Quantitative classification of breast tumors in digitized mammograms.

The goal of this study was to develop a technique to distinguish benign and malignant breast lesions in secondarily digitized mammograms. A set of 51 mammograms (two views/patient) containing lesions of known pathology were evaluated using six different morphological descriptors: circularity, mu R/sigma R (where mu R = mean radial distance of tumor boundary, sigma R = standard deviation); compactness, P2/A (where P = perimeter length of tumor boundary and A = area of the tumor); normalized moment classifier; fractal dimension; and a tumor boundary roughness (TBR) measurement (the number of angles in the tumor boundary with more than one boundary point divided by the total number of angles in the boundary). The lesion was segmented from the surrounding background using an adaptive region growing technique. Ninety-seven percent of the lesions were segmented using this approach. An ROC analysis was performed for each parameter and the results of this analysis were compared to each other and to those obtained from a subjective review by two board-certified radiologists who specialize in mammography. The results of the analysis indicate that all six parameters are diagnostic for malignancy with areas under their ROC curves ranging from 0.759 to 0.928. We observed a trend towards increased specificity at low false-negative rates (0.01 and 0.001) with the TBR measurement. Additionally, the diagnostic accuracy of a classification model based on this parameter was similar to that of the subjective reviewers.