Abstract
About 15 to 20% of malignant masses are missed in routine screening mammography. At the same time, about 80% of cases sent to biopsy are found to be normal. This is due to the complexity of normal breast tissue in the images. Mammograms have power-law spectra, P(f) = K/f^b, similar to natural scenes. The average spectral exponent, b, is approximately 3 with a range from 1.5 to 3.5. I have shown that the log-log plot (CD diagram) of contrast threshold for detection versus mass size has a slope m = (b-2)/2, with positive slopes for b greater than 2. I verified this equation by human observer experiments using filtered noise over the above exponent range. The focus of this talk is the effect of tissue composition. It is well known that masses are much harder to detect in dense tissue regions. I have a large database of normal digital mammograms and have found average exponents of 2.4 and 3,2 for fatty and dense tissue regions respectively. Human CD diagrams for the two tissue types were determined by 2AFC experiments by adding masses extracted from specimen radiographs to normal tissue backgrounds. Slopes agreed with theory. One can plot the physically determined contrast of growing masses as a function of size on the CD diagrams to determine when they will become detectable with a selected accuracy. It will be shown that under some circumstances it may not be possible to detect masses in dense tissue.
This research is supported by NIH grant R01-CA87734