Similarity searching for chest CT images based on object features and spatial relation maps

In this paper, an object-based image retrieval system for chest CT image databases is proposed. Based on the scheme of the content-based image retrieval method, we proposed an image segmentation method which combines the anatomical knowledge of the chest and the well-known watershed segmentation algorithm. The purpose of segmentation is to identify the mediastinum and the two lung lobes in a chest CT image. The ARGs (attributed relational graphs) are chosen to describe the features of segmented objects. Then, image database is constructed by the feature vectors of images. In database searching, two searching modes are provided that are "query by example" and "query by object". Our system uses Euclidean distance to measure the similarity between the image in query and the image in database. The system output the 30 most similar images in the chest CT image database as query results. The experimental results show that the average precision of our system is about 80% which is impressive in a totally automatic medical image retrieval system. Moreover, query concentrated in certain objects features usually show better result than the regular query by example. The possible reasons are discussed.