Unsupervised iterative segmentation and recognition of anatomic structures in medical imagery using second-order B-spline descriptors and geometric quasi-invariants

Abstract

A geometric deformable model is presented for iterative segmentation and recognition of boundaries belonging to anatomic structures in medical imagery. The model utilizes a conventional edge detection algorithm for the extraction of potential boundaries. B-spline descriptors for the boundaries are then calculated. Next, geometric quasi-invariants of the control point sets, describing the B-splines are used to match potential boundaries with that of a prototype template stored in memory. Such a template is part of a novel second-order B-spline prototype templates library where the boundaries of anatomic structures are stored as sets of control points instead of storing the images themselves. The utilization of a control point set for segmentation and recognition reduces computational complexity and improves the accuracy and efficiency of the process. Once a match has been found, segmentation is done again with the parameters of the matching template. Utilizing these parameters minimizes noise and other unwanted features. This model does not suffer from many of the drawbacks associated with other deformable templates and snake models that are currently used, such as computational complexity, user interaction, sensitivity to initial conditions and others. Furthermore, unlike most deformable model templates, this algorithm is not limited to a few images and does not require huge storage space since control point sets are used to describe templates in the library. Experiments performed on medical images confirm the efficiency and robustness of this algorithm.