Computer-aided detection worsens screening
On 5 April 2007 NEJM published a comparison of 43 facilities in three states over 1998-2002: where computer-aided detection was adopted, the specificity of screening mammography fell from 90.2% to 87.2%, biopsies rose by 19.7%, and the rise in sensitivity from 80.4% to 84.0% was not significant.
Why it matters
A technology approved by the FDA and paid for by insurers was tested in practice rather than on a test set, and it did worse than reading without it. Whether a system meant to help actually helps was answered from practice.
Breast Cancer Surveillance Consortium data: 222,135 women, 429,345 mammograms, 2,351 cancers diagnosed within a year of screening. Seven facilities (16%) adopted the system; the positive predictive value fell from 4.1% to 3.2%, and the cancer detection rate did not change (4.15 and 4.20 per thousand). Across all 43 facilities the area under the ROC curve was 0.871 with computer-aided detection and 0.919 without. The paper says the FDA approved the technology in 1998 and that Medicare reimburses it. This is an association in an observational study, not a trial; the record does not claim which system each facility used.