M-reps Deformable Model And Application In Heart Segmentation

German physicist Wilhelm Conrad Rontgen discovered X-rays in 1895, for thefirst time making non-invasive inspection of human tissue possible, it was used as away of medical imaging only after months of its discovery. In the past decades, a lotof new imaging techniques have been developed and matured, offering unprecedentedhelp to clinical medicine. After acquiring the images, post-processing and analysisof them gives the direct basis of medical diagnosis. For years a variety of medicalimaging processing techniques have been proposed, and some of them have developedinto their own independent subjects. Nowadays, research on medical image processingis taking place in many laboratories wordwide.Medical image segmentation is an important field in image processing and anal-ysis, and it's the basis for computer assisted diagnosis and clinical treatment. Im-age segmentation has important sense for medical image processing, because manylater processes like structure analysis, movement analysis, 3D visualization, assistedsurgery, radiation therapy all make the assumption that the image has already been cor-rectly segmentated. Image segmentation is a fundamental problem, and at the sametime it's a classic hard problem. In the past two decades a variety of ideas and tech-niques appeared in this field.Recently deformable model based methods have gradually developed into one ofthe most active and successful field in image segmentation. Terms like active con-tour, deformable contour and deformable surface mentioned in papers all belong todeformable model. The basic idea of deformable model is to set up an energy func-tion, and make the contour move under the internal model force and the external imageforce to optimize the energy function, at the time of optimization the contour will con-verge to the boundary of target ROI. In this paper, we review the classic snakes deformable model, and some extendedmodels, emphasise on their basic theory and mathematics. Deformable model basedmethods have the advantage that image data, initialization and target contour are allunified into a feature extraction process, and the segmentation mechanism can utilizethe high level information (e.g., information from human-computer interaction andstatistical model) to guide the low level segmentation, and after proper initialization,the model is able to converge to the state of minimum energe. Deformable modelbased methods can not only maintain the integrity of the structure being segmentated,but also agree with the image feature.In this paper, we introduce a recently proposed deformable model, m-reps. Usingthe result of manual segmentation of heart data we got from hospital, we build a pa-rameterized m-reps deformable model, with this model we conducted 3 experimentsof segmentation using different dataset. Segmentations using m-reps work like this:first, we manually put the m-reps model into the heart data (initialization), followed bya series of coarse-to-fine deformations with a goal of optimization of an object energyfunction. The object function includes a geometric typicality term and a model-imagematch term. Transforming the convergence state of the model to the boundary repre-sentation, we get the segmentation result.In this paper we conduct scientific analysis of the segmentation result, compare itwith the results of manual segmentation. M-reps deformable model shows its advan-tage in its validity as an object representation descriptor, its stability in deformationprocess and its convenience in computation.In this paper, we also propose a new method to measure the significance of medialaxis branch, an important step in m-reps modeling process. Inspired from the fact thatpeople perceive shapes largely based on areas of them, this new method is based on theratio of areas of Delaunay triangles. The new method possesses a few good propertiesincluding being able to always preserve the topology structure of the shape in theprocess of threshold-based pruning. We also demonstrate the validity of our methodusing experiments of different shapes.