Using the forest to see the trees: Correlation pattern recognition and fractal context classification in an integrated featureless computer-aided diagnosis system for non-palpable breast cancer.

摘要

The vision of this work was to identify an area of breast cancer disease which is a major contributor to the benign outcome biopsy problem of diagnosis by mammography. Our analysis of a well-studied public domain database led to the discovery that pleomorphic microcalcifications presented the most challenging descriptive variability for the task of transcribing the clinician's ontological criteria to computational intelligence methods. These observations, together with new findings regarding the underlying medical mineralogy problem for microcalcifications, formed the basis of our approach. The novelties emanating from this perspective led us to integrate the design of computer-aided detection with computer-aided diagnosis. We feel vision and discrimination tasks are inseparable for cohesive modeling of the difficulties inherent in searching for, and distinguishing, the semantic information conveyed by the term, pleomorphic, or literally, many shapes.;Our path, therefore, led naturally to the comparison of other systems in analogous problem domains where a target under wide variation must be searched in an image and discriminated as interesting or uninteresting. Military problems, where automatic target recognition systems must recognize vehicles as friend or foe, face a similar problem when trying to recognize whether a noisy radar image contains a truck carrying scud missiles or is simply an oil tanker. Advanced correlation pattern recognition resolves many of the challenges in the military problem by extending an unheralded template matching technique to composite template matching. We borrow this technique.