Relevance feedback for shape-based pathology in spine x-ray image retrieval

摘要

Relevance feedback (RF) has become an active research area in Content-based Image Retrieval (CBIR). RF attempts to bridge the gap between low-level image features and high-level human visual perception by analyzing and employing user feedback in an effort to refine the retrieval results to better reflect individual user preference. Need for overcoming this gap is more evident in medical image retrieval due to commonly found characteristics in medical images, viz., (1) images belonging to different pathological categories exhibit subtle differences, and (2) the subjective nature of images often elicits different opinions, even among experts. The National Library of Medicine maintains a collection of digitized spine X-rays from the second National Health and Nutrition Examination Survey (NHANES II). A pathology found frequently in these images is the Anterior Osteophyte (AO), which is of interest to researchers in bone morphometry and osteoarthritis. Since this pathology is manifested as deviation in shape, we have proposed the use of partial shape matching (PSM) methods for pathology-specific spinal X-ray image retrieval. Shape matching tends to suffer from the variability in the pathology expressed by the vertebral shape. This paper describes a novel weight-updating approach to RF. The algorithm was tested and evaluated on a subset of data selected from the image collection. The ground truth was established using Macnab's classification to determine pathology type and a grading system developed by us to express the pathology severity. Experimental results show nearly 20% overall improvement on retrieving the correct pathological category, from 69% without feedback to 88.75% with feedback.