A shape-based model for segmentation of MR brain images

• Main Author: Er, Colin Wen-Jie
• Other Authors: Teoh Eam Khwang
• Format: Final Year Project
• Language: English
• Published: 2009
• Subjects: DRNTU::Engineering
• Online Access: http://hdl.handle.net/10356/18003
• Institution: Nanyang Technological University

Digital image segmentation is an important aspect of digital image processing, particularly in the field of medical image processing. The segmentation of anatomical structures from medical images will provide medical professionals with good visualization of certain region of interest. In this project, the implementation of two approaches, namely the non-parametric and shape-based parametric model, will be investigated. The non-parametric model aims to evolve a signed distance function towards the boundary of the object by manipulating the level set function according to the image data. The parametric model uses an implicit representation of the segmenting curve based on prior information obtained from the training samples. It then manipulates the parameters according to the image data in segmenting the object. Both approaches were obtained using images with simple objects as well as MR brain images. The results achieved were promising. It required only 14 iterations for the non-parametric curve to segment the ventricle with a mean square error of 2.02 pixels. It also required 14 iterations for the curve to segment the ventricle with a gray strip placed across but with a mean square error of 5.21 pixels. For the parametric approach, it required 28 iterations to segment the ventricle with a mean square error of 5.21 pixels and 55 iterations to segment the ventricle with a gray strip placed across with a mean square error of 5.58 pixels. The advantages of the non-parametric approach include being able to match object boundary accurately and being computationally efficient while the parametric approach is extremely robust to noise and foreign objects. Both sets of advantages may be integrated into the joint curve evolution approach to achieve a model which is robust to noise and able to match object boundaries very accurately.