage or expansion of contrast-enhanced regions, a phenomenon typically observed with standard fluid methods. Unlike existing methods (Haber et Modersitzki, 2005), this work dissociates the incompressibility constraint from the regularization, allowing us to deal with the large-scale problems in a reasonable time. This makes our approach pratical for perfusion studies in the clinical environment.
展开▼
