| Recently, a great deal of research has been undertaken to develop effective image processing algorithms for efficient browsing, searching, and retrieval of images from a large database. In this thesis, we explore the use of a content-based approach for retrieval of mammograms. We conjecture that by presenting images with known pathology that are “medically similar” to the image being evaluated, the application of such a retrieval system would provide the radiologists with a “second reader” opinion. A fundamental issue in the development of such a retrieval system is how to determine the similarity of a given pair of mammograms in a way that is consistent with their perceptual similarity. In this work, we first develop effective methods for extraction of medically relevant features from microcalcification clusters (MCC) in a mammogram. For the automatic detection of MCCs, the classical template matching method and several of its variations are first considered; we also propose the use of a new scheme for support vector machines (SVM) training. The SVM method is found to be very effective for this task and outperforms other methods. Next, for the determination of the similarity between mammogram pairs, we propose the use of several learning machine algorithms, namely, neural networks and SVMs. In our initial study, it is found that the proposed framework is quite promising when evaluated using two statistical evaluation methods. Extension to large-scale databases is made possible through using a hierarchical structure of a pre-classifier stage and a regression stage. Further improvement in retrieval effectiveness is obtained by incorporating relevance feedback. This thesis also presents a new method based on incremental learning of SVM for regression, which has been applied to relevance feedback. |
