<p>The HER2 (Human Epidermal Growth Factor Receptor 2) gene in breast cancer serves as a crucial biomarker for prognosis, treatment planning, and therapeutic monitoring, playing a pivotal role in guiding targeted drug selection. Differences in HER2 expression status significantly affect the prognosis and treatment response of breast cancer patients. Early recognition of HER2 expression status based on machine learning methods enables clinicians to adjust treatment regimens proactively and select the most suitable targeted therapies, thereby enhancing treatment efficacy. This paper proposes a model leveraging a Dual-modal Feature Attention Fusion Network (DFAFN) based on ultrasound images, designed to accurately identify HER2 status in breast cancer. Initially, DFAFN utilizes both a HER2 ultrasound image feature extractor and a radiomic feature extractor to capture deep features and radiomic features of the ultrasound images independently. Subsequently, a HER2 status-related feature learning module extracts ultrasound image features and radiomic features correlated with HER2 status. The HER2-related features of two modalities are then integrated through a two-step cross-attention fusion network. Ultimately, a classification module distinguishes among HER2-zero, HER2-low, and HER2-positive statuses. The model achieved an AUC score of 0.8080 for HER2 status identification, with a 95% confidence interval of (0.7753, 0.8262) on the real dataset. Comparative experiments also demonstrate the DFAFN’s superior performance across various metrics relative to both traditional and state-of-the-art methods. Results indicate that DFAFN can assist physicians in accurately identifying HER2 status through breast cancer ultrasound images. Our source code is available at <a href="https://github.com/didadiuouo/DFAFN">https://github.com/didadiuouo/DFAFN</a></p>