Answers are grounded in this paper and cite their sources.
Evaluates diagnostic performance changes in early breast cancer detection using AI-CAD.
Background: AI-based computer-aided detection (AI-CAD) updates aim to enhance performance, yet unintended changes may affect clinical reliability.Following a version update intended to improve overall accuracy and specificity in lesion localization, we evaluated changes in diagnostic performance and detection behavior.Methods: 632 asymptomatic early breast cancer cases (2019)(2020)(2021)(2022)(2023) were analyzed by two versions (v1.1 and v1.2) of FxMammo (FathomX, Singapore) and certified human readers (HR).Sensitivity, specificity, and AUC were calculated by separating cancerbearing breasts from contralateral cancer-free breasts for breast-level analyses.Since FxMammo outputs continuous malignancy probability score, version-specific cutoffs were calibrated to achieve 80% sensitivity in accordance with Japanese screening guidelines.Under this framework, subgroup-specific true positive rate (TPR) and true negative rate (TNR) were examined to assess factor-dependent detection performance. Results:In breast-level analyses, sensitivity, specificity, and AUC were 63.9%, 67.6%, and 0.703 for v1.1; 60.4%, 83.5%, and 0.783 for v1.2; and 85.3%, 84.8%, and 0.850 for HR.Under sensitivity-calibrated cutoffs, the resulting TNR increased from 41.3% in v1.1 to 56.3% in v1.2.In subgroup analyses, imaging findings and clinical T stage were associated with detection outcomes in both versions (p<0.05),whereas breast density, menopausal status, and age were associated only in v1.1.Density-stratified analyses demonstrated TNR improvement in both non-dense and dense breasts, with a greater gain observed in dense breasts (30.3% to 50.4%) than in non-dense breasts (57.7% to 65.2%), although a residual density-dependent difference persisted.Both AI versions detected distinct subsets of cancers missed by HR, suggesting altered detection behavior following the update. Conclusions:Overall diagnostic performance improved following a substantial gain in specificity.However, the update also altered factor-specific detection profiles, underscoring the importance of post-update validation beyond aggregate performance metrics.
No takes yet. Share an insight, caveat, or question.
Mabuchi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: