BY CHARLES BANKHEAD A computer-aided detection system correctly identified almost 80% of breast lesions initially overlooked on screening mammography, a California radiologist reported at the RSNA meeting on Monday. The CAD system proved capable of identifying both masses and calcifications that were missed on mammography. The results suggest that computer-aided detection might have a role in screening mammography, said Dr. Robyn Birdwell of Stanford University. The findings came from a study of 286 breast cancers that were not identified on screening mammography. A panel of radiologists blinded to the results retrospectively determined that 115 of the lesions were "actionable," requiring a callback or biopsy. Birdwell and her colleagues used the CAD system to evaluate the mammograms, which were digitized for the CAD assessment.

Several exhibits in the RSNA InfoRad area demonstrate use of computer-assisted diagnosis.

"The neural network detection system marks features that might be calcifications or lesions that might be masses or distortions," Birdwell said. "On a normal mammogram, there will be an average of 2.5 marks."

To be correct, a CAD mark had to match the lesion type and be in the correct area in at least one view of the breast. A mark for calcification or mass was considered correct if a lesion had features of both abnormalities. A CAD mark also was considered correct if it fell anywhere within a widespread area of calcification.

The 115 actionable masses comprised 80 lesions characterized as masses, masses with calcifications, a single case of dilated duct, and a single focal area of asymmetry. The remaining 35 lesions consisted only of calcifications.

Overall, the CAD system marked 73% of the mass-like lesions and 86% of the calcifications. The system added an average of 4.3 marks per mammogram.

In evaluating factors associated with missed lesions on mammography, Birdwell found that the CAD system identified 83% of calcifications in dense breasts, 91% of overlooked calcifications, and 80% of cases involving distracting lesions. For mass-like lesions, CAD correctly marked 77% of cases involving distracting lesions, 85% of lesions at the edge of glandular tissue, and 78% of so-called busy breasts, characterized by spotty areas of glandular tissue.

"From these results we conclude that at screening mammography, CAD might be useful in marking overlooked calcifications and masses that represent breast cancers, and that CAD might have potential for mitigating detection errors," Birdwell said.