Content-based Medical Image Classification and Retrieval

Abstract

Content-based medical image classification and retrieval are the tasks finding medical images, e.g., of the same category of healthy or abnormal organs, for disease diagnosis and treatment in an automated way. They play important roles in establishing quantitative relationships among images by analyzing their content features. However, the features are often hindered by the inter- and intra-category visual variations between anatomical structures. Thus, it is important to design descriptive and discriminative feature descriptors. In this thesis, we develop novel fundamental content-based image feature analysis algorithms in medical image classification and retrieval in terms of different feature granularities: low-level feature, contextual descriptor, bag-of-visual-words (BoVW) representation and semantic association. In the interest of this study, we particularly focus on the following two applications: a lung nodule differentiation from low dose computed tomography images, and a neurodegenerative progression staging of Alzheimer's Disease (AD) given magnetic resonance and positron emission tomography scans. The thesis is divided into two parts: classification and retrieval. For the first part of this thesis, we focus on a classification task of lung nodule differentiation. It includes two lung nodule structure focus studies, which explore classification improvement by using new image information and identifying inter-category overlapping nodule images, and two contextual structure features, which investigate the combination of nodules and their surrounding anatomical structures. For the second part of this thesis, we work on two retrieval applications of finding intra-type lung nodule and intra-stage AD progression, with two feature designs of a pruned dictionary-based BoVW representation and a pair-wise latent association feature. The experimental evaluations demonstrated the better performance of the proposed methods.