Mammographic density assessment by an artificial intelligence model for breast cancer detection in BreastScreen Norway.
Authors
Affiliations (12)
Affiliations (12)
- Department of Breast Cancer Screening, The Cancer Registry of Norway, Norwegian Institute of Public Health, Oslo, Norway.
- SPKI-The Norwegian Centre for Clinical Artificial Intelligence, University Hospital of North Norway, Tromsø, Norway.
- Department of Clinical Medicine, Faculty of Health Sciences, UiT The Arctic University of Norway, Tromsø, Norway.
- Department of Radiology, Møre og Romsdal Hospital Trust, Ålesund, Norway.
- Department of Health Sciences in Ålesund, Faculty of Medicine and Health Sciences, Norwegian University of Science and Technology (NTNU), Trondheim, Norway.
- Department of Radiology and Nuclear Medicine, St. Olavs University Hospital, Trondheim, Norway.
- Department of Radiology, Østfold Hospital Trust, Kalnes, Norway.
- Institute of Clinical Medicine, University of Oslo, Oslo, Norway.
- Department of Radiology, Innlandet Hospital Trust, Hamar, Norway.
- Department of Radiology, Vestfold Hospital Trust, Tønsberg, Norway.
- Department of Breast Cancer Screening, The Cancer Registry of Norway, Norwegian Institute of Public Health, Oslo, Norway. [email protected].
- Department of Health and Care Sciences, Faculty of Health Sciences, UiT The Arctic University of Norway, Tromsø, Norway. [email protected].
Abstract
The sensitivity of mammographic screening is lower for women with mammographically dense vs fatty breasts. We aimed to explore automated mammographic breast density and malignancy risk scores generated by an artificial intelligence (AI) model for breast cancer detection, stratified by mammography vendor. This retrospective study included information from 200,000 examinations within BreastScreen Norway. An automated volumetric breast density (VBD) assessment and risk scores were obtained from a commercial AI model. The continuous VBD output from the AI model was categorized into four groups, VBD 1-4, with 15%, 40%, 40%, and 5% of the examinations in each group for exam level and for right and left breast. Area under the receiver operating characteristic curve (AUC) for breast cancer detection was calculated using a continuous AI risk score. Analyses were performed for exam-level and breast-level measures. At exam level, the difference in AUC estimates between the VBD groups was statistically significant, with the highest AUC observed for VBD 1 (0.955, 95% CI: 0.937-0.992) and lowest for VBD 4 (0.857, 95% CI: 0.803-0.910). Stratified by vendor, 3.3% and 6.7% were classified as VBD 4 for A vs B. At breast level, 18.5% of the examinations were classified with different VBD groups for the right and left breast. Despite high AUC estimates across all VBD groups and the potential advantages of an automated density assessment, several factors, such as vendor and breast vs exam level, must be carefully considered before implementing automated density measures in screening programs. Question What are the benefits and drawbacks of mammographic density assessment using an AI model for breast cancer detection? Findings The AI model's performance for cancer detection was promising but decreased with increasing density category. Density distribution varied by equipment vendor and examination level. Clinical relevance AI risk assessment at the current examination was promising despite the variation by mammographic density. However, discrepancies in density by equipment vendors and examination level highlight the need for careful consideration before use in risk-stratified screening.