Study On Methods Of Co-segmentation In PET Image Segmentation

The segmentation of PET(Positron Emission Tomography) is an extremely important link in clinical medicine. It can help us locate the tumor regions accurately and is of great significance for the treatment to patients. Now there are a lot of methods for PET segmentation and they also have acceptable results. But these traditional segmentation methods are all for single image segmentation. They do not have any contact among these images. Sometimes they need some prior information and do not have satisfactory robustness. So we hope to find a method which can share some common information among images to help us segment multiple images simultaneously. Of course, we need the results have better performance index and robustness than traditional methods.According to the idea of co-segmentation in natural image, we got some inspiration. Refer to some methods, we applied the co-segmentation idea into PET segmentation. We attemptted two co-segmentation methods for PET. One was combining nomalized cuts and kernel method. The nomalized cuts characterized the separability of single image and the kernel method shared the common information among the images. Trade-off was achieved using an energy function, which can be efficiently optimized using the low-rank optimizing technique. The other was based on the active contour model. On the basis of the original model, some reward strategy was added to limit the similarity between the images. This question can be iteratively optimized by level sets and variation methods.This paper conducted some co-segmentation experiments with MATLAB R2012 a on some PET datasets. The performance indexs of the results were compared with the traditional single image segmentation. By comparison, we found that the co-segmentation of PET can enhance the veracity and robustness of the segmentation results. This means cosegmentation methods which make use of the similar information among PET images can help us improve segmentation performance of PET images.