Knowledge-Based Deformable Matching for Pathology Detection

Abstract

Image registration is the process of matching corresponding features between images. Medical images present a challenge for registration because the same anatomical features vary considerably in appearance across individuals, and any pathology may aggravate the variation. We propose a registration method that uses knowledge of anatomy to cope with variations in the appearance of corresponding features and to help detect pathologies. The same anatomical feature may have a different density, shape, and size across individuals, but there exists a normal range of variation. Various clinical conditions may affect the gross anatomical morphology, but certain pathologies have characteristic impacts on the anatomy. Traditional registration methods do not seek guidance from this domain knowledge, and they fail to yield accurate registrations because medical images have complicated shapes, non-uniform textures, and ill-defined boundaries. Our algorithm will attempt to apply domain knowledge in the registration process to cope with the variations in appearance. We are especially interested in the registration of images showing pathologies because of its medical importance. In pursuit of this, we have developed a hierarchical registration algorithm that permits variations in appearance. We will infuse domain knowledge to discern normal and abnormal variations in appearances. The hierarchical registration algorithm has three levels. First a 3-D transformation (rotation, uni-form scaling, and translation) globally aligns the different image sets. Secondly a smooth 3-D deformation approximately matches the anatomical structures in the image sets. This attempts to address the inherent differences in the appearances of their anatomical structures. Finally, the match is refined by a free-form deformation to adjust to the small variations overlooked by the smooth deformation. We have developed this algorithm and conducted experiments on real data sets with and without pathologies. The results from the above registration algorithm are promising, but the algorithm has some limitations. We intend to extract and encode domain knowledge (e.g. using Principal Component Analysis), and apply it to guide the registration process (only permitting normal variations or characteristic impact caused by certain pathologies). To handle intensity variation, we plan to use mutual information as the matching criterion. We will establish a standard validation method and test database to quantitatively evaluate the accuracy of our algorithm and related work. The system will help in detecting lesions, observing the development of pathologies over time, indexing and retrieval in medical databases, conducting cross-patient analysis, and studying the functionality of different parts of the brain.