Optimizing AI-assisted decision-making in mammography screening: focusing on uncertainties.
Authors
Affiliations (8)
Affiliations (8)
- Advanced Diagnosis and Interventional Radiology Research Center, Imam Khomeini Hospital, Tehran, Iran.
- Department of Radiology, Tehran University of Medical Sciences, Tehran, Iran.
- Department of Computer Engineering, Sharif University of Technology, Tehran, Iran.
- Science and Technology Park of Sharif University of Technology, Tehran, Iran.
- Advanced Diagnosis and Interventional Radiology Research Center, Imam Khomeini Hospital, Tehran, Iran. [email protected].
- Department of Computer Engineering, Sharif University of Technology, Tehran, Iran. [email protected].
- Advanced Diagnosis and Interventional Radiology Research Center, Imam Khomeini Hospital, Tehran, Iran. [email protected].
- Department of Radiology, Tehran University of Medical Sciences, Tehran, Iran. [email protected].
Abstract
To evaluate an AI-based decision-support system for mammography interpretation, assessing its impact on diagnostic performance. This retrospective study analyzed 130,838 mammograms from 35,709 patients with confirmed follow-up outcomes across multiple centers, using the Digital Database for Screening Mammography (DDSM) for external validation. Representing one of the largest datasets from the Middle East, it evaluates model generalizability in underrepresented populations. An AI super-model incorporated separate UNet-based networks for lesion segmentation, a high-resolution model (4096) for calcifications, a multi-resolution ensemble (4096, 2048, 1024) for macro-scale lesions, and a risk-assessment model for malignancy prediction. Radiologists interpreted images under five scenarios: unaided, AI-assisted, AI as the final decision-maker in uncertain cases, union of radiologist and AI decisions in uncertain cases, and standalone AI. AI assistance significantly enhanced performance across all scenarios. Based on aggregated reader performance, AUC improved from 83.04% (95% CI: 79.8-86.3) unaided to 88.40% (95% CI: 85.6-91.2) in the union scenario (p < 0.001). Sensitivity increased from 54.8% (95% CI: 48.6-61.0) to 83.9% (95% CI: 79.3-88.4) (p < 0.001), and specificity rose from 83.6% (95% CI: 82.0-85.3) to 91.1% (95% CI: 89.8-92.4) (p < 0.001). Multi-resolution segmentation effectively detected calcifications and challenging lesions. DDSM validation confirmed model robustness. AI showed particular promise in aiding less-experienced radiologists in underserved areas to improve early cancer detection. AI in mammography interpretation improves diagnostic accuracy and reduces uncertainty, supporting early breast cancer detection and optimized screening workflows. Question Can an AI-based decision-support system improve the accuracy of mammography interpretation, reduce diagnostic uncertainty, and support less-experienced radiologists? Findings In this retrospective multicenter study, AI-assisted mammography improved sensitivity and specificity, increased cancer detection, and reduced recalls and false positives. Critical relevance statement AI integration in mammography enhances diagnostic accuracy, reduces unnecessary recalls, and supports more confident and efficient clinical decision-making, particularly by improving the performance of less-experienced radiologists in underserved and resource-limited settings where expert interpretation may be limited.