EMR: A Scalable Graph-based Ranking Model for Content-based Medical Image Retrieval

EMR: A Scalable Graph-based Ranking Model for Content-based Medical Image Retrieval

Abstract

Graph-based ranking models have been widely applied in information retrieval area. In this paper, we focus on a well known graph-based model - the Ranking on Data Manifold model, or Manifold Ranking (MR). Particularly, it has been successfully applied to content-based image retrieval, because of its outstanding ability to discover underlying geometrical structure of the given image database. However, manifold ranking is computationally very expensive, which significantly limits its applicability to large databases especially for the cases that the queries are out of the database (new samples). We designed a novel scalable graph-based ranking model called Efficient Manifold Ranking (EMR), trying to address the shortcomings of MR from two main perspectives: scalable graph construction and efficient ranking computation. Specifically, we build an anchor graph on the database instead of a traditional k-nearest neighbor graph, and design a new form of adjacency matrix utilized to speed up the ranking. An approximate method is adopted for efficient out-of-sample retrieval. Experimental results on some large scale image databases demonstrate that EMR is a promising method for real world retrieval applications.

EMR for Content-Based Image Retrieval

Fig. 1. Extend matrix W (MR) and Z (EMR) in the gray regions for an out-of-sample.

In this part, we make a brief summary of EMR applied to pure content-based image retrieval. To add more information, we just extend the data features. First of all, we extract the low-level features of images in the database, and use them as coordinates of data points in the graph. Secondly, we select representative points as anchors and construct the weight matrix Z with a small neighborhood size s. Anchors are selected off-line and does not affect the on-line process. For a stable data set, we don’t frequently update the anchors. At last, after the user specifying or uploading an image as a query, we get or extract its low-level features, update the weight matrix Z, and directly compute the ranking scores by equation 

. Images with highest ranking scores are considered as the most relevant and return to the user.

 

Simulation Video Demo