Artificial Intelligence Could Help Prioritize Mammograms for Faster Follow-Up
A new Breast Cancer Surveillance Consortium (BCSC) study using data from the BCSC Sacramento Area Breast Imaging Registry finds that the 2% of mammograms with the highest AI scores included half of all screen-detected breast cancers.
Millions of people in the United States are called back each year for additional testing after a screening mammogram. Because diagnostic evaluation usually requires a separate visit, an abnormal screening result can lead to delays, anxiety, additional costs, and time away from work or other responsibilities. These burdens may be especially difficult for people living in rural or underserved communities.
One way to reduce these burdens is to interpret screening mammograms while patients are still at the imaging facility, allowing those with abnormal findings to receive diagnostic imaging during the same visit. However, immediate interpretation of every screening mammogram is generally not feasible because of staffing and workflow constraints. Artificial intelligence (AI) could help by identifying the examinations most likely to benefit from immediate interpretation and prioritizing them for review.
In a new study published in the Journal of the American College of Radiology, researchers evaluated whether AI could support this type of targeted workflow. The study included 3,468 consecutive screening digital breast tomosynthesis examinations performed at UC Davis between October 17 and December 31, 2022. Two FDA-cleared AI algorithms assigned each examination a score reflecting the level of suspicion for breast cancer.
Patient characteristics, mammography results, and cancer outcomes were obtained from the Sacramento Area Breast Imaging Registry (SABIR), a registry within the Breast Cancer Surveillance Consortium (BCSC). SABIR links breast imaging data to the California Cancer Registry, allowing researchers to identify breast cancers diagnosed within one year after each screening examination.
Among the 3,468 screening mammograms, 213, or 6.1%, resulted in a recall for additional imaging. Twenty-five breast cancers were diagnosed within one year, including 20 cancers detected following a positive screening mammogram.
The researchers ranked the examinations from highest to lowest AI score and assessed the recalls and screen-detected cancers included at different thresholds. The 2% of mammograms with the highest AI scores included 50% of all screen-detected cancers and 7.5% to 9.0% of all recalled examinations. Lowering the AI-score threshold would identify more patients who could benefit from expedited evaluation but would also increase the number of examinations requiring immediate interpretation. For example, prioritizing the 10% of examinations with the highest AI scores would include 25.8% to 28.8% of all recalls and 75% to 90% of all screen-detected cancers.
The AI scores were substantially better at identifying examinations associated with cancer than at predicting every radiologist recall. This finding is expected because mammography AI algorithms are designed primarily to detect cancer or cancer-associated imaging features. Radiologists may also recall patients for benign findings that appear suspicious or because of differences in interpretive practice.
These findings support a potential workflow in which AI flags a small number of high-suspicion screening examinations for priority image transfer and immediate, potentially remote, interpretation. Patients with abnormal results could then receive same-day diagnostic imaging, reducing delays and eliminating the need for some patients to return for another visit. Such a strategy may be particularly valuable for mobile mammography programs and for patients who travel long distances or face other barriers to completing follow-up care.
The study was retrospective and included only two AI algorithms, one institution, and 25 breast cancers. Prospective implementation would require radiologist availability, rapid image transfer, information technology support, and sufficient appointment capacity for same-day diagnostic imaging. Facilities would also need flexible workflows because the number of examinations exceeding a selected AI-score threshold may vary from day to day.
The results demonstrate that the value of AI in breast imaging may extend beyond helping radiologists detect cancer. By helping health systems determine which examinations should be interpreted first, potentially even while patients wait, AI could also improve the timeliness, convenience, and equity of breast cancer screening follow-up.
Miglioretti DL, Ponzini M, Ramwala OA, de Bie E, Aminololama-Shakeri S, Morris EA, Lee CI. Using Artificial Intelligence to Improve Timeliness of Follow-Up in Breast Cancer Screening. J Am Coll Radiol. Published online May 2, 2026. doi:10.1016/j.jacr.2026.04.023. PMID: 42082064; PMCID: PMC13265059. [link]
The full article can be found here: