# Artificial intelligence–assisted breast ultrasound: modest AUROC improvement and shorter interpretation time without significant change in diagnostic accuracy

## TL;DR

AI assistance in breast ultrasound exams slightly improves diagnostic accuracy and halves interpretation time, but does not significantly change overall diagnostic accuracy.

## Contribution

Demonstrates that AI assistance modestly improves diagnostic discrimination and interpretation speed in breast ultrasound without compromising accuracy.

## Key findings

- AI assistance increased AUROC from 0.921 to 0.953 (p=0.002) and reduced median reading time from 6.0 to 3.0 seconds (p<0.001).
- Accuracy, sensitivity, and specificity did not differ significantly with AI assistance (all p>0.06).
- AUROC improvements were significant for dense breasts and tumors ≤ 2 cm (p<0.001).

## Abstract

To evaluate whether Vis-BUS, a commercial artificial intelligence (AI) breast ultrasound detection and analysis software, improves diagnostic discrimination and interpretation efficiency in breast ultrasound examinations.

This retrospective multi-reader study included 258 breast ultrasound examinations (129 malignant and 129 benign lesions). Six radiologists independently interpreted all cases without AI and, after a two-week washout, with AI assistance. Diagnostic performance metrics, including the area under the receiver operating characteristic curve (AUROC), area under the precision–recall curve (AUPRC), accuracy, sensitivity, and specificity, were compared using multi-reader analysis. Median interpretation time per case was recorded and compared using paired statistical tests.

Vis-BUS assistance modestly increased the pooled AUROC (0.921 vs. 0.953, p = 0.002) and reduced median reading time (6.0 vs. 3.0 s, p < 0.001), whereas AUPRC, accuracy, sensitivity, and specificity did not differ significantly (all p > 0.06). Accuracy (79.1% vs. 83.9%, p = 0.061), sensitivity (94.2% vs. 96.3%, p = 0.243), and specificity (64.0% vs. 71.6%, p = 0.069) showed no significant differences. Median interpretation time decreased from 6.0 to 3.0 s (p < 0.001). Subgroup analyses demonstrated significant AUROC improvements for dense breasts and tumors ≤ 2 cm (p < 0.001 for both).

Vis-BUS AI assistance was associated with improved diagnostic discrimination and shorter interpretation time. However, accuracy, sensitivity, and specificity did not differ significantly. These findings suggest potential efficiency benefits, while the clinical impact remains to be confirmed in prospective multi-center studies.