Medical Image Segmentation Based On Map-em Framework

With the development of medical imaging modalities that have been widely used in clinics, image processing techniques have been playing an increasingly important role in medical imaging. As an essential tool to extract quantitative information of specific tissues from medical images, image segmentation is prerequisite for further visualization and analysis of different tissues. However, due to the complexity of human anatomy and the multiplicity of medical imaging modalities, currently there is no segmentation algorithm that could be applicable to all kinds of medical images. Most segmentation algorithms are only suitable to the solution of specific problems. Therefore, accurate segmentation has been emerged as an important and hot topic in medical image processing.Due to the partial volume (PV) effect caused by limited resolution of imaging modalities and other degradation factors, a voxel near boundary may be considered as a mixture of different types of tissue rather than assuming that voxel being fully filled by a single tissue type. Therefore, conventional segmentation that assigns a single label to each voxel could cause a significant error in quantitative image analysis. The MAP-EM algorithm (MAP:Maximum a Posterior, EM: Expectation Maximization) utilizes the EM algorithm to estimate tissue mixture percentages in each voxel and statistical model parameters of the acquired image data simultaneously under the principle of MAP, using a tissue mixture model that accounts for the PV effect. Then final segmentation is accomplished according to the mixture percentage. Generally, the MAP-EM algorithm adopts invariable weights and penalty factors, making it difficult to balance between noise suppression and boundary preservation. To increase the accuracy of MAP-EM segmentation, in this paper, a MAP-EM algorithm has been proposed based on adaptive weights and penalty factors for mixture segmentation.Cardiovascular disease and bladder cancer are fatal diseases which severely threaten the human health. Accurate segmentation of coronary plaque and bladder wall out of surrounding tissues is very useful for the analysis of plaque components and the invasion of bladder tumor into the wall, and therefore would be beneficial to prevent, diagnose and treat the diseases in clinic. To validate the effectiveness of the improved MAP-EM algorithm with clinical data, it has been applied to segment coronary plaque out from enhanced coronary CT images and to extract the bladder wall from MRI bladder images, with more accurate initialization and the utilization of neighboring information. Comparison with the results outlined manually by the experienced radiologists has shown that the improved algorithm outperforms the original one with better noise suppression and edge preservation.