… or “Read this mammogram, HAL.” “I’m sorry, Dave. I’m afraid I can’t do that.”
I went to college in the late 1980s, at perhaps the peak of optimism about computer intelligence. Personal computers had just become available and there was a general expectation that computers would soon be driving our cars, accepting our commands in spoken English, and generally doing everything better than humans could.
The reality has been much less consistent. There have been impressive gains in computer intelligence applied to some specific tasks, like chess. But there has been remarkably little progress in others fields, like transcribing spoken language. Transcription software is still notoriously error-prone, and transcription by humans remains much in use.
At first glance, reading a mammogram seems like the perfect task for a computer program. The software would just need to recognize the characteristic appearance of breast cancer and the appearance of normal breast tissue. It would not be biased by factors that can affect radiologists, like fatigue or anxieties about making an error.
Indeed, such software exists. Computer-aided detection (CAD) technology is computer software that performs a second reading of a mammogram which is supposed to point out abnormalities on the mammogram the radiologist may have missed. It does not replace a radiologist’s reading, but was intended to help the radiologist detect more cancers and perhaps detect cancers earlier. It was FDA approved and is currently used in the reading of about three quarters of mammograms in the US.
Except it might not help.
A study published this week in the Journal of the National Cancer Institute looked at 1.6 million mammograms done at 90 facilities over 8 years. Some facilities used CAD, and some did not. The study found that CAD did not lead to increased detection of cancers or to detection of cancers at earlier stages. Worse, CAD led to an increase in false positives – mammograms read as abnormal that led to normal biopsies. That means that CAD led to an increase in biopsies without actually helping patients.
That’s not exactly what we hoped for from intelligent machines. That’s much less like Rosie, the Jetsons’ unflappable household robot, and more like HAL, the computer in 2001: A Space Odyssey.
Of course this study shouldn’t be the last word on CAD. Technologies improve all the time, and the fact that it’s not helpful now doesn’t mean it won’t be in a few years. But until some improvements are made, the best software for reading mammograms is still behind the eyes of a radiologist.
Computers Still Not a Big Help With Reading Mammograms (Wall Street Journal Health Blog)
Mammograms: Computer-aided detection doesn’t help (LA Times Booster Shots)