Determining the Correct Number of Clusters in the CT Image Segmentation

Abstract

Clustering algorithm plays an essential role in CT image segmentation, and cluster validity index is an essential component in clustering analysis. There are a lot of validity indices used for assessing clustering results, that is, determine the optimal cluster number. But the existing validity indices are often ineffective for the datasets with irregular-shaped clusters and corrupted by noise. This study aims to define a novel validity index which cannot be affected by the shapes of clusters and corrupted by noise of the investigated datasets. Chain-based distance different from original Euclidean distance is defined first, then by a multidimensional scaling (MDS) transformation, all points are mapped into a new data space. After evaluation of compactness and separation twice in datasets, a novel validity index is proposed. A lot of synthetic datasets and several typical CT images were used for validating the proposed validity index. Experimental results validate the proposed index and this index is applicable to the datasets with arbitrary-shaped clusters and corrupted by noise, which is helpful in clustering analysis and computer-aided detection system.