The development of a novel parametric imaging technique for quantitative ultrasound breast tissue characterization

This dissertation examines the problem of breast tissue characterization using ultrasound radio frequency (RF) data. The main goal of this study is to develop a novel parametric imaging technique that is able to identify breast cancers at an early stage, and thereby serve as a complement to conventional mammography and sonography to reduce high benign biopsy rate.; Ultrasound is an attractive diagnostic medical imaging modality. It is used routinely to reduce the mammography false-positive rate since it can distinguish solid tumors from benign cysts. One of the most important contributions of this dissertation is the evaluation of the differential diagnostic value of parameters extracted from ultrasound radio-frequency (RF) data of breast masses (mostly solid lesions) with various benign and malignant pathologies, to improve specificity of breast cancer diagnosis. Algorithms for the extraction of ultrasonic tissue characterization (UTC) parameters, closely related to breast pathologies, were introduced. The resulting parameter set was processed further to obtain the minimum number of parameters with the most powerful discriminatory ability to differentiate malignant lesions from benign and normal tissues. Specifically, fifty-nine patients with a total of 295 in vivo breast scans were studied. The most decisive parameter subset was identified, which includes backscattering intercept (binter), weighted average of magnitudes of cepstral peaks (mcep) and effective cross-section (alpha). Several classification algorithms based on statistical and neural network methods were presented and tested. The best performance was achieved using the back-propagation neural network (BP-NN) classifier with the recognition rate of 87%, sensitivity of 100% and specificity of 81%. Although individually, each of these three parameters did not provide a needed diagnostic accuracy, a composite parameter, formed by BP-NN, obtained remarkable result of A_z = 0.95, with A_z indicating the area under the receiver operating characteristic (ROC) curve.; The optimal feature set was nonlinearly combined by a functional link network to form a single feature parameter as the cancer indicator, called parametric imaging index (PII) for parametric images. These resulting PIIs were gray-level mapped or color-coded to form final parametric images with diagnostic ability to differentiate benign (including normal) from malignant breast lesions. (Abstract shortened by UMI.)...