Mammogram analysis by artificial intelligence (AI) identified a third of interval breast cancers not seen on the original review, a retrospective study of more than 200 interval cancers showed.
AI correctly spotted 73 of 224 interval breast cancers on digital breast tomosynthesis (DBT) exams. In a separate review of 1,000 screening DBT exams, AI accurately located 84% of true-positive cancers, 86% of true negatives, and 73% of false positives.
The results suggest AI has the potential to decrease the rate of interval breast cancers, which tend to be more aggressive and have worse outcomes, reported Manisha Bahl, MD, of Massachusetts General Hospital in Boston, and co-authors in Radiology.
"The results are very promising," Bahl told MedPage Today. "What's unique about our study is that it was applied to digital breast tomosynthesis, which has largely become the standard of care in the United States. Prior studies have focused on 2D mammography, so we didn't know what results to expect, what percentage of cancers could be detected by AI. I was surprised that about a third of cancers were detected on retrospective evaluation, and I think there's real promise and potential for AI to help us in the clinic."
The introduction of DBT was expected to help reduce the rate of interval breast cancers, she added. Though long-term data are not yet available, several studies have not shown a meaningful decrease in interval cancers. The lack of improvement suggests that women at risk of interval breast cancer would benefit from additional imaging, such as contrast-enhanced MRI or contrast-enhanced mammography.
"The problem is, we don't necessarily know which patients will develop interval cancers and should undergo that type of imaging," said Bahl. "A strength of AI is that it could be applied to every screening mammogram."
Though limited by data from a single institution, the study provided valuable insight into the question of whether AI has the ability to identify interval cancers, according to the authors of an accompanying editorial.
"The meticulous lesion-specific validation represents a critical strength of the study," wrote Christoph Lee, MD, of the University of Wisconsin in Madison, and Hannah S. Milch, MD, of UCLA Health in Los Angeles. "Previous validation studies have often relied on examination-level AI scores without confirming lesion-level accuracy. Such an approach risks attributing successful cancer detection to AI even if the algorithm flagged unrelated benign areas on the mammogram, potentially overstating clinical utility."
"Ultimately, this retrospective analysis highlights the critical need for prospective randomized controlled trials to rigorously validate the role of AI in breast cancer screening," Lee and Milch added. "These trials would elucidate the impact on interval cancer detection in real-world practice where the radiologist interacts with the AI and includes the AI inputs in the broader context of their imaging interpretation. Prospective studies are essential to understand the complexities of these AI-radiologist interactions and accurately measure clinical outcomes associated with the use of AI."
Summarizing the background of the study, Bahl and colleagues noted the aggressive biology, rapid growth, and poor prognosis for interval breast cancers. Several prior studies explored AI to detect false-negative cancers at screening with digital mammography, but none had examined use of AI in the setting of DBT mammographic screening.
To add contemporary data to the issue, investigators retrospectively used an AI algorithm to analyze screening DBT images for women who subsequently developed interval breast cancers, referring to cancers that arose between screening 3D mammograms after a negative screen. The study included women who had interval breast cancers diagnosed from February 2011 to June 2023.
Using an FDA-cleared AI algorithm, Bahl and colleagues assigned a score of 0-100 for AI-marked lesions on DBT sections. Each AI-positive result, defined as a score ≥10, was reviewed independently by two breast imaging radiologists to confirm that AI annotations corresponded with the site of the subsequent interval breast cancer. Investigators used the same threshold (AI score ≥10) in the analysis of 1,000 DBT images that were categorized into true positive, true negative, and false positive.
AI correctly identified 32.6% of the interval breast cancers. Features of AI-detected cancers included larger size at surgical pathology (37 vs 22 mm for AI-negative results, P<0.001) and positive axillary lymph nodes (41.3% vs 22.8%, P=0.01).
"I reviewed each of the cases and felt that the algorithm was better at detecting certain types of findings, for example, masses and architectural distortion that could be masked by dense breast tissue," said Bahl.
With regard to the AI-negative results, the analysis shed little light on whether AI missed a lesion in hiding or no lesion existed at the time of screening.
"I was impressed with AI's ability to detect some very subtle findings that, prospectively, I don't know that I would have called a finding, and whoever read the mammogram didn't call a finding," she said. "I do think there is a proportion of exams not detected by AI where there simply wasn't a finding to detect."
Disclosures
The study was supported by Lunit.
Bahl disclosed relationships with Hologic, NIH, and 2nd.MD, as well as being a member of the Radiology editorial board. One co-author reported receiving institutional grants from the Radiological Society of North America and GE.
Lee disclosed relationships with DeepHealth/RadNet, UpToDate, Oxford University Press, McGraw Hill, NIH, and the American College of Radiology
Milch disclosed relationships with the Wyeth Foundation, California Breast Cancer Research Program, as well as a grant pending from the Patient-Centered Outcomes Research Institute.
【全文结束】
