Artificial intelligence decision support in automated breast ultrasound: improving diagnostic accuracy and reducing unnecessary biopsies

Abstract

Background: The emerging roles of artificial intelligence (AI) support in the imaging of the breast have led to improved radiologist performance.

Objective: This study assessed the diagnostic performance of the AI decision support in the evaluation of breast masses using automated breast ultrasound (ABUS).

Methods: One hundred eighty-two patients (415 breast masses) who received ABUS were included. Two readers, including an experienced breast radiologist (reader 1) and the breast imaging fellow (reader 2), separately reviewed the ABUS images and the AI decision support according to the American College of Radiology BI-RADS 5th edition guidelines.

Results: In the 415 masses that were evaluated, 395 (95.2%) were benign, and 20 (4.8%) were malignant. The area under the receiver operating curve (AUC) of the AI decision support was 0.74 (95% confidence interval Original article (CI) 0.72–0.77) with a sensitivity and specificity of 100.0% and 48.6%, respectively. The integration of AI decision support significantly increased the AUC for both readers, from 0.82 (95% CI 0.74–0.91) to 0.85 (95% CI 0.76–0.93) for reader 1 (P < 0.001) and from 0.79 (95% CI 0.71–0.88) to 0.81 (95% CI 0.73–0.89) for reader 2 (P < 0.001). Furthermore, the AI decision support led to a 14.2% and 16.9% alteration in BI-RADS, with a 22.2% and 10.7% reduction in biopsies of benign masses for reader 1 and reader 2, respectively.

Conclusion: AI decision support demonstrates diagnostic performance comparable to that of radiologists, exhibiting high sensitivity and a high negative predictive value. Integrating AI into the diagnostic workflow may potentially enhance the diagnostic performance of radiologists across various experience levels and thereby contribute to a reduction in unnecessary biopsies of benign masses.