Screening mammograms miss close to one in eight breast cancers. But early research suggests artificial intelligence (AI) could close this detection gap and markedly improve early diagnosis of the disease. Still, questions remain regarding how to best incorporate AI into screenings and whether it’s too soon to deploy the technology.
Already, some radiology clinics are offering AI analysis of mammograms through an add-on cost method.
Mammography patients who visit RadNet facilities, for example, have the option of an additional AI screening of their images. RadNet, the largest national owner and operator of fixed-site diagnostic imaging centers in the United States with more than 370 locations, first launched its AI program in the Northeast. The company has now rolled out its product across all regions in the country.
Because the AI is not reimbursed by insurers, patients must pay a $40 out-of-pocket fee if they want the AI analysis.
“RadNet practices have identified more than 400 women whose cancer was found earlier than it would have been had the AI not been present,” said Greg Sorensen MD, chief science officer for RadNet.
How RadNet’s AI Program Works
Patients coming to RadNet facilities for screening mammography undergo 3D high-resolution mammography that includes the use of 70-micron resolution detector technology, said Dr. Sorensen. The mammogram is reviewed by a qualified radiologist with assistance from two Food and Drug Administration–cleared AI programs, Saige-Q and Saige-Density. The radiologist then makes an interpretation.
Saige-Q is an AI tool that helps identify more suspicious mammograms by providing a quick signal to radiologists if the AI considers a given mammogram to be in a suspicious category, according to Dr. Sorensen. Saige-Density provides a density rating for each mammogram using one of the four standard categories:
- A. Almost entirely fatty
- B. Scattered areas of fibroglandular density
- C. Heterogeneously dense
- D. Extremely dense
Starting in September 2024, the FDA will require all mammogram reports to indicate density.
For patients who choose the $40 add-on service, called Enhanced Breast Cancer Detection, two other FDA-registered AI programs are also applied: Saige-Dx and Saige-Assure. These AI programs go a step further by placing marks on areas within the images that they find suspicious. Mammograms flagged as “high-suspicion” by the AI are then reviewed by a second human radiologist. The first and second radiologists confer to agree on a final diagnosis, Dr. Sorensen explained.
“Our research shows that approximately 20% more cancers are found when the safeguard review process is in place,” Dr. Sorensen said. “We also have seen [30%] decreases in recall rates” — the percentage of screening cases in which further tests are recommended by the radiologist.
Bethesda radiologist Janet Storella, MD, has used the AI program for about 3 years and said the technology has improved her screening performance.
The AI is linked to her practice’s imaging software, and radiologists have the option of turning the AI on at any time during their reading of screening mammograms, Dr. Storella explained. Some radiologists review the mammogram first and then initiate the AI, while others like Dr. Storella turn it on at the start, she said. Once initiated, the AI draws bounding boxes — or outlines — around areas that it deems suspicious.
The AI helps focus Dr. Storella’s attention on suspicious areas and grades the level of suspicion into one of four categories: high, intermediate, low, and minimal, she said.
“I find it especially useful in patients who have dense breast tissue,” said Dr. Storella, medical director of women’s imaging at Community Radiology Associates, a RadNet practice. “In these situations, the tissue on the mammogram is a field of white, and cancers are also white, so you’re looking for that little white golf ball on a sea of snow. The AI really helps hone that down to specific areas.”
About 35% of RadNet’s screening mammography patients have enrolled in the Enhanced Breast Cancer Detection program, according to RadNet data. In a recent study of nine general radiologists and nine breast imaging specialists, all radiologists improved their interpretation performance of DBT screening mammograms when reading with RadNet’s AI versus without it. (An average AUC [area under the receiver operating characteristic curve] of 0.93 versus 0.87, demonstrating a difference in AUC of 0.06 (95% CI, 0.04-0.08; P < .001)