Modeling clinician medical-knowledge in terms of med-level features for semantic content-based mammogram retrieval

Abstract

The huge volume of variability in real-world medical images such as on dimensionality, modality and shape, makes necessary efficient medical image retrieval systems for assisting physicians to perform more accurate diagnoses. However, the major limitation of these systems is the semantic gap, which is the difference between low-level features of images and their high-level semantics in a given situation. This paper deals with this problem and proposes a content-based image retrieval method based on med-level descriptors. These descriptors are automatically generated from low-level image features by exploiting the semantic concepts based on the clinician medical-knowledge. In fact, the proposed method is based on three main steps: (1) low-level feature extraction, (2) med-level model extraction and (3) online retrieval based on med-level feature vectors. The main contributions reside firstly in the integration of clinician medical-knowledge in terms of med-level features without needing radiologists interaction. Secondly, the determination of the query high-level features can be performed through the predicted query med-level descriptors, in addition to retrieve the most relevant images to the query one. Proposed method was validated in the context of mammogram retrieval, on the MIAS dataset, and the results prove its effectiveness and its superiority to the compared methods.