Real-time AI-based computer-aided detection/diagnosis (AI-CAD) for breast ultrasound: A prospective, multicenter, multinational study.

Abstract

1569 Background: To evaluate the effectiveness of a real-time AI-based computer-aided detection/diagnosis (AI-CAD) system as a diagnostic decision support tool for breast ultrasound in a real-world clinical setting, conducted as a prospective, multicenter, and multinational study. Methods: From May to December 2024, a total of 75 patients undergoing breast ultrasound were enrolled in a prospective study conducted in Korea (n = 38) and Hong Kong (n = 37). In this study, six experts operated a real-time AI-CAD system (CadAI-B, BeamWorks Inc., Korea) on a tablet PC connected to a handheld ultrasound device during breast ultrasound examinations. Image and clinical data were collected from patients with established ground truth through follow-up, biopsy, or surgery. The AI-CAD system highlights suspicious areas during scanning to assist physicians in detecting breast cancer and supports big data-driven differential diagnosis by providing BI-RADS categories and malignancy scores (0–100%) when the user freezes the image. The diagnostic performance of experts and the real-time AI-CAD system was evaluated using the area under the receiver operating characteristic curve (AUC), along with sensitivity and specificity. Results: The analysis included 75 patients (mean age 55 years, IQR 46–66) with 24 malignancies (32.0%), 45 benign lesions (60.0%), and 6 normal cases (8.0%). The mean breast mass size was 1.2 cm (±1.0 cm): benign 0.8 cm (±0.7 cm), malignant 1.8 cm (±1.3 cm). The BI-RADS category distribution was as follows: for experts—category 1 (4.0%), 2 (21.3%), 3 (24.0%), 4a (16.0%), 4b (18.7%), 4c (4.0%), 5 (12.0%); and for AI-CAD—category 1 (32.0%), 2 (5.3%), 3 (9.3%), 4a (17.3%), 4b (21.3%), 4c (13.3%), 5 (1.3%). The overall diagnostic performance of experts and AI-CAD, as AUCs calculated by BI-RADS, were 0.801 and 0.751, respectively (P = .679). The sensitivity and specificity were 91.7% (22/24) and 68.6% (35/51) in experts and 87.5% (21/24) and 57.8% (32/51) in AI-CAD, respectively (P = .481). Conclusions: In this real-world clinical study conducted across multiple centers and countries, CadAI-B demonstrated performance comparable to that of experts and showed its potential as a valuable diagnostic tool. Clinical trial information: NCT06622967 .