Research Of Multi Level Threshold And Multi-Atlas Methods On Medical Image Segmentation

Medical imaging techniques have been rapidly developed and are serving as an important accessibility tool for medical staff when doing disease diagnosis,risk evaluation and rehabilitation program making.At the same time,huge quantity of medical image data calls for more powerful computer-aided image processing techniques.Image segmentation is a vital image processing technique and its role is to partition every pixel in the image to different classes so as to provide medical staff with valuable anatomical information,such as the accurate position,size and shape of the lesion.Moreover,the output of image segmentation can be applied in other image processing techniques.Therefore,to some degree,the quality of image segmentation determines the quality of 3D restoration,visualization,quantity analysis and other processes and that makes the research on medical image segmentation great practical significance.During the last 30 years,a great number of image segmentation algorithms have been proposed.For example,intensity based methods and deformable model based methods are employed to segment lung,heart,breast,kidney and other organs or tissues automatically.These techniques make the work of medical staff much simpler and improve the chance of patients' rehabilitation.However,for some special medical images,there is still no available method to operate segmentation.For example,MR images of brain always suffer from a un-uniformity of intensity(also known as bias field)because of the electric noise of MR scanner.Also,these images are contaminated by volume effect which is caused by the limited resolution of MR equipment.These bias field and effect make the intensity of different tissues look so similar and the edges so overlapping that almost no segmentation method based solely on intensity or deformable model can offer a decent segmentation result;another example is the CT image of the aorta where the intensity of aorta is very similar to neighboring tissues,such as the spine.The thin wall of the aorta makes it even worse as only quite an extremely blurring edge is presented in the image.Hence,almost no useful information of intensity,edge or texture can be used when doing segmentation.With the abovementioned reasons,we use the fuzzy entropy theory and the concept of membership function to handle the challenges of MR images of brain.With the help of membership function,each pixel(or voxel)of the image can be softly partitioned into each class and avoid the risk of wrong labeling.When optimizing the objective function of fuzzy entropy,we proposed an improved backtracking search optimization algorithm(BSA)whose local search ability is much better than the original BSA to retrieve the best thresholds in a more efficient and accurate way.The experiments demonstrate that fuzzy entropy-integrated method performs well than the conventional Otsu's and Kapur's methods.To tackle the low discriminability of aorta CT images,we employ the multi-atlas segmentation method and combine the joint label fusion strategy for the first time to extract the aorta structure.Joint label fusion strategy takes the similarity between atlas and the target image as well as the correlation between atlases into consideration which overcomes the interference of redundant information and improves the label propagation process.Experimental results show that the proposed method can automatically segment the CT images with a precision of more than 90%and its performance and robustness are better than other label fusion techniques.