Schematic block diagram of both the previous RFPR and new CAD systems.
Ranklet transform, invariance to linear∕nonlinear monotonic gray-scale transformations of the original image. Ranklet decomposition of an image (a) and its histogram-equalized version (b) from resolution 2 up to 16. Regardless of the images’ gray-scale histogram, the resulting ranklet decompositions are nearly identical.
Difference between the gray-scale dynamics used for training and test both the previous RFPR and new CAD systems: average gray-scale histogram of analog digitized DDSM mammograms used for training (a) and full-field digital in-house mammograms used for test (b).
Analog digitized DDSM mammograms and full-field digital in-house mammograms provided as inputs to the CAD system. Digital mammograms present a wider dynamic range, which allows one to perceive a better contrast. Also, tissue close to the skin and the nipple can be observed well on digital mammograms, whereas it is barely visible on analog mammograms.
Per-mammogram and per-case FROC curve analysis of both the previous RFPR and new CAD systems. FPs per mammogram are reported at different sensitivities.
Visual analysis of some FP marks rejected by the RTFPR module: FP marks caused by the overlapping of breast tissues in fatty breasts (top) and FP marks in dense breasts (bottom).
Summary of the per-mammogram and per-case FROC curve analysis for both the previous RFPR and new CAD systems. For the two CAD systems, FPs per mammogram and FP reduction are reported at different per-mammogram and per-case sensitivities.
JAFROC analysis of both the previous RFPR and new CAD systems. For the two CAD systems. JAFROC’s figure of merit and -value associated with the statistical significance of the difference between the two figures of merit are reported on per-mammogram and per-case basis.
Article metrics loading...
Full text loading...