Research On Automatic Liver Segmentation Based On Statistical Shape Model

With the rapid development of computer technology,computer-aided diagnosis technique has gained rapid development in the medical field,especially in medical imaging.It has been found that computer-aided diagnosis techniques play a great positive role in improving diag-nostic accuracy,reducing missed diagnosis,and improving work efficiency.Liver segmentation from 3D CT images is indispensable for volume measurement,resec-tion planning and postoperative effect evaluation etc.The main challenges in automatic seg-mentation are in that similar intensity values with surrounding issues,large shape variability and pathology,the existence of liver lesion and artifacts in the scanning process.In recent years,model-based segmentation approaches have been established as one of the most successful methods for image analysis.The model contains information about the expected shape and ap-pearance of the structure of interest.Due to the inherent prior information,this approach is more stable against local image artifacts and perturbations.The statistical shape model gathers the information about common variations among all shapes.Therefore,statistical shape model is an adequate model for research on liver segmentation in spite of considerable natural variability.Modeling the statistics of a class of shapes requires a set of appropriate training shapes with well-defined correspondences.Based on point distribution model,the geometric features of Simplex mesh is easy to calculate,so simplex mesh is employed to represent shape.To es-tablish dense mesh point correspondences among all shapes of training set is generally the most challenging part of 3D model construction,and is the major factor influencing model quality.Manual landmarking is not only time-consuming,but also expert work required.A much larger amount of landmarks is needed and the correspondences are much harder to pinpoint on 3D shapes.In practice,all algorithms that automatically compute correspondences actually perform a registration.An approach based on volume-to-volume registration is proposed to establish mesh point correspondences in this study.An improved deformable Simplex mesh model is used to optimize the mesh,and the mesh surface can be consistent with the liver boundary.After the model is fitted to new data,the segmentation process needs to be executed under the guidance of the appearance of the structure of interest.Boundary,internal,external profile features are extracted to train the k-Nearest Neighbor(kNN)classifications,and constructing the appearance model.In the segmentation phase,model initializations far from the structure of interest can result in a complete failure.AdaBoost classifier is trained to localize liver position robustly.The seg-mentation algorithm is a coarse-to-fine process based on active shape model search.Under the hard constraint of statistical shape model,the template shape is fitted to the contour predicted by the appearance model,and the coarse segmentation terminates.For liver soft tissue segmen-tation,as the large amount of natural variation in these structures often cannot be captured ad-equately with a limited amount of modes of variation,the shape constraints are relaxed after initial segmentation.Local free-form deformation is used to adapt the mesh to the optimized liver boundary.This study is tested on the benchmark of Visceral anatomy of grand challenge.The results demonstrate automatic segmentation method based on statistical shape model is accurate and robust.For further testing the accuracy and stability,a test database from Zhujiang Hospital,Southern Medical University,containing 42 contrast-enhanced CT images is built.All-round comparison with probabilistic atlas method is finished.The results show this approach is adap-tive in various liver surgery cases,and has potential in clinical application.