AI in Mammography: A Turning Point in Breast Cancer Detection
Despite a 50% reduction in breast cancer deaths over the past 3 decades, the disease still claims tens of thousands of lives each year—a toll that arguably could be reduced through the integration of AI into mammographic screening. That case is being compellingly made by Rachel F. Brem, MD, Director of Breast Imaging and Intervention and Professor and Vice Chair of the Department of Radiology at George Washington University in Washington, DC, who has been deploying AI in routine clinical practice, at no additional charge to her patients, for 3 years.
“We still lose 46,000 women a year. We have to do better,” Dr. Brem said in her presentation at the 43rd Annual Miami Breast Cancer Conference. She covered the full scope of AI’s application in breast imaging—from improving interpretive accuracy on standard 2D and 3D mammography, to leveling the performance gap between general and subspecialist radiologists, to detecting cancer years before it is visible to the radiologist’s eye.
Despite mounting evidence of such advances, AI is currently used in only 20% of U.S. mammography centers, she indicated. To further illustrate the point, Dr. Brem asked her audience of oncology providers whether their imaging centers had incorporated AI, and only a few raised their hands.
The Fundamental Need: Dense Breasts and Diagnostic Limits
Mammography does have well-documented limitations—in particular, falling short in women with dense breasts, who account for much of the screening population. “In women with dense breasts, we know that more than 50% of breast cancers are not mammographically visible. That keeps us up at night as breast imagers,” Dr. Brem said.
“Right now, for any woman with dense breasts, we recommend additional screening, often with MRI, but with 39 million mammograms performed and half the women having dense breasts, that’s not a feasible solution,” she said. The problem is compounded by the low positive predictive value (PPV) of biopsy, roughly 20%, translating into many unnecessary procedures.
Detection Gains: 2D and 3D Mammography
Clinical evidence for AI’s diagnostic contribution has grown steadily since early studies demonstrated significant improvements in cancer detection when radiologists used AI assistance vs standard reading alone.1 The gains are particularly pronounced in 3D mammography, or tomosynthesis, which now accounts for at least 90% of exams in U.S. mammography facilities.
When AI has been applied to digital breast tomosynthesis reads, the data are striking, she said, citing a University of Pennsylvania study in which the area under the receiver operating curve (ROC) improved by 5.7%, sensitivity increased by 8.4 percentage points, specificity rose by 6.9 percentage points, recall rates fell by 7.2 percentage points, and reading time was reduced by 52.7% with AI, all statistically significant improvements over unaided readings.2
“It just gives you the confidence to get through that dense mammogram quicker, faster,” Dr. Brem said. “It used to be that many of us wouldn’t read screens later in the afternoon, but with AI support, it makes it much easier, makes it faster, and makes for more comfortable reading.”
The efficiency implications extend well beyond individual radiologist comfort. The United States faces a well-documented radiologist shortage, with imaging volume growing at 3% to 4% annually while the radiologist workforce expands at just 1%. Some institutions are reporting 6-month waits for mammograms. AI’s capacity to shorten read times translates directly into expanded access.
Closing the Gap: AI and Health Equity
One of the greatest benefits from this gain in detection is the boost it gives to the general radiologist, elevating the generalist’s performance to the level of a breast subspecialist. General radiologists interpret the vast majority of mammograms in the United States, particularly in underserved communities and in the military.
A National Mammography Database Study drives home that need. In the study, only 63% of radiologists read within an acceptable recall rate range; the cancer detection rate fell within acceptable bounds for only 77% of radiologists and the PPV for biopsy was within the acceptable range for just 52%.3
Against that backdrop, the University of Pennsylvania study is informative, having examined AI’s benefits specifically for general and subspecialty radiologists.2 For breast subspecialists, sensitivity rates without AI were 81.7%, improving to 85.9% with AI assistance. For general radiologists, they were 71.5% without AI, improving to 84.0%. General radiologists attained enough benefit from AI to reach a level of sensitivity comparable to their subspecialty colleagues.
“General radiologists perform like breast subspecialists with the addition of AI. That will really help us with health-care disparity,” Dr. Brem commented.
Finding Cancer Before It Can Be Seen
Indisputably, AI improves accuracy as much as a second pair of eyes. In one of the seminal studies, the magnitude of improvement in AUC would have resulted in missing 18% fewer cancers.1 In detecting interval cancers, a Swedish retrospective study of 429 interval cancers found a significant correlation between interval cancer classification groups and AI risk score (score of 1-10 derived from a deep learning model).4 AI gave one in three (143/429) interval cancers a risk score of 10, of which 67% (96/143) were either classified as having minimal signs or as false negatives, but 58% (83/143) were correctly located by AI and could therefore potentially have been detected at screening with the aid of AI. A review of the preceding screening exam showed that 19% of interval cancers had at least minimal signs of malignancy and were correctly localized by AI. “With AI, we will be able to decrease the number of interval cancers,” she predicted.
Perhaps most strikingly, AI offers the ability to identify malignancy years before the trained human eye can see it. In four independent studies, AI-assisted mammography detected tumors up to 5 years earlier than standard mammographic interpretation.5–8
Risk Stratifying With AI
AI’s emerging role in individualized risk assessment may prove to be as consequential as its diagnostic contributions. A retrospective study on almost 3,000 prior screening exams of 1,602 women with screen-detected and interval cancers showed that about 30% of those women would have been designated at highest risk prior to screening by an AI risk model.9 The findings suggest that the use of AI in mammography screening might lead to earlier detection of breast cancers.
Looking ahead, a deep learning model to assess breast cancer risk can be part of a more streamlined risk-based screening approach. In fact, this approach was superior to traditional models in identifying patients destined to develop cancer in a large screening cohort in which mammographic biomarkers alone outperformed the Tyrer-Cuzick and other widely used risk assessment tools.10 A 2D mammography AI platform, Clairity Breast, has already received FDA clearance for 5-year short-term risk assessment, opening the door to a more precision approach to supplemental screening.
“We will be able to markedly decrease the number of women that we recommend MRI or adjunctive screening at all in the United States, with far-reaching ability to find more cancers with fewer exams,” Dr. Brem predicted.
The MASAI Trial and the European Model
The clearest evidence for AI’s clinical viability at a population level came from the MASAI trial, a randomized, prospective study conducted in Sweden on 105,934 women and published last year.7 The trial integrated AI (Transpara version 1.7.0) into a national screening program and produced results that drew attention from the broader oncology community: a 29% increase in cancer detection, with the difference accounted for almost entirely by small, node-negative invasive cancers. The PPV of recall was significantly higher, and radiologist workload fell by 44%.
“The MASAI study demonstrates that integrating AI into screening mammography improves accuracy, reduces workload, and identifies cancers missed by human readers alone,” she said, noting that six European countries now use AI as a replacement for the second reader traditionally required by national screening programs.
Dr. Brem does not, however, endorse fully autonomous AI reading at this time, noting that stand-alone AI performs comparably to breast subspecialists but that the regulatory and practice standard frameworks are not yet in place. She also raised a more nuanced point: the cancers detected by AI and by human radiologists are not always the same entities. “Together, we’ll be able to find more cancers than each approach individually,” she explained.
She added that she is frequently asked whether AI poses a threat to the radiologist’s role. “AI is not going to replace me or other radiologists,” she foresees. “AI will replace radiologists that don’t incorporate AI.”
DISCLOSURE: Dr. Brem reported personal disclosures for BeSound, Breathe Biomedical, Screenpoint Medical, and EosDx.
REFERENCES
Rodríguez-Ruiz A, Krupinski E, Mordang JJ, et al: Detection of breast cancer with mammography: Effect of an artificial intelligence support system. Radiology 290:305-314, 2019.
Conant EF, Toledano AY, Periaswamy S, et al: Improving accuracy and efficiency with concurrent use of artificial intelligence for digital breast tomosynthesis. Radiol Artif Intell 1:e180096, 2019.
Lee CS, Moy L, Hughes D, et al: Radiologist characteristics associated with interpretive performance of screening mammography: A National Mammography Database (NMD) Study. Radiology 300:518-528, 2021.
Lång K, Hofvind S, Rodríguez-Ruiz A, Andersson I: Can artificial intelligence reduce the interval cancer rate in mammography screening? Eur Radiol 31:5940-5947, 2021.
Gjesvik J, Moshina N, Lee CI, et al: Artificial intelligence algorithm for subclinical breast cancer detection. JAMA Netw Open 7:e2437402, 2024.
Dembrower K, Crippa A, Colón E, Eklund M, Strand F; ScreenTrustCAD Trial Consortium: Artificial intelligence for breast cancer detection in screening mammography in Sweden: a prospective, population-based, paired-reader, non-inferiority study. Lancet Digit Health 5:e703-e711, 2023.
Hernström V, Josefsson V, Sartor H, et al: Screening performance and characteristics of breast cancer detected in the mammography screening with artificial intelligence trial (MASAI): a randomised, controlled, parallel-group, non-inferiority, single-blinded, screening accuracy study. Lancet Digit Health 7:e175-e183, 2025.
McKinney SM, Sieniek M, Godbole V, et al: International evaluation of an AI system for breast cancer screening. Nature 577:89-94, 2020.
Larsen M, Olstad CF, Koch HW, et al: AI risk score on screening mammograms preceding breast cancer diagnosis. Radiology 309:e230989, 2023.
Lehman CD, Mercaldo S, Lamb LR, et al: Deep learning vs traditional breast cancer risk models to support risk-based mammography screening. J Natl Cancer Inst 114:1355-1363, 2022.
ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.
