Abstract
Objectives: To develop a deep learning (DL) model for the automated detection and diagnosis of breast cancer utilizing automated breast volume scanner (ABVS) images, and to compare its diagnostic performance with that of radiologists in screening ABVS examinations. Methods: In this multicenter diagnostic study, ABVS data from 1,368 patients with breast lesions were collected across three hospitals between November 2019 and April 2024. The DL model (VGG19, DenseNet161, ResNet101, and ResNet50) was developed to detect and classify lesions. One-tenth of the cases from Hospital A were randomly selected as a fixed internal test set; the remaining data were randomly divided into training and validation sets at an 8:2 ratio. External test sets were derived from Hospitals B and C. Pathological findings served as the gold standard. Clinical applicability was assessed by comparing radiologists' diagnostic performance with and without DL model assistance. Results: For breast cancer detection, the DL model achieved an area under the receiver operating characteristic curve (AUC) of 0.984 (95% CI: 0.965-0.995) on the internal test set, 0.978 (95% CI: 0.951-0.994) on the external test set 1 (Hospital B), and 0.942 (95% CI: 0.902-0.978) on the external test set 2 (Hospital C). The model demonstrated significantly higher sensitivity (98.2%) and specificity (90.3%) than junior radiologists (P < 0.05), while exhibiting comparable diagnostic reliability and accuracy to senior radiologists. Interpretation time was significantly reduced for all radiologists when using the DL model (P < 0.05). Conclusion: The DL model based on ABVS images significantly enhanced diagnostic performance and reduced interpretation time, particularly benefiting junior radiologists.