ABSTRACT
Tissue classification by examining sets of ultrasound parameters is an elusive goal. We report analysis of measurements of ultrasound speed, attenuation and backscatter in the range 3 to 8 MHz in breast tissues at 37 C. Statistical discriminant analysis and neural net analysis were employed. Data were acquired from 24 biopsy and 7 mastectomy specimens. Best separation of the classes normal, benign, and malignant occurred in the 18 cases where two tissue classes were present in the same specimen and parameters were corrected for within-patient mean; then 85-90% of cases in test sets were correctly classified. Most errors comprised misclassified benign cases. The neural net was comparable to discriminant analysis and slightly superior in separating normal and malignant classes.
Subject(s)
Breast/pathology , Image Processing, Computer-Assisted/methods , Ultrasonography, Mammary , Breast Neoplasms/diagnostic imaging , Breast Neoplasms/pathology , Discriminant Analysis , Female , Humans , In Vitro Techniques , Neural Networks, Computer , Sensitivity and SpecificityABSTRACT
Results of measurements of ultrasound speed and absorption coefficients in the range 3 to 8 MHz in breast tissues at 37 C are reported and analyzed in attempts to identify a set of ultrasound parameters capable of discriminating normal, benign, and malignant tissues. We analyzed 118 tissue regions, comprising 47 normal, 55 benign, and 16 malignant by straight-line fitting of frequency dependence of attenuation. Data for ten additional regions, for a total of 128, became available and were added to the cohort when we subsequently fitted quadratic curves. Sound speed consistently emerged as the variable with greatest discriminating power, particularly for separating normal from benign and malignant tissue. Great difficulty was encountered in discriminating benign from malignant, even when the jackknife technique was used. More success was found with classification and regression trees (CART), although results were sensitive to assigned misclassification costs. Best results from straight-line fits were obtained when discriminating malignant from combined normal/benign data after randomly assigning 75 percent of the data to the learning set and 25 percent to the test set. Then, 23 out of 25 normal/benign and 4 out of 4 malignant cases in the test set were correctly classified. With quadratic fitting, best results were obtained in the three-class case--the false positive rate for malignancy was reduced to zero in the learning (0/31) and test (0/10) sets. Nevertheless, the false negative rate increased to 13 out of 31 (42 percent) in the learning set, while attaining zero (0/4) in the test set.