基于局部稀疏形状表示的医学图像分割模型

Medical Image Segmentation Model Based on Local Sparse Shape Representation

针对人体器官计算机断层扫描(CT)图像边缘模糊、难以自动分割的问题,提出了一种基于局部先验形状信息和主动轮廓模型的分割方法。针对一个形状与训练集中样本相似的器官目标,在基于图像灰度信息进行底层分割的同时,利用形状字典中的先验形状表示目标,将其作为高层监督,引导变分目标分割。在已有形状字典稀疏表示的基础上,利用掩模矩阵对字典形状进行局部分解,以生成补充字典,通过对局部先验的稀疏形状的约束实现对目标形状的局部描述。通过对字典中相似形状局部分解的重组,替代传统整体稀疏形状的表示方法,实现对与形状字典中仅存在部分相似目标的分割,扩大了字典形状的适用范围。分割实验表明,所提模型可准确地从边缘模糊的图像中提取并分割所需目标,从而可应用于医学图像分割。

With respect to the problems of the fuzzy edge and difficulty in automatic segmentation of human organs during the computed tomography （CT） scanning, a local prior shape information and active contour based model is proposed. For an object whose shape is similar with the shapes in the dictionary, the prior shape in the shape dictionary is used to supervise and guide the high-level object segmentation while underlying segmentation is performed based on the image information. On the basis of the existed shape dictionary sparse representation, dictionary shapes are decomposed by mask matrix, and supplemental dictionary is generated, so the local shapes of the object can be described by the constraint of sparse shape of partial prior shapes. By the decomposition and recombination of the local shapes instead of the traditional prior shapes, shapes which are not included directly in the dictionary can be segmented, and the application range is extended. Experimental results of the segmentation experiments show that even if the edge of the object is fuzzy, the image can be recovered and segmented accurately with the proposed method, and the proposed method can be applied to medical image segmentation.