An AI mammography tool did not lead to significant benefits in a real-world single-reader study published March 18 in the American Journal of Roentgenology.
Researchers led by Emily Ambinder, MD, from Johns Hopkins Medicine in Baltimore, MD, found that a federally approved AI interpretation tool for mammograms did not lead to significant changes in breast cancer detection and recall.
“The present study failed to identify such benefits from AI assistance,” Ambinder and colleagues wrote.
Recent studies have shown the potential benefits of using AI-assisted mammography in breast cancer screening. One study published in January analyzed results from the Mammography Screening with Artificial Intelligence (MASAI) trial, which found that AI-assisted mammography led to more interval cancers being found and reduced radiologist workload. Another prospective study published in March found that AI-supported breast cancer screening that excludes low-risk mammograms from radiologist reading may be safe and effective.
However, these European studies took place in double-reading settings, while U.S. practices employ single reading. The researchers called for more assessments in the U.S. with this in mind.
Ambinder and colleagues studied the performance of radiologists when using a U.S. Food and Drug Administration-approved AI interpretation tool (Transpara Version 1.7.4-A, ScreenPoint Medical) in a real-world single-reading setting.
The study included all screening mammograms performed using digital breast tomosynthesis (DBT) in 2024 and 2025. The team also formed a comparison group from DBT exams performed in 2022 and 2023.
Final analysis included 24,520 screening mammograms from the same number of women after the AI tool was implemented. The comparison group consisted of 21,630 screening mammograms from the same number of women.
The team found no significant differences between the two groups in terms of recall rate or cancer detection rate.
|
Comparison between pre-, post-AI implementation in breast cancer screening |
|||
|
Measure |
Pre-AI |
Post-AI |
p-value |
|
Recall rate |
12.3% |
12.8% |
0.13 |
|
Cancer detection rate per 1,000 women |
6.7 |
7 |
0.69 |
AI categorized 61.3% of exams as low risk, 34.9% as intermediate risk, and 3.8% as elevated risk. For the low-, intermediate-, and elevated-risk categories, the recall rate was 8.5%,17.9%, and 35.7%, respectively (p < 0.001). And the cancer detection rate per 1,000 women was 0.5, 8.4, and 98, respectively (p < 0.001).
All 91 cancers diagnosed in women classified by AI as elevated risk corresponded to AI-marked findings. The same went for 65 of 72 cancers diagnosed in women classified by AI as intermediate risk.
The AI tool missed 15 of the 171 diagnosed cancers. These included all eight cancers deemed low-risk and the seven cancers for intermediate-risk exams that did not correspond with AI-marked findings. The missed cancers in the low-risk group included six invasive ductal carcinomas (IDCs) and two cases of ductal carcinoma in situ (DCIS). The intermediate-risk group included four IDCs and three cases of DCIS.
With the slight increases seen in the study, the authors suggested that the results may support the use of AI for patient stratification.
Still, they noted that breast cancer measures “did not significantly change after implementation of AI for mammography interpretation.”
Read the full study here.