The Medical Image Segmentation Based On Boundary Feature Learning

With the development of the medical imaging techniques, medical imagesegmentation is a very important task in computer-aided diagnosis and therapy. Inorder to alleviate the burdensome work and avoid inter-observer variations for thedoctors or experts, using the algorithms in computer vision, machine learning, andimage processing can segment the organ from medical images automatically.Although the statistical shape model (SSM) has been widely used in medical imagesegmentation, the accuracy of prostate segmentation in MR images should beimproved.Due to the incorporation of shape prior in shape modeling, SSM based methodshave been successfully used in medical image segmentation. However, there are somedisadvantages in SSM based methods, which affect the accuracy of the segmentationresults. SSM based methods consist of several modules, such as shape representation,shape model, shape correspondence, appearance model, and search algorithm. Inbuilding the appearance model, the existing features cannot describe the boundary ofthe organ exactly and mislead the deformation of the shape. The principal componentanalysis (PCA) based shape modeling can capture the global shape variation, butignore the local shape variation. As a result, the shape is over-constrained and cannotfit the new image.In order to solve the mentioned problems, a new segmentation method forprostate in MR images is proposed. The segmentation method consists of three steps.Firstly, the distinctive feature is proposed to describe the boundary in building theappearance model. The scale and variance adaptive SIFT (SVA-SIFT) feature isemployed to capture the information of the local patch surrounding the boundary. Asthe gray values and gradients vary significantly over the boundary of the prostate,separate appearance descriptors are built for each landmark and optimized. Based on SVA-SIFT, two types of truncated normal vector feature profile (NVFPt) are proposedand optimized using the distinctive function for features. Secondly, the global shapevariation and local shape variation is combined in a coarse-to-fine segmentation framework. Thirdly, in extracting the SVA-SIFT feature and optimized NVFPt, thediscriminant analysis is introduced to measure the distinctiveness of features. Thedistinctive function is achieved with minimizing the intra-class distance andmaximizing the inter-class distance.The proposed segmentation method is employed in prostate segmentation in MRimages. The experimental segmentation results validate the performance of thedistinctive function and the distinctiveness of SVA-SIFT and NVFPt, in addition tothe effect of the local variation on shape modeling. The proposed method is alsocompared with other segmentation methods. The experimental results improve thatthe proposed method performs better than other segmentation methods.