We will develop a computer-assisted system to improve the ability of breast sonography to distinguish malignant from benign lesions. These are lesions originally detected by mammography and sent to ultrasound for father evaluation. The lesions are biopsied when they are classified by ultrasound as solid or indeterminate, even though 75% to 80% are benign at biopsy. Because of the large numbers of breast biopsies being performed, even a small improvement in the ability of ultrasound to distinguish benign from malignant could result in significant decreases in numbers of benign biopsies, health care costs, and patient morbidity. Our approach is to apply advanced image analysis techniques for ultrasound breast lesion images. In this Phase I project, we will analyze each lesion for three zones: background texture, edge of the lesion, and lesion itself rather than convention methods using one zone. The artificial neural network will then integrate extracted features for the classification of the malignancy of the lesion. A newly developed artificial visual neural network will be constructed specifically for ultrasound images. This neural network, which simulates human eye, will be trained by experienced sonographers and will be tested for the analysis of ultrasound breast lesions.In Phase II, we will finalize all possible features and analytical techniques, test our methods on a large database, and develop automatic process on a workstation to make the system usable in a clinical environment.National Cancer Institute (NCI)
