Query dependent multiview features fusion for effective medical image retrieval

Abstract

Multiple features have been employed for content-based medical image retrieval. To reduce curse of dimensionality, subspace learning techniques have been applied to learn a low-dimensional subspace from multiple features. Most of the existing methods have two drawbacks: first, they ignore the fact that multiple features have complementary properties, and thus have different contributions to construct the final subspace; second, they construct the optimal subspace without considering user's query preference, i.e., for a same query example, different users want different query results. In this paper, we propose a new method termed Query Dependent Multiview Features Fusion (QDMFF) for content-based medical image retrieval. Inspired by ideas of multiview subspace learning and relevance feedback, QDMFF iteratively learns an optimal subspace by fusing multiple features obtained from user feedback examples. The method operates in the following four stages: first, in local patch construction, local patch is constructed for each feedback example in different feature space; second, in patches combination, all patches within different feature spaces are assigned different weights and unified as a whole one; third, in linear approximation, the projection between original high dimensional feature spaces and the final low-dimensional subspace is approximated by a linear projection; finally, in alternating optimization, the alternating optimization trick is utilized to solve the optimal subspace. Experimental results on IRMA medical image data set demonstrate the effectiveness of QDMFF.