Validation of an artificial intelligence program in the characterization of breast nodules by ultrasound.

INTRODUCTION: Breast ultrasound is a widely accessible imaging method but highly operator-dependent. Artificial intelligence (AI) may improve breast lesion characterization, aiding in diagnostic decisions. OBJECTIVE: To validate an AI system (Koios DS v3.1) in the BI-RADS classification of breast lesions on ultrasound. METHODS: This cross-sectional diagnostic study included 100 women with breast lesions on ultrasound (July 2022-July 2023), later submitted to histopathology. BI-RADS classifications by conventional ultrasound were compared with AI-based classifications and histopathological findings. Diagnostic agreement and validity measures were calculated. RESULTS: The median patient age was 58.5 years (range, 31-89). AI identified lesions in 93 % of cases. Moderate agreement (Kappa 0.41-0.60) was found between AI and conventional ultrasound BIRADS classification (Kappa = 0.405). When compared with histopathology, AI showed a Kappa of 0.626, with 95.6 % sensitivity, 68.0 % specificity, and a 2.1 % false-negative rate. Disagreement was significantly higher for lesions situated in the lower or central quadrants of the breast (OR = 4.55; p = 0.009) and in irregular heterogeneous areas (OR = 8.27; p < 0.001). CONCLUSION: AI demonstrated high sensitivity and low false-negative rates in classifying breast lesions by ultrasound, showing potential as a complementary diagnostic tool. However, limitations persist, especially in irregular heterogeneous areas and for lesions situated in the lower or central quadrants of the breast.