Abstract
OBJECTIVES: In view of inherent attributes of breast BI-RADS 3, benign and malignant lesions are with a subtle difference and the imbalanced ratio (with a very small part of malignancy). The objective of this study is to improve the detection rate of BI-RADS 3 malignant lesions on breast ultrasound (US) images using deep convolution networks. METHODS: In the study, 1,275 lesions out of 1,096 patients were included from Southwest Hospital (SW) and Tangshan Hospital (TS). In which, 629 lesions, 218 lesions and 428 lesions were utilized for the development dataset, the internal and external testing set. All malignant lesions were biopsy-confirmed, while benign lesions were verified through biopsy or stable (no significant changes) over a three-year follow-up. And each lesion had both B-mode and color Doppler images. We proposed a two-step augmentation method, covering malignancy feature augmentation and data augmentation, and further verified its feasibility on a dual-branches ResNet50 classification model named Dual-ResNet50. We conducted a comparative analysis between our model and four radiologists in breast imaging diagnosis. RESULTS: After malignancy feature and data augmentations, our model achieved a high area under the receiver operating characteristic curve (AUC) of 0.881 (95% CI: 0.830-0.921), the sensitivity of 77.8% (14/18), in the SW test set, and an AUC of 0.880 (95% CI: 0.847-0.910), a sensitivity of 71.4% (5/7) in the TS test set. Compared to four radiologists with over 10-years of diagnostic experience, our model outperformed their diagnoses. CONCLUSIONS: Our proposed augmentation method can help the deep learning (DL) classification model to improve the breast cancer detection rate in BI-RADS 3 lesions, demonstrating its potential to enhance diagnostic accuracy in early breast cancer detection. This improvement aids in a timely adjustment of subsequent treatment for these patients in clinical practice.