Understanding three-dimensional images: the recognition of abdominal anatomy from computed axial tomograms (cat)
Uri Shani · 1980
Understanding 3-D images is conceived as the automatic extraction of the 3-D geometry of objects presented therein. This particular approach is employed in analyzing Computed Axial Tomograms (CAT) of the human abdomen. A collection of CAT scans is searched as a 3-D image and matched against a detailed geometrical model of the abdominal anatomy. Detected organ boundaries serve to construct an instance of the model that reflects the anatomical variations of a particular patient as revealed by the scan. The importance of this work is gained from the major role that CAT images played in diagnostic medicine during the last decade. No less important is their growing application in industry, especially for nondestructive inspection of manufactured products. Because manual examination of CAT scans is tedious and time consuming, automatic analysis is a promising means for easing the growing burden on radiology departments, and for increased efficiency in industry. In the medical application field, several efforts have developed systems that employ data-directed approaches by boundary-following and region-growing techniques for semiautomatic detection of organ boundaries. These approaches, however, tend not to work for organs like kidneys, which can have obscured boundaries. In contrast, a model-directed approach, as employed here, overcomes these difficulties by capitalizing on two important characteristics of the problem. First, the digital image is a direct representation of a 3-D domain (i.e., it samples the density of corresponding locations in a [limited] 3-D domain) that is unlike computer vision, in which images are samples of the 2-D projection of the domain (cf. analysis of outdoor scenes or chest radiology). Second, organs in CAT images can be described by geometrical and relational types of knowledge that are usually repeatable from patient to patient. This knowledge is used in a model-directed approach in several ways: for example, knowledge of the gross anatomy and the location of easy-to-find organs (e.g., the spinal column) and help in locating hard-to-find organs (e.g., the kidney). . Key technical contributions of the thesis are the following: • A decentralized organization of knowledge for image understanding by specializing search strategies to the various expected types of boundaries in the image; • A demonstration of a technique for organization and application of 3-D geometrical knowledge to the analysis of 3-D images, based on the use of Generalized-Cylinders; and • A promotion of the technique of Generalized-Cylinders to a degree never tried before, by the use of parametric cubic uniform B-splines. The system is suitable for fast execution, since the hierarchical model is inherently parallel. Other difficulties that result from the huge amount of redundant information presented in the data are managed by concentrating on local searches in the vicinity of expected organ locations, in small subimages, and in a pyramid-of-resolutions data organization. Expected benefits of the system (in…