The rationale for CAD
There is no doubt that eventually (50 years?) CAD will approach expert-level performance. CAD has access to the pixel values in the image, a tremendous amount of information that no human can assimilate. As an example, most humans can resolve only 64 shades of gray (26). Pixels are typically stored at 16 bit precision, meaning the computer can distinguish 65536 (216) shades of gray. The computer is capable of performing, in seconds, complex analysis that would take years for the human.
Principled analysis of the information should allow CAD to approach ideal observer performance. Ideal observer performance is not to be confused with perfect performance, which is only attainable in zero-noise or infinite lesion-contrast images.
The key is how the algorithm is designed and trained to find lesions while avoiding finding non-lesions (this sentence defines search performance, a fundamental concept conspicuously missing in the extensive CAD literature). The premise of ExpertCAD™ is that the current approach to CAD training and evaluation is fundamentally flawed. While we obviously cannot share trade secrets, we can at least show that CAD, in the specific context of finding malignant breast masses in screening mammography, does not work. Unless the CAD community acknowledges that there is problem, the "eventually" mentioned above, could be a long time, and in the meantime lots of women will die unnecessarily.
Unfortunately, the CAD community, for the most part, appears to have given up on improving CAD. Most troubling, they have convinced the NIH to pump new money into the field to work on how radiologists "interact" with CAD, and NIH specifically bars new research in CAD. Believe it or not, they are blaming radiologists for not using CAD properly! They believe it a user-interface problem!
Funding is desparately needed in this field, but not for non-scientific approaches such as improving the GUI.*
*They are careful not to use the term "GUI", instead they use nebulous concepts of "perception" and "interaction", but "what looks like a GUI and walks like a GUI is a GUI".