Chest X-ray retrieval based on modified Homogeneous Texture Descriptor and local feature fusion

An effective image retrieval system applied to chest X-ray image database is presented in this paper. In this system, image features are represented and matched by a modified MPEG-7 Homogeneous Texture Descriptor, whose Gabor filter bank is redesigned to be more suitable for chest X-ray images, and the similarity metric strategy is replaced by a boosting distance metric learning method. A local feature fusion technique, which contains an automatic subdivision method and a self-adaptive local feature weighing algorithm, is proposed to detect and analyze lesions in small areas. The system is tested on a database with 120 chest radiographs and the average retrieval accuracy of this system reaches 83.67%. The experimental results indicate that the modified Homogeneous Texture Descriptor improves the retrieval effectiveness compared with the standard version, and the local feature fusion enhances the retrieval accuracy significantly.