COMPUTER-ASSISTED DIAGNOSIS - THE CLASSIFICATION OF MAMMOGRAPHIC BREAST PARENCHYMAL PATTERNS

We have developed a method for the quantification of breast texture by using different algorithms to classify mammograms into the four patte rns described by Wolfe (N1, P1, P2 and Dy). The computerized scheme em ploys craniocaudal views of conventional screen-film mammograms, which are digitized by a laser scanner. We used discriminant analysis to se lect among different feature-extraction techniques, including Fourier transform, local-contrast analysis, and grey-level distribution and qu antification. The method has been evaluated on 117 clinical mammograms previously classified by five radiologists as to mammographic breast parenchymal patterns (MBPPs). The results show differences in agreemen t among radiologists and computer classification, depending on the Wol fe pattern: excellent for Dy (kappa = 0.77), good for P2 (kappa = 0.52 ) and N1 (kappa = 0.52) and poor for P1 (kappa = 0.22). Our quantitati ve texture measure as calculated from digital mammograms may be valuab le to radiologists in their assessment of MBPP and therefore useful in establishing an index of risk for developing breast carcinoma.