Isolating an anatomical or pathological structure in a medical image is a task performed every day by hospital practitioners. In many cases, however, this task proves incredibly complex when it comes to precisely delineating this structure in the image, to the extent that the contours traced by two experts can vary significantly. This difference in estimation may not be significant for a quick visual inspection of medical images, but this lack of reproducibility and the time required to isolate each structure is a major obstacle to a more quantitative and personalized practice of medicine.
This is why, for almost 20 years, image segmentation algorithms have been developed to detect and isolate anatomical structures, often semi-automatically to take advantage of the user's expertise. These algorithms incorporate and combine explicitly or implicitly information on the appearance and shape of organs, as well as on the anatomical relationships between these structures. The existence of a characteristic shape for a large number of organs is the basis of computational anatomy, and is certainly an important feature for recognizing and isolating them in a medical image.
The aim of this presentation is to explore different ways of incorporating organ geometry to segment medical images. Thus, when a simple assumption of geometric regularity of shape is considered, this can be translated into a constraint of connectedness between the nodes of a graph. On the other hand, when the existence of an average shape is assumed, more complex geometric representations are required in the form of implicit functions or meshes. The choice of this shape representation is important, and the use of simplex meshes for organ reconstruction will be particularly illustrated. Finally, the presentation will conclude with a discussion of the subtle relationship between shape and image and its formulation from a deterministic paradigm to a probabilistic framework.
