Machine Vision and Applications (2010) 21:865–877 DOI 10.1007/s00138-009-0201-3 ORIGINAL PAPER A perceptual similarity method by pairwise comparison in a medical image case M. Luisa Durán · Pablo G. Rodríguez · J. Pablo Arias-Nicolás · Jacinto Martín · Carlos Disdier Received: 25 November 2007 / Revised: 24 February 2009 / Accepted: 15 May 2009 / Published online: 6 June 2009 © Springer-Verlag 2009 Abstract The evolution of image techniques in medicine has improved decision making based on physicians’ expe- rience by means of computer-aided diagnosis (CAD). This paper focuses on the development of content-based image retrieval (CBIR) and CAD techniques applied to bronchos- copies and according to different pathologies. A novel pair- wise comparison method based on binary logistic regression is developed to determine those images must alike to a new image from incomplete property information, after account- ing for the physicians’ appreciation of the image similarity. This method is particularly useful when problems with both a large number of features and few images are involved. Keywords Computer vision · CBIR · Logistic regression · Computer-aided diagnosis · Similarity 1 Introduction Research into image retrieval has steadily gained high rec- ognition over the past few years as a result of the great M. Luisa Durán · Pablo G. Rodríguez (B ) · J. Pablo Arias-Nicolás Escuela Politécnica, 10071 Cáceres, Spain e-mail: pablogr@unex.es M. Luisa Durán e-mail: mlduran@unex.es J. Pablo Arias-Nicolás e-mail: jparias@unex.es J. Martín Facultad de Ciencias, 06071 Badajoz, Spain e-mail: jrmartin@unex.es C. Disdier Hospital San Pedro de Alcántara, 10003 Cáceres, Spain e-mail: cdisdier@separ.es increase in the volume of digital image production. How- ever, little use can be made of the information unless it is organized so as to allow efficient browsing, searching, and retrieval. Two fundamental problems remain largely unsolv- ed: how to best learn from users’ query concepts, and how to measure perceptual similarity. The research communities in Database Management and Computer Vision analyze image retrieval from different angles, one being text-based and ano- ther visual-based. A very popular framework in the 1970s was to first annotate the images by text, and then to use text- based database management systems (DBMS) to perform image retrieval [37]. The essential difficulty in this method results from rich content in the images and subjectivity in hu- man perception: the same image content may be perceived differently by other users. Perception subjectivity and anno- tation inaccuracy may cause unrecoverable mismatches in the subsequent retrieval processes [35]. Then, the computer vision community enables feature extraction methods com- putationally, without accounting for human perception. Dif- ficulty results from differences between computational and human similarities [45], this discrepancy is known as the semantic gap. Computational features do not have as high a semantic content as human perceptual features do [26]. Frequently in content-based image retrieval (CBIR), system query results are a set of images selected by feature similar- ities related to the query image [7, 15, 46]. Then, a decision- making conflict arises about which images are most similar to the query. Image retrieval by pictorial content is a new type of strat- egy in database systems. A global perspective of CBIR is presented in [40]. Because of different application domains, e.g. spatial photogrammetric [16], biological modelling [32] and Medicine, specific systems are suitable for adaptabil- ity design. In the case of Medicine, picture archiving and communication systems (PACS) are general-use computer 123