Computer-aided mammography now in use at Hospitals

Since mid-April, screening mammograms have been read by a radiologist specializing in mammography, who has been assisted by guidance from the computer system. Computer-aided diagnosis is expected to reduce by as much as one-half the number of breast cancers missed by radiologists.

Like the spell-check program on a word processor, which helps writers avoid typographical errors, the computer provides "an automated second opinion, alerting the radiologist to take a closer look at subtle warning signs that might have been missed on the first pass," said project leader Kunio Doi, the Ralph W. Gerard Professor in Radiology and Director of the Kurt Rossman Image Research Laboratories. "We hope this system can significantly reduce the number of undetected breast cancers, even in the best screening programs."

Although mammography is currently the best method for the early detection of breast cancer, even the most skilled mammographers fail to spot between 10 percent and 30 percent of the cancers present in women who have had screening mammograms. In as many as twothirds of these "false negative" mammograms, the radiologist missed signs of cancer that, when the mammogram was more closely examined months or years later, were identified as present and visible.

Eye fatigue and inattentiveness can explain some of those reading errors, but in other cases the mammographic abnormalities caused by the cancer were subtle enough to escape detection, even by top experts in the field.

Up to half of early breast cancers include little clusters of microcalcifications, tiny deposits of calcium that on a mammogram look like a constellation of faint stars. Although these minute specks can be obscured by the wealth of information contained in a typical mammogram, the Chicago research team has taught the computer to recognize these stars, even through the "clouds" of normal tissue, and to alert the radiologist to their presence.

Another 40 percent of early breast cancers will appear on a mammogram as a mass. The computer can detect such masses by contrasting the architectural symmetry of the right and left breasts.

A series of preliminary studies has demonstrated the power of the computer. In the first study, an earlier version of the current system detected 85 percent to 90 percent of breast lesions in a set of 200 subtly abnormal mammograms -- a performance level comparable to that of an experienced radiologist.

In a second study, the computer correctly identified more than 50 percent of breast cancers that had been missed by radiologists. In a third study, the combination of computer and radiologist outperformed a radiologist alone, increasing accuracy from 85 percent up to 92 percent.

A fourth preclinical study, which analyzed 1,200 consecutive screening patients who had mammograms at the Hospitals, confirmed the system's usefulness and reliability in the clinical setting.

"The computer makes its own mistakes," said Robert Nishikawa, Assistant Professor in Radiology and one of the principal investigators in the project. The computer typically directs the radiologist's scrutiny to one or two benign areas on each mammogram, what physicians call false positives. "But radiologists can distinguish between the computer-selected false positives and malignancy," he said.

The computer's weaknesses are sufficiently different from radiologists' shortcomings that the two complement each other well. "It doesn't make human errors," said Robert Schmidt, Associate Professor in Radiology and a member of the research team. "The computer never forgets to look at the edges of the film. It doesn't get distracted by the phone, or by frequent interruptions. It doesn't get tired or cranky, or worry about its career choice."

Maryellen Giger, Associate Professor in Radiology and another principal investigator in the project, said, "Since the computer serves as a second opinion rather than acting alone, it doesn't have to be perfect to make a substantial difference. Radiologists will continue to make the final diagnostic decisions."

At an estimated cost of $100,000, about the same as a full-featured mammography machine, the system may in the future be available even to community radiologists.

This project, which now involves more than 40 researchers, has been under way at the University for more than 10 years, with primary support from the National Cancer Institute, the American Cancer Society, the U.S. Army and the Whitaker Foundation.

The American Cancer Society estimates that 206,000 women will be diagnosed with breast cancer in the United States in 1995. About one in eight American women will develop the disease during their lifetimes. Mammography is the only known method that can find small, curable cancers before they can be felt, and its use has been shown to significantly decrease the number of deaths from breast cancer.