Abstract:

Objective To evaluate the diagnostic performance of a deep learning model based on YOLOv5 in assisting ultrasound physicians with different levels of experience in diagnosing breast imaging reporting and data system (BI-RADS) category 4 breast nodules. Methods A total of 100 patients with suspicious breast malignant lesions who underwent breast ultrasonography at Affiliated Zhongshan Hospital of Dalian University from August 2023 to October 2024 were included, yielding 135 nodules. According to the ultrasonographic (US) BI-RADS classification, and with histopathology as the reference standard, three ultrasound physicians with different seniority levels and an artificial intelligence model independently assessed the benignity or malignancy of the nodules. The diagnostic performance of physicians of different seniority levels, with and without assistance from the artificial intelligence model, was then analyzed. Results Among the 135 nodules, 76 were classified as BI-RADS 4A, 41 as 4B, and 18 as 4C; histopathology confirmed 57 malignant nodules and 78 benign nodules. Using histopathological findings as the reference standard, the YOLOv5 model achieved higher sensitivity, specificity, accuracy, and AUC than the low-seniority physicians, with statistically significant differences (P<0.05). Its sensitivity and accuracy were lower than those of the high-seniority and mid-seniority physicians, with statistically significant differences (P<0.05). After combining the YOLOv5 model with physicians of different seniority levels, the low-seniority group showed an increase in sensitivity from 64.91% to 80.70%, accuracy from 65.18% to 74.81%, specificity from 65.38% to 70.51%, and AUC from 0.618 to 0.782 (P<0.001). In the mid-seniority group, sensitivity increased from 80.70% to 91.22%, accuracy from 77.03% to 80.74%, and AUC from 0.767 to 0.872 (P=0.013). In the high-seniority group, no statistically significant differences were observed in sensitivity, accuracy, specificity, or AUC (P=0.679). Subtype analysis showed that the combined approach was most effective for BI-RADS 4A nodules. In the low-seniority group, diagnostic accuracy for 4A nodules increased from 59.21% to 71.05%; in the mid-seniority group, it increased from 73.68% to 82.89%. Conclusion The YOLOv5 deep learning model can significantly improve the differential diagnostic ability of low- and mid-seniority physicians for BI-RADS category 4 breast nodules, with the greatest added value for 4A nodules. This approach may help reduce unnecessary biopsy and surgery.

Key words: breast neoplasms, ultrasonography, deep learning, BI-RADS, physicians of different seniority