Minimizing the semantic gap in biomedical content-based image retrieval

A major challenge in biomedical Content-Based Image Retrieval (CBIR) is to achieve meaningful mappings that minimize the semantic gap between the high-level biomedical semantic concepts and the low-level visual features in images. This paper presents a comprehensive learning-based scheme toward meeting this challenge and improving retrieval quality. The article presents two algorithms: a learning-based feature selection and fusion algorithm and the Ranking Support Vector Machine (Ranking SVM) algorithm. The feature selection algorithm aims to select 'good' features and fuse them using different similarity measurements to provide a better representation of the high-level concepts with the low-level image features. Ranking SVM is applied to learn the retrieval rank function and associate the selected low-level features with query concepts, given the ground-truth ranking of the training samples. The proposed scheme addresses four major issues in CBIR to improve the retrieval accuracy: image feature extraction, selection and fusion, similarity measurements, the association of the low-level features with high-level concepts, and the generation of the rank function to support high-level semantic image retrieval. It models the relationship between semantic concepts and image features, and enables retrieval at the semantic level. We apply it to the problem of vertebra shape retrieval from a digitized spine x-ray image set collected by the second National Health and Nutrition Examination Survey (NHANES II). The experimental results show an improvement of up to 41.92% in the mean average precision (MAP) over conventional image similarity computation methods.