To compare the efficacy of conventional ultrasound and S-Detect in the differential diagnosis of benign and malignant breast lesions.

A total of 468 lesions were identified from 367 patients with breast lesions confirmed by surgery and pathology from June to July in 2018. Both the man-made BI-RADS classifications (still images and dynamic videos identified by three specialist physicians with 1, 4, and 7 years of experience, respectively) and computer S-Detect classification were performed. By plotting the ROC curves of different BI-RADS classification groups, the optimal diagnostic cutoff values were determined. Using pathological results as the gold standard, and the diagnostic test four grids were used to calculate the sensitivity, specificity, accuracy, positive predictive value, and negative predictive value of different BI-RADS classifications and S-Detect classification in the diagnosis of benign and malignant breast lesions, and chi-square tests were used to compare the diagnostic efficacy between groups. The ROC curves of each group were plotted, and the area under the ROC curve (AUC) of each BI-RADS classification group was compared with the S-Detect classification group by the Z-test.

Of 468 breast lesions, 313 were confirmed to be benign lesions and 155 confirmed to be malignant lesions by pathological biopsy. The optimal diagnostic cut-off value was determined to be BI-RADS 4a by plotting the ROC curves for different BI-RADS classification diagnostic groups. The sensitivity of S-Detect classification diagnosis was 93.5%, which was significantly higher than those (69.0% and 72.3%) of the 1-year physician by BI-RADS classifications (still images and dynamic videos, respectively, χ²=30.627 and 24.785, respectively, P=0.000 for both). The specificity of S-Detect classification diagnosis was 83.7%, which was significantly lower than that (92.0%) of the 4-year physician by BI-RADS classification (dynamic videos, χ²=10.124, P=0.001). The differences in other diagnostic comparisons were not statistically significant (P>0.05). The AUC of S-Detect classification was 0.917, significantly higher than those (0.790 and 0.803) of the 1-year physician by BI-RADS classifications (still images and dynamic videos, respectively, Z=5.271 and 4.693, respectively, P<0.0001 for both). The difference between the AUC of S-Detect and that of 4-year physician by BI-RADS classification (still images) was not statistically significant (P>0.05). The AUC of S-Detect was lower than that (0.941) of 4-year physician by BI-RADS classification (dynamic videos, Z=4.327, P<0.0001). The AUC of S-Detect classification was lower than those (0.946 and 0.959) of the 7-year physician by BI-RADS classifications (still images and dynamic videos, respectively, Z=4.225 and 5.477, respectively; P<0.0001 for both).

The S-Detect classification can achieve the diagnostic level of the 4-year physician by BI-RADS classification of still images, but is lower than that by dynamic videos.