BREAST TISSUE CLASSIFICATION USING STATISTICAL PATTERN RECOGNITION ON BACKSCATTERED ULTRASOUND.

Abstract

Diagnoses using images made with non-ionizing ultrasound are based on qualitive criteria and are not more accurate than those made with mammography. Information about tissue state is lost in the processing required to produce ultrasound images, and textural information may not be perceptible to a human observer. This study uses statistical pattern recognition to classify ultrasound A-scans, before any processing other than amplification occurs. A U. I. Octoson was used to collect data from normal, benign, and malignant, in vivo breast tissues. Features based on textural or frequency content of received sound were computed from digitized A-scans. Most textural features have been used previously in image processing, while frequency features assumed differences in frequency-dependent attenuation. Data were collected at the University of Arizona from 17 malignant masses, 8 benign masses, and 7 normal tissues. Univariate and multivariate statistical tests were used to find combinations of features which discriminated best between the classes of tissue. Equal a priori probabilities were used in a Bayesian classifier to classify malignant vs. nonmalignant. Specificity of 76% (13 of 17 malignant masses correct) was found with a sensitivity of 80% (12 of 15 masses correct). A linear combination of one frequency feature and three textural features was used. For malignant vs. benign, sensitivity of 88% (15 of 17 masses) and specificity of 75% (6 of 8 masses) were found. Features used were the same as for classification of malignant vs. nonmalignant, except for modification of one textural feature. The inability to visually detect and gather data from some palpable masses means that further study is needed to determine the effectiveness of applying the method to all breast masses. A set of A-scans from Thomas Jefferson Hospital in Philadelphia was gathered using similar procedures, and analysed with the following results: 18 of 21 (86%) malignant masses, and 45 of 66 (68%) nonmalignant masses were classified correctly, using a linear combination of one textural feature and five frequency features. Confidence limits on the results show that the majority of masses can be classified correctly with this procedure, but success rates are not high enough for breast cancer screening.