Content-Based Retrieval of Focal Liver Lesions Using Geometrical and Textural Features of Multi-Phase CT-Scan Images

The most critical step in content-based medical image retrieval systems is feature extraction. The objective of the feature-extraction phase is to discriminate between different types of lesions. Preparing useful features improves diagnosis by an image retrieval system and successful treatment. We propose a new set of geometrical features based on the description of different types of tumors and integrate them with textural elements to enhance the discrimination of five kinds of focal liver lesions. The abnormal region is divided into three sections, including central, middle, and border regions. The textural features are obtained in each part individually, while the geometrical characteristics are calculated for the whole zone of a lesion. Evaluation of the results shows an improvement of prec@6 and mAP, which is attributed to the proposed geometric characteristics. Moreover, our method increases prec@6 by 3.9% compared to recent researches.