Artificial Intelligence Detection Scores in Screening Mammography for Early Breast Cancer Alerts.
Authors
Affiliations (7)
Affiliations (7)
- Barts Health NHS Trust, London, United Kingdom.
- Department of Oncology and Pathology, Karolinska Institutet, Stockholm, Sweden.
- Linköping University, Linköping, Sweden.
- Linköping University Hospital, Linköping, Sweden.
- Diagnostic Radiology, Department of Translational Medicine, Lund University, Malmö, Sweden.
- Skåne University Hospital, Malmö, Sweden.
- Medical Diagnostics Karolinska, Karolinska University Hospital, K7 Onkologi-Patologi, K7 Forskning Strand, 171 77 Stockholm, Sweden.
Abstract
Background Artificial intelligence (AI)-based computer-aided detection (CAD) systems have demonstrated clinically significant performance in mammographic screening in both retrospective and prospective studies. However, the role of AI in early cancer detection requires further evaluation. Purpose To demonstrate how AI scores in individuals diagnosed with breast cancer may increase up to 10 years before their diagnosis, compared with individuals who remain cancer-free. Materials and Methods A retrospective study using the Validation of Artificial Intelligence for Breast Imaging (VAI-B) database, covering four regions in Sweden from January 2008 to April 2019, was carried out. For each individual, all mammographic examinations were included (maximum of 10 years before diagnosis in those who developed cancer). Upsampling was performed to provide a representative 1% cancer proportion. Three commercial AI CAD systems (AI-1, AI-2, and AI-3) provided raw examination-level AI scores, which were converted to rank-order percentile AI scores. Temporal changes in AI scores were analyzed, and receiver operating characteristic curve analyses were conducted. Results A total of 31 394 individuals (with 88 963 examinations) were included, of whom 12 072 (38.5%) were diagnosed with cancer. The median age at screening was 57.6 years (IQR, 48.9-65.2 years). At 90% specificity, the proportion of cancers potentially flagged by AI-1, AI-2, and AI-3 was 12.7%, 13.8%, and 17.0%, respectively, at 10 years before diagnosis; 19.0%, 19.6%, and 19.7%, respectively, at 6 years before diagnosis; and 24.2%, 23.3%, and 25.2%, respectively, at 4 years before diagnosis. For all time points combined, excluding screen detection, the area under the receiver operating characteristic curve (AUC) ranged from 0.63 to 0.67 for the three AI CAD systems, compared with an AUC of 0.57 for mammographic density as the predictor (all <i>P</i> < .001). Conclusion AI scores from sequential mammograms in individuals diagnosed with breast cancer showed elevated scores up to 10 years before diagnosis, which could potentially be used as an early alert for supplemental imaging. © RSNA, 2026 <i>Supplemental material is available for this article.</i>