A hybrid reading strategy that combines AI with radiologist input reduced the workload for mammogram interpretation by nearly 40% without affecting recall or cancer detection rates, according to a study published in Radiology.

Researchers from Radboud University Medical Center in the Netherlands tested the strategy on more than 41,000 screening mammograms collected between 2003 and 2018 as part of the Dutch National Breast Cancer Screening Program.

The system relies on an AI model that not only estimates the probability of malignancy (PoM) for each case but also quantifies its confidence in that prediction. When the AI confidently identified an exam as normal or suspicious, it either excluded the case from further review or initiated a recall without radiologist input. If the AI was uncertain, the case was reviewed by human readers.
“Although the overall performance of state-of-the-art AI models is very high, AI sometimes makes mistakes,” said Sarah D. Verboom, a doctoral candidate in medical imaging at Radboud. “Identifying exams in which AI interpretation is unreliable is crucial to allow for and optimize use of AI models in breast cancer screening programs.”

Out of the 41,469 mammograms from 15,522 women, 332 screen-detected and 34 interval cancers were identified. Researchers found that about 38% of the exams were confidently classified by AI, allowing those cases to bypass radiologist review. The recall and cancer detection rates were comparable to standard double-reading by radiologists, at 23.6 per 1,000 and 6.6 per 1,000, respectively.

Verboom emphasized the importance of building uncertainty quantification into commercial AI tools. “The key component of our study isn’t necessarily that this is the best way to split the workload, but that it’s helpful to have uncertainty quantification built into AI models,” she said.

The research is part of the aiREAD project and was supported by the Dutch Research Council, Dutch Cancer Society, and Health Holland.

Artificial Intelligence Homepage