Application value of a deep learning model based on 3D ultrasound videos in assisting radiologists of different experience levels to differentiate benign and malignant breast masses

Objective To establish a deep learning model based on 3D ultrasound videos (DL-3DUV) and investigate its application value in assisting radiologists with different levels of experience to differentiate benign and malignant breast masses. Methods The ResNet50 model was employed to develop DL-3DUV for the classification of benign and malignant breast masses. A retrospective study was conducted using automated breast volume scanner (ABVS) dynamic videos from 400 patients with breast masses (a total of 525 lesions), which were randomly divided into training and testing sets at an 8∶2 ratio. The diagnostic performance of DL-3DUV was compared with that of senior and junior radiologists, both independently and in combination. Results When diagnosing independently, DL-3DUV demonstrated comparable sensitivi-ty (83.33% vs. 81.77%), accuracy (82.50% vs. 84.60%), and area under the curve (AUC) (0.83 vs. 0.85) compared to senior radiologists (all P>0.05), though its specificity was significantly lower (81.58% vs. 87.73%, P<0.05). Compared with junior radiologists, DL-3DUV showed significantly higher sensitivity (83.33% vs. 78.60%), specificity (81.58% vs. 57.00%), accuracy (82.50% vs. 68.37%), and AUC (0.83 vs. 0.68) (P<0.05). The combination of senior radiologists and DL-3DUV achieved higher sensitivity (89.70% vs. 81.77%) and AUC (0.91 vs. 0.85) than senior radiologists alone (all P<0.05), with no significant differences in specificity (91.23% vs. 87.73%) or accuracy (89.17% vs. 84.60%) (all P>0.05). Similarly, the combination of junior radiologists and DL-3DUV significantly improved diagnostic performance compared to junior radiologists alone, with statistically significant differences in sensitivity (88.10% vs. 78.60%), specificity (82.47% vs. 57.00%), accuracy (85.47% vs. 68.37%), and AUC (0.85 vs. 0.68) (all P<0.05). Conclusions DL-3DUV exhibits significant value in differentiating benign and malignant breast masses and is expected to become a useful tool to assist ultrasonographers, particularly for junior radiologists.