Improving Retrieval of Pathological Liver Images in Multi-phase CT Data Using a Correlation Graph Distance

摘要

Accurate retrieval of similar tumor images helps a clinician to suggest a successful treatment planning. In this paper, we employ a correlation graph distance to measure the similarity of the images. After reading a multi-phase input data, we extract the ROI of the corresponding tumors. We represent each phase by individual features of the GLCM texture descriptor and therefore describe the three phases by a 9×1 vector. Instead of a local correlation, we employ a global correlation to initially measure the similarity of two multi-phase data. The correlation graph distance is next constructed using the strongly connected component algorithm through an iterative process. In each iteration, the links between a node and the graph are modified. Then, the connected components are recognized, and unattached nodes are connected to the nearest clusters. Our results revealed that the non-linear approach of the correlation graph improves the Recall index by at least 10% and individual selection of the features enhances the output by at least 4 %.