Artificial intelligence-assisted triage of screening mammograms following breast-conserving therapy: a comparative simulation study.

PURPOSE: To evaluate the diagnostic performance and potential of artificial intelligence (AI)-based triage for screening mammograms following breast-conserving therapy (BCT). METHODS AND MATERIALS: In this retrospective study, consecutive post-BCT mammograms obtained between January and May 2021 were analyzed and divided into ipsilateral and contralateral breasts. Triage was simulated using three models with a commercial AI-based computer-aided detection (CAD), and outcomes were categorized as recall or no recall: (1) original report-based triage, (2) standalone AI, and (3) decision referral AI-triage. Cancer detection rate (CDR), recall rates, and diagnostic performance were evaluated. RESULTS: A total of 1190 women were enrolled. For the ipsilateral breast, 10 mammography-visible recurrences were identified. All three models-original report, standalone AI, and decision referral AI-achieved equivalent CDR (6.6 per 1000) and sensitivity (80%), with recall rates of 3.4%, 23.0%, and 2.8%, respectively. AI-CAD classified 77% of examinations as negative without a reduction in CDR or sensitivity. For the contralateral breast, three recurrences were identified. The original report yielded a CDR of 1.8 per 1000, a recall rate of 1.9%, and 66.7% sensitivity. While AI-CAD triaged 90% of examinations as negative, standalone AI achieved a CDR of 2.6 per 1000, recall rate of 9.8%, and 100% sensitivity. Decision referral AI maintained CDR and sensitivity with a lower recall rate (2.0%). CONCLUSIONS: Our simulation suggests that AI-based triage can exclude a substantial portion of negative mammograms following BCT without a reduction in CDR or sensitivity. Nonetheless, radiologist expertise remains crucial, particularly in interpreting the ipsilateral breast, given the higher false positive rate of AI-CAD.