Welcome to ExpertCAD™, where we tell the truth about CAD's dismal performance in breast mass screening
The purpose of this website is to tell the truth about CAD and why it is not performing in the clinic as expected. Methodology and software downloads are in the JAFROC website.
CAD is an acronym for Computer Aided Detection, algorithms that analyze digital mammograms to find cancers that radiologists might miss. Breast cancer claims about 40,000 lives in the US annually. The basic idea of CAD is valid but 25+ years of NIH funded research has not yielded the desired results. Unlike researchers struggling to obtain NIH funding, and who cannot afford to "rock the boat", I am in a position to speak out. Having worked 30 years from within the system it is my duty to do so. Here are my credentials...
In the early 1980s, as soon as computing became common, researchers began applying computers to a clinical problem, namely, the wide variability in radiologist performance, as much as 40% (!) measured by sensitivity, with a similar variability in specificity, see Beam et al study. The idea was that with computer assist, variability should decrease: e.g., if the radiologists always took CAD's advice, their variability would be zero. The problem is that then their collective performance would be identical to that of CAD, which is currently quite poor.
Shown is how CAD is currently used in a mammographic interpretation.
- The radiologist interprets the image without any help from CAD, Fig. (a).
- The CAD prompts (or "cues") are turned on, indicated by the yellow arrows, meaning, effectively, "have you looked at this region" and "how about this?", Fig. (b).
- The radiologist is free to change their initial interpretation or ignore CAD's advice.
This is called the second reader mode. CAD is not FDA-approved for use as a first reader. In the US, CAD is used in practically every screening mammographic interpretation and is reimbursed under Medicare.
The focus of ExpertCAD™, is developing improved methods for finding breast cancer. As implicit in the name, the intent is to develop an expert level CAD system, not one that is only suitable as a second reader. This will be done by novel CAD training and assessment methods.