Knowledge-based method for segmentation and analysis of lung boundaries in chest X-ray images Matthew S. Brown a, 1, * , Laurence S. Wilson b , Bruce D. Doust c , Robert W. Gill b , Changming Sun d a Department of Radiological Sciences, School of Medicine, University of California, Los Angeles, CA, USA b CSIRO Telecommunications and Industrial Physics, Australia c St. Vincents Hospital, Sydney, Australia d CSIRO Mathematical and Information Sciences, Australia Received 15 June 1998; received in revised form 17 September 1998; accepted 17 September 1998 Abstract We present a knowledge-based approach to segmentation and analysis of the lung boundaries in chest X-rays. Image edges are matched to an anatomical model of the lung boundary using parametric features. A modular system architecture was developed which incorporates the model, image processing routines, an inference engine and a blackboard. Edges associated with the lung boundary are automatically identified and abnormal features are reported. In preliminary testing on 14 images for a set of 18 detectable abnormalities, the system showed a sensitivity of 88% and a specificity of 95% when compared with assessment by an experienced radiologist.  1999 Elsevier Science Ltd. All rights reserved. Keywords: Knowledge-based segmentation; Chest X-ray; Blackboard; Anatomical model; Edge detection 1. Introduction Identification of the lung boundaries in chest radiographs is a necessary step for detecting abnormalities such as inter- stitial disease [1,2], pneumothorax [3], cardiomegaly [4] and pulmonary nodules [5,6]. The aim of this work is to develop an experimental system which demonstrates a knowledge-based approach to segmentation and analysis of the lung boundaries in chest X-ray images. A clear distinction is made between ‘‘high-level’’ and ‘‘low-level’’ processing. We define low-level processing as operating on raw image data. In such algorithms, the input is considered simply as an array of pixel values. We define high-level processing as operating on data which are represented in a symbolic, knowledge-based domain. If high-level analysis is to be applied to image data, then a high-level representation must be derived. In our approach, segmentation involves matching low-level image objects to high-level objects described in a mathematical model of the relevant anatomy (anatomical model). To compare low- and high-level objects, a common, intermediate representation is required, for example using parametric features. Numerous systems were reported that use of anatomical knowledge, in the form of constraints on features such as expected size, shape, texture and relative positions of struc- tures, to perform image interpretation [7–20]. A number of systems have dealt specifically with segmenting the lung fields in chest X-rays [5,6,21–27]. Typically segmentation is based around thresholding, edge detection and feature- based pixel classification. Duryea and Boone [21] used a heuristic edge-tracing approach with validation against hand-drawn lung contours. Armato et al. [27] used a combi- nation of gray-level thresholding (both global and local) and contour smoothing. McNitt-Gray et al. [23] developed a method using feature-based classification of pixels into regions such as heart, lung and axilla. These systems are effective and useful but do not provide a high-level repre- sentation of the image content. Also, domain knowledge is embedded as heuristics within the segmentation algorithms, making it difficult to reapply or extend to other problems. Images of abnormal anatomy pose a problem for all auto- mated analysis schemes, particularly cases where expected anatomical structures are missing, or altered to the point where they cannot be detected by the available segmenta- tion routines. Most knowledge-based approaches do not Computerized Medical Imaging and Graphics 22 (1998) 463–477 PERGAMON Computerized Medical Imaging and Graphics 0895-6111/99/$ - see front matter  1999 Elsevier Science Ltd. All rights reserved. PII: S0895-6111(98)00051-2 1 Formerly at CSIRO Telecommunications and Industrial Physics, Australia and School of Computer Science and Engineering, The University of New South Wales, Sydney, Australia. * Corresponding author. Tel.: + 1-310-267-1820; Fax: + 1-310-206- 2967; e-mail: mbrown@endeavour.radsci.ucla.edu