AI Mammography With Uncertainty Quantification Reduces Screening Workload

Background: Artificial intelligence (AI) algorithms in mammography can be used for decision support, triage of screening examinations, and can be used for stand-alone interpretation that is equal or superior to that of a single radiologist reading. Despite this, AI implementation in screening practice is limited because some cancers are still missed. The ability to identify examinations in which the AI interpretation is not reliable is crucial because in these cases, human interpretation is needed. This could substantially reduce workloads by allowing standalone AI interpretation only in cases in which the model performs as well as or better than screening radiologists. The AI model would need to provide not only an assessment of the probability of malignancy but also a rating of how certain it is about that assessment. Uncertainty estimation of neural network predictions is a growing area of research and usually limited to a single network. In contrast, breast cancer detection is ty more...