AI-enhanced ultrasound video analysis for BI-RADS 4–5 breast lesions: a multicenter study of diagnostic performance

Abstract

Abstract Purpose This study aims to develop a computer-aided diagnosis model for aiding ultrasonographers in classification of benign or malignant breast lesions in BI-RADS categories 4–5 using ultrasound videos. Materials and methods This prospective multicenter study enrolled 532 breast lesions categorized as breast imaging reporting and data system (BI-RADS) 4–5 by ultrasound. Of them, 334 breast lesions were included as training and validation sets. The remaining 198 lesions were assigned to a test set. We acquired both static and dynamic video ultrasound images of lesions for each patient and utilized them to develop a deep learning model. Temporal-guided multi-scale multi-instance learning network (TGMIL) with the video images was also developed. In this study, four metrics were calculated: area under the curve (AUC), accuracy, precision and sensitivity. In addition, pairwise DeLong and McNemar were used as statistical tests. Results A total of 120 patients with 198 videos (benign, 45.8 ± 11.5 years; malignant, 50.7 ± 11.0 years) were evaluated. On the patient-based test set, TGMIL achieved an AUC of 0.845 (95% CI, 0.747–0.944), significantly outperforming junior ultrasonographers (AUC, 0.682; 95% CI, 0.583–0.782; p < 0.05) and attending ultrasonographers (AUC: 0.802; 95% CI, 0.713-0.891; p < 0.05). With TGMIL assistance, diagnostic accuracy improved from 66.7% to 89.2% for juniors ( p < 0.05) and from 80.8% to 90.0% for attendings ( p < 0.05); Sensitivity increased from 67.4% to 94.2% for juniors ( p < 0.05) and from 80.2% to 89.5% for attendings ( p < 0.05). Sensitivity was significantly improved for senior ultrasonographers, from 86.0% to 90.7% ( P < 0.05) with TGMIL assistance. Conclusions The TGMIL model significantly improved the diagnostic performance of junior and attending ultrasonographers for BI-RADS 4-5 breast lesions.