Judging by their marketing, companies are gearing up for widespread adoption of their artificial intelligence (AI) algorithms by breast screening clinics. Hologic claims, “A breakthrough in early breast cancer detection … empowered by GeniusAI,” while Kheiron entices, “Radiologists – discover how AI is revolutionizing breast cancer screening.” ScreenPoint says it has reached a major milestone and “now is first of its kind to help radiologists read more than 1 million mammograms,” while CureMetrix believes it is providing radiologists with advanced technology that “will support improved clinical and financial outcomes.” A recent review in the British Medical Journal suggests these claims might be premature.
Scientists from the University of Warwick, led by Karoline Freeman (seen below left), were commissioned by the National Screening Committee in the UK “to determine whether there was sufficient evidence to use AI for mammographic image analysis in breast screening practice.” They performed a systematic review of test accuracy and identified 12 studies, all published between 2019 and 2021, that met their inclusion criteria. Five studies used AI as a standalone system, four used AI for triage, and three used AI as a reader aid. No prospective studies were found.
Studies were conducted in a range of geographic regions, including the United States, Sweden, Holland, Germany, and Spain. In most studies, the mammography systems were manufactured by Hologic or Siemens, while the AI systems consisted of both in-house and commercially available algorithms. The authors concluded that the studies were of poor methodological quality, and that “Current evidence for AI does not yet allow judgement of its accuracy in breast cancer screening programmes, and it is unclear where on the clinical pathway AI might be of benefit.”
Researchers from Lund University in Sweden, led by Victor Dahlblom, have recently published a paper in Radiology in which they studied the ability of AI to improve cancer detection in full-field mammograms (FFDM) compared to digital breast tomosynthesis (DBT). The AI algorithm missed 25% of the cancers detected by FFDM, although it did identify 44% of cancers detected by DBT alone. The authors concluded that, while AI did not reach the performance of two readers, it had the potential to close the gap between FFDM and DBT.
Another paper just published in Radiology showed that an in-house AI algorithm predicted the risk of cancer on future screening-detected mammograms better than clinical risk factors, including breast density. Unfortunately, the algorithm performed poorly in predicting interval cancers. As the Warwick group emphasized, well-designed prospective studies will be required to measure the impact of AI in clinical practice.