%0 Journal Article
%T Semiautonomous Medical Image Segmentation Using Seeded Cellular Automaton Plus Edge Detector
%A Ryan A. Beasley
%J ISRN Signal Processing
%D 2012
%R 10.5402/2012/914232
%X Segmentations of medical images are required in a number of medical applications such as quantitative analyses and patient-specific orthotics, yet accurate segmentation without significant user attention remains a challenge. This work presents a novel segmentation algorithm combining the region-growing Seeded Cellular Automata with a boundary term based on an edge-detected image. Both single processor and parallel processor implementations are developed and the algorithm is shown to be suitable for quick segmentations (2.2£¿s for 2 5 6 ¡Á 2 5 6 ¡Á 1 2 4 voxel brain MRI) and interactive supervision (2¨C220£¿Hz). Furthermore, a method is described for generating appropriate edge-detected images without requiring additional user attention. Experiments demonstrate higher segmentation accuracy for the proposed algorithm compared with both Graphcut and Seeded Cellular Automata, particularly when provided minimal user attention. 1. Introduction Segmentation, also known as labeling, of medical images is an important task for quantitative analyses, custom intervention planning such as localized radiotherapy, and design of patient-specific tools such as orthotics or jigs for joint replacement. Manually labeling images takes a prohibitive amount of time due to the large number of voxels in most medical images, while autonomous segmentations can fail to reach the required accuracy due to interpatient morphological variability. Supervised segmentation algorithms are a promising solution because they allow the user to guide the segmentation without requiring the user¡¯s attention for each voxel. Popular supervised segmentation algorithms include active contours (snakes) [1], Level Sets [2], intelligent scissors (live wire) [3, 4], Graphcut [5], and to some degree Seeded Cellular Automata (SCA) [6]. For active contours and Level Sets, the user initializes the boundary near the desired contour and the algorithm moves the boundary to a local minimum determined by an energy functional. This approach requires that the user solve two optimization problems: setting the impact of the terms in the functional, and placing the initial boundary such that the results settle into a desirable minimum [7]. Intelligent scissors connect-the-dots between user-placed boundary points, but either image noise or inaccurate placement of those points can generate inaccurate segmentations. Graphcut uses graph-based operations to separate user-placed ¡°seeds¡± so as to satisfy the max-flow/min-cut theorem, but demonstrates a bias toward small segmentations and is inherently a two-label approach (with
%U http://www.hindawi.com/journals/isrn.signal.processing/2012/914232/