Integrated ultrasound-based radiomics and deep learning models in screening breast intraductal high-risk lesions or carcinoma: a multicenter retrospective study.

Purpose: The aim of this study was to explore the diagnostic performance of ultrasound (US)-based radiomics combined with deep learning (DL) in the screening of high-risk and malignant intraductal breast lesions. Methods: This multicenter retrospective study included patients with breast intraductal lesions from January 2022 to June 2024 from five hospitals in China. In the training set, conventional US images were segmented and radiomics features were extracted. After feature selection using least absolute shrinkage and selection operator (LASSO) regression, a radiomics model was developed using logistic regression, and the DL model was constructed based on ResNet-50. An integrated model was constructed by fusing the predicted probabilities from single models. The diagnostic performance of US, radiomics, DL, and integrated models was compared in the internal and external validation sets. Results: A total of 785 lesions were collected, including 486 benign lesions and 299 high-risk or malignant lesions. In the training set (520 lesions), the integrated model achieved superior performance (area under the curve (AUC) = 0.946 [0.923, 0.964]) to that of the US model (AUC = 0.774 [0.732, 0.816]; Conclusion: The integrated radiomics and DL model demonstrated potential clinical value in screening the high-risk or malignant breast intraductal lesions.