A similarity learning approach to content-based image retrieval: application to digital mammography

TLDR: A new approach to content-based retrieval of medical images from a database is described, in which similarity is learned from training examples provided by human observers, and the use of neural networks and support vector machines to predict the user's notion of similarity is explored.

Abstract: In this paper, we describe an approach to content-based retrieval of medical images from a database, and provide a preliminary demonstration of our approach as applied to retrieval of digital mammograms. Content-based image retrieval (CBIR) refers to the retrieval of images from a database using information derived from the images themselves, rather than solely from accompanying text indices. In the medical-imaging context, the ultimate aim of CBIR is to provide radiologists with a diagnostic aid in the form of a display of relevant past cases, along with proven pathology and other suitable information. CBIR may also be useful as a training tool for medical students and residents. The goal of information retrieval is to recall from a database information that is relevant to the user's query. The most challenging aspect of CBIR is the definition of relevance (similarity), which is used to guide the retrieval machine. In this paper, we pursue a new approach, in which similarity is learned from training examples provided by human observers. Specifically, we explore the use of neural networks and support vector machines to predict the user's notion of similarity. Within this framework we propose using a hierarchal learning approach, which consists of a cascade of a binary classifier and a regression module to optimize retrieval effectiveness and efficiency. We also explore how to incorporate online human interaction to achieve relevance feedback in this learning framework. Our experiments are based on a database consisting of 76 mammograms, all of which contain clustered microcalcifications (MCs). Our goal is to retrieve mammogram images containing similar MC clusters to that in a query. The performance of the retrieval system is evaluated using precision-recall curves computed using a cross-validation procedure. Our experimental results demonstrate that: 1) the learning framework can accurately predict the perceptual similarity reported by human observers, thereby serving as a basis for CBIR; 2) the learning-based framework can significantly outperform a simple distance-based similarity metric; 3) the use of the hierarchical two-stage network can improve retrieval performance; and 4) relevance feedback can be effectively incorporated into this learning framework to achieve improvement in retrieval precision based on online interaction with users; and 5) the retrieved images by the network can have predicting value for the disease condition of the query.