医学图像自动分割的混合聚类方法

摘要

Segmentation result and automation level of medical images impact significantly on those aspects including computer-aided diagnosis and visualisation. In light to the features of medical images with low-contrast and high noisy influences, a hybrid clustering method is proposed. After the pre-processing step, the neighbouring eigenvectors of each pixel are fed into a self-organising map (SOM) for training. The output prototype vectors of SOM, as the preliminary clustering outcomes, are then filtered with hits-map, and followed by further processing with a hierarchical agglomerative clustering method. A quantitative index of image segmentation is selected to determine the best sort number of the clustering after comparing the Davies-Bouldin ( DB) clustering evaluation index with two other image segmentation evaluation indices. The final segmentation results are obtained from the post-processing. Analyses show that the proposed method is effective. Meanwhile, its limitations and the research work in the future are also discussed.%医学图像的分割效果和自动化程度对计算机辅助诊断和可视化等方面有重要影响.针对医学图像低对比度、噪声影响大的特点,提出一种混合聚类方法:在预处理图像之后,将每个像素的邻域特征向量送入自组织特征映射网络SOM(self-organizing map)中进行训练;作为初步聚类的结果,SOM的输出典型向量根据命中图(Hits-Map)过滤,再由层次合并聚类方法进一步处理.在比较了一种聚类评价指数和两种图像分割评价指数之后,采用图像分割量化指数来确定聚类的最佳类别数;再通过后处理得到最后分割结果,分析表明这个方法是有效的.同时,也指出其不足之处和进一步研究的方向.