AI software modestly improves AUROC and reduces interpretation time in breast ultrasound reading

In a retrospective multi-reader cohort study, six radiologists interpreted 258 breast ultrasound examinations (129 malignant and 129 benign lesions) with and without assistance from Vis-BUS, a commercial AI detection and analysis software. The study compared diagnostic performance and interpretation time between AI-assisted and unassisted reads.

With AI assistance, the pooled area under the receiver operating characteristic curve (AUROC) increased modestly from 0.921 to 0.953 (p = 0.002). Median interpretation time per case decreased from 6.0 to 3.0 seconds (p < 0.001). However, key diagnostic accuracy metrics showed no significant differences: accuracy was 79.1% vs. 83.9% (p = 0.061), sensitivity was 94.2% vs. 96.3% (p = 0.243), and specificity was 64.0% vs. 71.6% (p = 0.069).

Safety and tolerability data were not reported. Key limitations include the retrospective design and the use of a multi-reader study with a washout period, which may not reflect real-world clinical workflow. The study demonstrates an association between AI use and improved AUROC with faster interpretation, but the lack of significant change in accuracy, sensitivity, or specificity suggests the clinical impact on diagnostic performance may be limited. Prospective studies are needed to determine if these findings translate to improved patient outcomes.