High performance computing for 3D image segmentation

• Main Author: Lenkiewicz, Przemyslaw
• Publication Date: 2013
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10400.6/3983
• topic: Processamento de imagem digital; Segmentação de imagens 3D - Métodos; Segmentação de imagens 3D - Modelos; Segmentação de imagens - Modelos deformáveis; Segmentação de imagens - Whole Mesh Deformation model (WMD)

Digital image processing is a very popular and still very promising eld of science, which has been successfully applied to numerous areas and problems, reaching elds like forensic analysis, security systems, multimedia processing, aerospace, automotive, and many more. A very important part of the image processing area is image segmentation. This refers to the task of partitioning a given image into multiple regions and is typically used to locate and mark objects and boundaries in input scenes. After segmentation the image represents a set of data far more suitable for further algorithmic processing and decision making. Image segmentation algorithms are a very broad eld and they have received signi cant amount of research interest A good example of an area, in which image processing plays a constantly growing role, is the eld of medical solutions. The expectations and demands that are presented in this branch of science are very high and dif cult to meet for the applied technology. The problems are challenging and the potential bene ts are signi cant and clearly visible. For over thirty years image processing has been applied to different problems and questions in medicine and the practitioners have exploited the rich possibilities that it offered. As a result, the eld of medicine has seen signi cant improvements in the interpretation of examined medical data. Clearly, the medical knowledge has also evolved signi cantly over these years, as well as the medical equipment that serves doctors and researchers. Also the common computer hardware, which is present at homes, of ces and laboratories, is constantly evolving and changing. All of these factors have sculptured the shape of modern image processing techniques and established in which ways it is currently used and developed. Modern medical image processing is centered around 3D images with high spatial and temporal resolution, which can bring a tremendous amount of data for medical practitioners. Processing of such large sets of data is not an easy task, requiring high computational power. Furthermore, in present times the computational power is not as easily available as in recent years, as the growth of possibilities of a single processing unit is very limited - a trend towards multi-unit processing and parallelization of the workload is clearly visible. Therefore, in order to continue the development of more complex and more advanced image processing techniques, a new direction is necessary. A very interesting family of image segmentation algorithms, which has been gaining a lot of focus in the last three decades, is called Deformable Models. They are based on the concept of placing a geometrical object in the scene of interest and deforming it until it assumes the shape of objects of interest. This process is usually guided by several forces, which originate in mathematical functions, features of the input images and other constraints of the deformation process, like object curvature or continuity. A range of very desired features of Deformable Models include their high capability for customization and specialization for different tasks and also extensibility with various approaches for prior knowledge incorporation. This set of characteristics makes Deformable Models a very ef cient approach, which is capable of delivering results in competitive times and with very good quality of segmentation, robust to noisy and incomplete data. However, despite the large amount of work carried out in this area, Deformable Models still suffer from a number of drawbacks. Those that have been gaining the most focus are e.g. sensitivity to the initial position and shape of the model, sensitivity to noise in the input images and to awed input data, or the need for user supervision over the process. The work described in this thesis aims at addressing the problems of modern image segmentation, which has raised from the combination of above-mentioned factors: the signi cant growth of image volumes sizes, the growth of complexity of image processing algorithms, coupled with the change in processor development and turn towards multi-processing units instead of growing bus speeds and the number of operations per second of a single processing unit. We present our innovative model for 3D image segmentation, called the The Whole Mesh Deformation model, which holds a set of very desired features that successfully address the above-mentioned requirements. Our model has been designed speci cally for execution on parallel architectures and with the purpose of working well with very large 3D images that are created by modern medical acquisition devices. Our solution is based on Deformable Models and is characterized by a very effective and precise segmentation capability. The proposed Whole Mesh Deformation (WMD) model uses a 3D mesh instead of a contour or a surface to represent the segmented shapes of interest, which allows exploiting more information in the image and obtaining results in shorter times. The model offers a very good ability for topology changes and allows effective parallelization of work ow, which makes it a very good choice for large data-sets. In this thesis we present a precise model description, followed by experiments on arti cial images and real medical data.