Curr Probl Diagn Radiol. 2025 Jun 3:S0363-0188(25)00114-8. doi: 10.1067/j.cpradiol.2025.06.001. Online ahead of print.
ABSTRACT
BACKGROUND: Breast lesions are a common but often missed incidental finding. We evaluated whether artificial intelligence (AI) algorithms can efficiently detect radiologically significant incidental breast lesions (RSIBLs) missed by original interpreting radiologists (OIRs) on chest CT examinations.
METHODS: This retrospective multi-institutional study analyzed chest CT examinations performed in June 2017 by a national teleradiology practice. Visual classifier (VC) and natural language processing (NLP) algorithms flagged potential RSIBLs, which were reviewed independently by two primary readers; disagreements were adjudicated by a third reader. Sizes and margins of confirmed RSIBLs were evaluated similarly. Differences in size and margin obscuration between RSIBLs missed versus identified by OIRs were statistically assessed (alpha, 0.05). A workflow efficiency analysis was performed.
RESULTS: 3279 of 3541 examinations (92.6 %) were marked negative by both algorithms (i.e., VC-/NLP-) and not reviewed. The two primary readers assessed 262 examinations for RSIBLs, with substantial agreement (kappa, 0.77). After adjudication, 76 RSIBLs were confirmed (73 females, 3 males). Compared with the OIRs, the VC algorithm identified more RSIBLs (90.8 % [69/76] vs 39.5 % [30/76]) though with more false positives (67.9 % [178/262] vs. 3.4 % [9/262]). Among the OIRs, missed RSIBLs had smaller diameters than identified RSIBLs (1.4 cm vs. 3.0 cm; P < 0.001). Our reader workflow reduced the number of images viewed by 97.3 % relative to a hypothetical full double-read approach.
CONCLUSION: An AI-based approach enhanced RSIBL detection rates. Although the AI-based approach also increased the number of false positives, our targeted review process allowed for efficient detection of missed RSIBLs.
PMID:40517117 | DOI:10.1067/j.cpradiol.2025.06.001
