Information-Theoretic Based CAD in Mammography

Duke University, Durham NC

Investigators

Linked publications & trials

Abstract

This competing renewal application aims to continue ongoing development of a novel, knowledge-based computer-assisted decision (CAD) system to improve the detection of malignant masses in mammography. The principle guiding our research is that knowledge-based systems are the future of CAD technology because i) they are adaptive, taking advantage of rapidly growing digital image libraries without requiring continuous retraining or recalibration, and ii) they can be interactive, providing evidence-based decision support for patient management as well as for training and continuing education. The novel aspect of our CAD system is that it relies on mutual information (MI) to assess regional similarities between a query image and images stored in the knowledge library. By studying the pixel-based information content of images, our CAD system is completely featureless, thus eliminating the painstaking and platform-specific image preprocessing steps required for feature extraction and selection. During the original funding period we established the potential of our CAD system for screen-film mammograms and demonstrated its feasibility in full-field digital mammography and digital breast tomosynthesis. The current application aims to improve upon the stand-alone performance and clinical integration of our CAD prototype guided by visual perception modeling principles and eye-tracking technology. Specifically, we propose the following objectives. First, we will enhance the standard MI similarity measure by incorporating "bottom-up" (i.e., stimulus-driven) saliency maps to capture more effectively the diagnostic and visual similarity of two images. Second, we will investigate state-of-the-art decision algorithms that exploit the relational structure of the knowledge database, thus imitating the "top-down" (i.e., goal-oriented) cognitive model of decision making. Third, we will integrate eye-tracking technology into the CAD system to provide decision support that is tailored to the individual needs and reading patterns of each radiologist, a dramatic departure from the static, "black-box" CAD paradigm that is currently used in clinical practice. We will develop and evaluate the system in digital mammography but we will also assess its adaptability to digital breast tomosynthesis. The expected outcome of this research is a perception-driven, evidence-based decision support system customized to each radiologist's visual search pattern that improves robustly the breast cancer detection accuracy of radiologists while bridging the performance gap among them. Although this proposal targets specifically the detection of malignant masses in mammography, the proposed CAD methods will be applicable to other breast imaging modalities as well.