Abstract
OBJECTIVES: This study aimed to investigate the potential of ultrasound (US) radiomics approaches based on machine learning (ML) and deep learning (DL) in the early management from diagnosis to biopsy suggestion of Breast Imaging Reporting and Data System (BI-RADS) 4 breast lesions. METHODS: From January 2020 to December 2022, this multicenter study prospectively evaluated a dataset of dual-view greyscale US images in 1054 patients with 1062 BI-RADS 4 lesions from three centers. Three types of radiomic approaches were developed and compared: (i) the ML-based radiomics; (ii) the DL-based radiomics; and (iii) the dual-core driven hybrid radiomics (Dc-HR) approach. The performance and unnecessary biopsy rate of each model and shear wave elastography (SWE) or contrast-enhanced ultrasound (CEUS) assessment was assessed and compared. RESULTS: The diagnostic performance of Dc-HR (area under the curve [AUC]: 0.944–0.980) outperformed DL-based radiomics (AUC: 0.839–0.964) and ML-based radiomics (AUC: 0.544–0.882). The Dc-HR achieved a better performance in identifying malignancy than the assessment with SWE (AUCs: 0.584–0.725) and CEUS (AUCs: 0.813–0.840). The Dc-HR approach could reduce 90.70–100% unnecessary biopsies in BI-RADS 4a lesions without extra SWE or CEUS examinations, while maintaining a high positive predictive value of 96.43–100% in BI-RADS 4b and 4c lesions. The Dc-HR approach applied with dual-view greyscale US images yielded the highest net benefit according to the Decision curve analyses in comparison to other approaches. CONCLUSION: The Dc-HR approach might be a promising method in diagnosing BI-RADS 4 breast lesions and reducing unnecessary biopsy rate significantly, especially in areas where extra SWE or CEUS examinations are not accessible easily. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s40644-025-00973-y.