AI in Mammography and Breast Cancer Screening

Screening mammography looks for early breast cancer in women without symptoms. Radiologists compare views of each breast, look for masses, calcifications, asymmetries and architectural distortion, and decide whether to recall the woman for more tests. In much of Europe, two radiologists read each exam independently. In the United States, one radiologist usually reads it. Computer-aided detection is not new. Traditional CAD was approved in the late 1990s and became widely used in the US, but large studies found it did not clearly improve accuracy and added false marks. Modern deep learning systems, such as Lunit INSIGHT MMG, iCAD ProFound AI and ScreenPoint Transpara, learn from very large sets of images with known outcomes and give each exam a suspicion score as well as region markers. The strongest evidence comes from prospective trials in real screening programs. Sweden's MASAI randomized trial, first reported in The Lancet Oncology in 2023, used AI to triage exams: low-risk exams got one reader and high-risk exams got two. It found cancer detection was at least as good as standard double reading, with a large reduction in radiologist reading workload and a similar false-positive rate. Later MASAI results and studies elsewhere, including a large German real-world study, reported higher detection rates with AI support. Important limits remain. Finding more cancers does not automatically mean saving more lives, because some detected cancers are slow-growing, and a key long-term measure is interval cancers, which appear between screenings. Dense breast tissue appears white on mammograms, the same as many tumors, which can hide cancers from both humans and AI. A common misconception is that AI reads mammograms by itself. In current programs, radiologists make the recall decisions.