Soft SVM and Novel Sampling Rule Based Relevance Feedback for Medical Image Retrieval

In content-based image retrieval, understanding the user's needs in the process of retrieval is a challenging task. Relevance feedback (RF) has been proven to be an effective method for integrating the user's knowledge into the retrieval process to eliminate the semantic gap between the high level semantic concept and the low level features of an image. In this paper we present a framework of content-based medical image retrieval with RF based on support vector machine (SVM). In the framework, we design two novel sampling methods, i.e., nearest positive margin sampling algorithm (NPMSA) and positive margin sampling algorithm (PMSA), which can select informative images to feedback to user; and we adopt 10-level soft label instead of 2-level hard label, which increases the annotation accuracy. The results of experiments on medical image database show that the proposed sampling methods, especially NPMSA one, both outperform SVMactive sampling method, and the soft SVM classifier based on the framework behaves better than SVMactive. The convergence speed of RF based on the proposed framework and the sampling methods is faster than that of SVMactive.