CBIR of spine X-ray images on inter-vertebral disc space and shape profiles using feature ranking and voting consensus

Abstract

Very limited research is published in the literature that applies content-based image retrieval (CBIR) techniques to retrieval of digitized spine X-ray images that combines inter-vertebral disc space and vertebral shape profiles. This paper describes a novel technique for retrieving vertebra pairs that exhibit a specified disc space narrowing (DSN) and inter-vertebral disc shape. DSN is characterized using spatial and geometrical features between two adjacent vertebrae. In order to obtain the best retrieval result, all selected features are ranked and assigned a weight to indicate their importance in the computation of the final similarity measure. Using a two phase algorithm, initial retrieval results are clustered and used to construct a voting committee to retrieve vertebra pairs with the highest DSN similarity. The overall retrieval accuracy is validated by a radiologist and proves that selected features combined with voting consensus are effective for DSN-based spine X-ray image retrieval.

Introduction

Osteoarthritis affects a significant portion of the elderly population in the United States [1]. Osteophytes, disc space narrowing (DSN), subluxation and spondylolisthesis are typical radiographic hallmarks characterizing this condition on the spine. The ability to retrieve spine X-ray images on these conditions could be very valuable to clinicians (radiologists), researchers of arthritis and musculoskeletal diseases, and educators. This paper focuses on the problem of retrieval of digitized X-ray images of the spine based on disk space narrowing coupled with vertebral shape content analysis.

Manually finding reference images from a large image database is a tedious and error prone process. An automatic CBIR system can significantly alleviate the problem of retrieving relevant images with specified DSN. Content-Based Image Retrieval (CBIR) techniques have been studied for nearly two decades. The techniques have been used for searching images in digital libraries, on the World Wide Web, and other applications such as trademark search [2]. Research on medical image retrieval, however, has been fairly recent [3], [4], [5], [6], [7], [8]. These efforts can be broadly categorized into two themes: (i) retrieval of biomedical images from a heterogeneous collection (images of different anatomy, modality, and detail) with little importance given to localized pathology, and (ii) retrieval of images from a homogenous collection (images of single modality, anatomy, and detail) with particular focus on the localized pathology. Our research [9] has been of the latter category.

Retrieval of medical images started with text-based retrieval and has grown to include image content-based retrieval with explosive growth in the acquisition and use of biomedical images. Mao and Chu [10] studied the vector space model (VSM) to automatically retrieve medical documents. Following the development of image processing and computer vision techniques, indexing and retrieval of the medical images based on content analysis became possible. Muller et al. [11] gave an overview of available literature in the field of content-based access to medical image data and on the technologies used in this field.

The Lister Hill National Center for Biomedical Communications, an intramural R&D division of the National Library of Medicine (NLM) at the National Institutes of Health (NIH), maintains an archive of digitized spine X-rays collected from the second National Health and Nutrition Examination Surveys (NHANES II) [12] which can serve as a reference collection for study of DSN. A prior study [13] proposed use of four scale-invariant, distance transform-based features to characterize spacing between adjacent vertebrae. K-means clustering and self-organizing map (SOM) were used to classify inter-vertebral disc space and assigned it a degree of DSN severity with an overall accuracy of 82.1%. A shortcoming of using this approach which has proven to be robust for automatically classification of severity level for shape-based CBIR is the lack of disc shape profiles. Using DSN severity level classification alone is insufficient for shape-based CBIR which is the focus of this work.

Vertebral shape is valuable in expressing the spine conditions described earlier. Fig. 1a shows two adjacent vertebrae outlined on a lumbar spine X-ray. As seen in the sagittal view, the inferior and superior edges of vertebrae adjacent to the disc can serve as the disc shape profile. Experienced radiologists use several criteria when evaluating DSN similarity between a candidate case and references from an atlas, for example. These criteria include the top to bottom size of the inter-vertebral gap, the length of the gap, and its configuration, i.e., whether there are spurs, concavities, convexities, irregularities, etc. Many of these disc space characteristics can be computed from disc shape profiles. In this paper, we present an approach that combines disc shape profile with computed inter-vertebral disc space features and uses voting consensus for finding similar images. In addressing this important problem, this effort makes advances the state of the art in CBIR taking advantage of clustering ensemble based machine learning methods [1], [14], [15], [16], [17].

The proposed algorithm and DSN similarity measures are discussed in Section 2. Feature ranking for weight computation of extracted features is presented in Section 3. Section 4 introduces the proposed voting consensus mechanism. Experimental results and analysis are presented in Section 5. Conclusions and future directions are described in Section 6.