br Risk stratification will be used to guide women to
Risk-stratification will be used to guide women to an adequate screening modality
The modality developments described above could increase the proportion of cancers that are detected at an early, curable, stage. However, they also increase the time and cost of screening. Therefore, it is possible that their use might be limited to women at elevated risk of breast cancer and at elevated risk of non-detection at mammography. In the classic screening study by Rose two ap-proaches to prevention and screening are described: focus on high-risk individuals and screening an entire Daptomycin . Both ap-proaches have advantages and disadvantages. The main disadvan-tage with a focus on high-risk individuals is the inherent difficulty in predicting if and when a woman will have cancer. It has been estimated that only a third of all cancers are due to environmental or genetic risk factors, and the remainder are due to random events correlated with the number of stem cell divisions . Neverthe-less, we believe that risk stratification algorithms will be used, and we believe that they will combine a measure of breast cancer risk with a measure of the risk of mammographic non-detection for that particular woman. The cost-effectiveness perspective of risk strat-ified screening should be an integrated part of future studies in the area. Communicating risk to individual women puts a high demand on robust risk models, communication strategies and counseling-areas that currently are explored and will need attention before setting the full sails for risk-based screening.
Artificial intelligence will potentially enhance many aspects of breast imaging
Artificial intelligence (AI) is a recent buzz word which is used for computer software that performs functions that people would normally associate with an intelligent human mind. The commer-cial advantage of claiming that products “contain AI” has made the abbreviation close to meaningless. In the following we use AI in a narrow sense meaning computer software that has been trained on numerous real-world examples rather than being controlled by
human-specified rules. The rapid development of AI was enabled by massive processing power having rapid access to vast amounts of digital data. Mammography is one of the key application areas for AI developers in radiology due to its large volumes of training cases and the large market size. At the last RSNA annual meeting in November 2018, one could visit more than eight companies offer-ing AI applications to help with tumor detection and decision-making in screening mammography. AI systems can potentially act as an assistant to the radiologist, as one of two independent readers, or as the only reader. Many vendors and academic groups show promising results in their own datasets . A recent study showed similar levels of performance of AI as that of radiologists . The published research has so far been based on datasets enriched with cancer cases and not representative of a true screening population. We are still waiting for neutral performance evaluations in common external datasets representative of a normal screening population. Most AI applications in the area of breast imaging are focused on traditional radiologist tasks, such as tumor detection and differential diagnosis. However, there is also work on-going to assist technologists with instant feedback on positioning of the breast and other image acquisition parameters as well as workflow optimization including automated triage for screening MRI . Other research groups are investigating the use of AI in breast cancer risk prediction to optimize the selection of women to whom MRI-based screening should be offered [23,24]. Currently, we perceive AI applications in the field to be in early phases and we expect that leaf primordia will take another 3e5 years for the field to mature. Regulatory agencies involved in the traditional medical technology field may have to establish new approval pro-cesses. There is also a need to rethink how to distribute the medical responsibility, and failure analyses, between health professionals and medical AI companies. It might be necessary for the NIH and its counterparts in other regions to establish a number of officially validated test databases in which companies are required to run their algorithms before applying for certification for use in screening. A decisive factor for AI adoption will be how it is perceived by the women attending screening. Unpublished results from a large screening participant survey from Sweden indicates that many women, once human-level performance has been proven, would be willing to receive computerized assessments of their mammograms for tumor detection and for breast cancer risk assessment. It is important for breast radiologists to put themselves in the driving seat to best direct the possibilities offered by AI.