ABSTRACT
OBJECTIVE: The use of artificial intelligence has potential in assisting many aspects of imaging interpretation. We undertook a prospective service evaluation from March to October 2022 of Mammography Intelligent Assessment (MIA) operating "silently" within our Breast Screening Service, with a view to establishing its performance in the local population and setting. This evaluation addressed the performance of standalone MIA vs conventional double human reading of mammograms. METHODS: MIA analyzed 8779 screening events over an 8-month period. The MIA outcome did not influence the decisions made on the clinical pathway. Cases were reviewed approximately 6 weeks after the screen reading decision when human reading and/or MIA indicated a recall. RESULTS: There were 146 women with positive concordance between human reading and MIA (human reader and MIA recalled) in whom 58 breast cancers were detected. There were 270 women with negative discordance (MIA no recall, human reader recall) for whom 19 breast cancers and 1 breast lymphoma were detected, with 1 cancer being an incidental finding at assessment. Six hundred and four women had positive discordance (MIA recall, human reader no recall) in whom 2 breast cancers were detected at review. The breast cancers demonstrated a wide spectrum of mammographic features, sites, sizes, and pathologies, with no statistically significant difference in features between the negative discordant and positive concordant cases. CONCLUSION: Of 79 breast cancers identified by human readers, 18 were not identified by MIA, and these had no specific features or site to suggest a systematic error for MIA analysis of 2D screening mammograms.
