Background: There is considerable evidence that one of the strongest risk factors for breast carcinoma can be assessed from the mammographic appearance of the breast. However, the magnitude of the risk factor and the reliability of the prediction depend on the method of classification. Subjective classification requires specialized observer training and suffers from inter- and intraobserver variability. Furthermore, the categoric scales make it difficult to distinguish small differences in mammographic appearance. To address these limitations, automated analysis techniques that characterize mammographic density on a continuous scale have been considered, but as yet, these have been evaluated only for their ability to reproduce subjective classifications of mammographic parenchyma.
Methods: In this study, using a nested case-control design, the authors evaluated the direct association between breast carcinoma risk and quantitative image features derived from automated analysis of digitized film mammograms. Two parameters, one describing the distribution of breast tissue density as reflected by brightness of the mammogram (regional skewness) and the other characterizing texture (fractal dimension), were calculated for images from 708 subjects identified from the Canadian National Breast Screening Study.
Results: These parameters were evaluated for their ability to distinguish cases (those women who developed breast carcinoma) from controls. It was found that both the skewness and fractal parameters were significantly related to risk of developing breast carcinoma.
Conclusions: Although the relative risk estimates were moderate (typically > 2.0) and less than those from subjective classification or for an interactive computer method the authors have previously described, they are comparable to other risk factors for the disease. The observer independence and reproducibility of the automated methods may facilitate their more widespread use.