Purpose: To demonstrate that the improved signal-to-noise ratio (SNR) of objects imaged with full-field digital mammography (FFDM) versus digitized screen-film mammography (SFM) results in improvements to the feature extraction stage of a computer-aided diagnosis (CAD) scheme.
Methods and Materials: Images of a Lucite contrast-detail phantom were obtained with FFDM (Senographe 2000D, GE Medical Systems) and SFM (Senographe DMR, GE Medical Systems; Min-R 2000, Kodak) using matched imaging techniques. The contrast-detail phantom contained disks with a range of sizes and contrasts. Films were digitized with a Lumisys 85 digitizer at 100 micron pixel size. We obtained multiple samples of each disk by imaging the phantom 34 times with both FFDM and SFM. The distribution of SNRs was measured experimentally for each disk in the phantom. For every disk, we also extracted features currently used in our CAD scheme for microcalcification detection, including object area, contrast, edge gradient and texture. For each feature, the accuracy of feature extraction was computed as the mean of the measured feature values divided by the true feature value. The precision of feature extraction was measured using the relative error, computed as the standard deviation of the measured feature values divided by its mean value.
Results: The measured SNR values in the FFDM images were 15-50% higher than in the SFM images for matched imaging techniques. The accuracy and precision of feature extraction improved for objects imaged with FFDM versus SFM. For 0.44, 0.61, and 0.88 mm diameter disks, the measured accuracies for the area feature were 0.88, 0.90, and 0.93 respectively in FFDM images, and 0.83, 0.87, and 0.89 respectively in SFM images. For these same disks, the precision of the measurements was better for FFDM (0.30, 0.09, and 0.09) than for SFM (0.35, 0.12, and 0.10). Similar trends were observed for the other computer features. The accuracy and precision of feature extraction improved as the object SNR increased. We are currently increasing the number of object samples in order to test the statistical significance of these differences.
Conclusion: We demonstrated that the accuracy and precision of CAD feature extraction improved for objects imaged with FFDM versus digitized SFM due to the higher SNR. We expect that CAD schemes will perform better and more reproducibly when applied to FFDM than to SFM due to the increased accuracy and precision of feature extraction. (R.M.N. is a shareholder in R2 Technology, Inc.)
Questions about this event email: email@example.com