Content-based medical image retrieval of CT images of liver lesions using manifold learning

Accurate retrieval of liver CT images can help a specialist to decide on the type of lesion and treatment planning. However, the complex texture of the abnormality and its nonlinear characteristic reduces the recognition rate of a retrieval system. In this paper, we propose how to represent an abnormal region of a liver by individual attributes of a multi-phase CT image. The indexing of a medical image database is represented by a correlation graph distance, which considers nonlinear behavior of the feature space as well. The results showed that the average recall was improved by 7.5% using the proposed feature vector. Concerning a complex scheme for lesion representation and the manifold indexing technique, the recall of the system was increased by twice. The proposed indexing and feature representation prove the potential of our method in content-based medical image retrieval systems.