Multi-feature analysis for automated breast lesion classification from ultrasonic data

We have developed quantitative descriptors of lesions for reliable, operator-independent breast cancer identification using ultrasound. These include acoustic features as well as morphometric features related to lesion shape. Acoustic features include "echogenicity," "heterogeneity," and "shadowing," computed from radio-frequency (RF) spectral-parameter images of the lesion and surrounding tissue. Morphometric features were computed by geometric and fractal analysis of manually-traced lesion boundaries. Initial results show that no single parameter can precisely identify cancerous breast lesions and that the use of multiple features can substantially improve discrimination. Our analysis produced an ROC-curve area of 0.9164 ± 0.0346