Automated 3D segmentation using deformable models and fuzzy affinity

Abstract

We have developed an algorithm for segmenting objects with closed, non-intersecting boundaries, such as the heart and the lungs, that is independent of the imaging modality used (e.g., MRI, CT, echocardiography). Our method is automatic and requires as initialization a single pixel/voxel within the boundaries of the object. Existing segmentation techniques either require much more information during initialization, such as an approximation to the object's boundary, or are not robust to the types of noisy data encountered in the medical domain. By integrating region-based and physics-based modeling techniques we have devised a hybrid design that overcomes these limitations. In our experiments we demonstrate across imaging modalities, that this integration automates and significantly improves the object boundary detection results. This paper focuses on the application of our method to 3D datasets.