Computer screening of mammographic data can identify half the breast cancers that radiologists initially overlook, results of a Chicago study indicate.
Computer analysis picked up 12 of 23 cancers that were not detected by an initial screening mammogram. Even on second observation, radiologists picked out only two of the 12 cancers, said Robert Nishikawa, Ph.D., at an RSNA scientific session.
The findings came from an evaluation of an automated detection program for breast masses and clustered microcalcifications. Among 10,000 women screened by the computer program, 79 subsequently developed breast cancer. Fourteen cases initially were true negative. Of the remaining 65 cancers, the computer program identified 44, as compared to 61 for radiologists who reviewed the mammograms.
The computer program excelled, however, on a subsequent analysis that focused on 42 women whose initial screening mammograms were negative. A second review showed than 19 of the cancers were truly occult, while the remaining 23 were visible on second observation. The program retrospectively identified 12 of the missed lesions, as compared to 2 by radiologists. Nishikawa said that prospective use of the computer program could have identified the cancers up to a year prior to the lesions' ultimate detection.
For the entire cohort of 79 patients, the computer program picked up 15% more cancers (12), which is comparable to the increased detection rate reported in the literature for double reading by radiologists, said Nishikawa, a researcher in the department of radiology at the University of Chicago.
During the evaluation, the false-positive rate for the computer program ranged between 1.4 per image during the first 1,000 cases and 2.2 during the most recent 1,000 cases. Nishikawa said an update of the computer algorithms should decrease the rate at which false results occur.
"The computer program detected 50% of cancers that were initially overlooked by radiologists," said Nishikawa. "Several large clinical trials are being planned, which should give us a much better idea about the accuracy of the computer program as compared to radiologists experienced in reading mammograms."
