Methods for Characterizing Mammographic Density

Methods for characterizing mammographic density can be classified as subjective, like the SCC technique described earlier, interactive, or fully automatic. Furthermore, they can be based on anatomical or mathematical image features or on the underlying physics of image acquisition. In this chapter, we will use some of the research that has been carried out in our laboratory to exemplify these methods.

Subjective methods such as SCC can provide strong risk predictions; however, several limitations have been identified. Use of subjective criteria contributes to considerable inter- and intraobserver variation in making the classification. For example, in one study [5], where three radiologists made classifications of mammographic density, there was considerable variation in the magnitude of the risk prediction between them; agreement between radiologists was as low as 60%. Subsequently it has been found that cooperative training of radiologists can improve their agreement; however, this may not be possible or practical in general. A further limitation is associated with coarse categorization scales, which make it difficult to distinguish small differences in mammographic density. The ability to measure small differences in mammo-graphic density is particularly important in studies investigating changes in density with time, or with a potential intervention that may alter the breast. The use of more categories to provide a finer scale of density assessed in the same subjective manner would likely only increase variability in classification.

It is desirable to be able to characterize mammographic density on a continuous scale (quantitative), with particular consideration for magnitude of the risk prediction, and the reliability of classification. Because many clinical studies are based on patients for whom mammograms have already been obtained, it is also important to use image features that can be obtained from images without further information, i.e., features that can be calculated retrospectively.

0 0

Post a comment