Digital Mammography: Advanced Computer-Aided Breast Can*

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): The major goals of the proposed research are (1) to develop a computer-aided diagnosis (CAD) system for full field digital mammography (FFDM) using advanced computer vision techniques and (2) to evaluate the effects of CAD on interpretation of DMs. Previous CAD methods for lesion (mass and microcalcification) detection and characterization have been designed for digitized film mammograms and have generally been based on image features extracted from a single view. Our proposed approach is distinctly different from the previous approaches in that image information from two-view mammograms and bilateral mammograms will be fused using machine intelligence techniques. This fundamental change will expand the amount of information utilized in CAD and is expected to improve lesion detection and characterization. New computer vision techniques will be specifically designed for FFDM in order to exploit the advantages offered by digital detectors. This will produce a CAD system that is integrated with and takes full advantage of the latest imaging technologies to further improve the health care of women. We hypothesize that these advanced multiple-image information fusion techniques will lead to a more effective CAD system for FFDMs in comparison to a single-image approach, and that the CAD system will significantly improve radiologists' accuracy in the four most important areas of mammography: (i) detection of masses, (ii) classification of masses, (iii) detection of microcalcifications, and (iv) classification of microcalcifications. A database of digital mammograms (DMs) with malignant and benign lesions and a set of normal cases will be collected. We will first adapt our current film-based CAD algorithms to DMs in each of the four areas, taking into account the differences in the imaging characteristics between DMs and digitized mammograms. New computer vision techniques will then be developed to improve upon the current methods and to exploit the potential advantages of the high contrast sensitivity, high detective quantum efficiency, wide dynamic range, and the linear response to x-ray intensity of digital detectors. Novel regional registration methods for identifying corresponding lesions on CC and MLO views and for comparing the density symmetry on bilateral mammograms will be developed. Innovative fuzzy classification schemes will be designed to fuse multiple-image information and one-view information to reduce false positives and to improve detection sensitivity. Multiple-view morphological and texture features of a lesion will be merged using neural networks or other statistical classifiers for characterization of malignant and benign lesions. To test the hypotheses, we will (1) compare the performance of the multiple-image fusion CAD algorithm for DMs in each area to that of the corresponding one-view algorithm, (2) compare the detection accuracy of masses and microcalcifications on DMs with and without CAD by observer ROC studies, and (3) compare the classification accuracy of masses and microcalcifications on DMs with and without CAD by observer ROC studies. It is expected that this research will not only lead to an effective CAD system for FFDM, the multiple-image fusion approach and the new computer vision techniques will also advance CAD technology for mammography in general.