Evaluation of a deep transfer learning-based ultrasound model for predicting HER2-positive breast cancer

This study aimed to evaluate the accuracy of an ultrasound imaging-based deep transfer learning model for predicting human epidermal growth factor receptor 2 (HER2)-positive breast cancer and to explore its potential advantages and clinical applications. This retrospective study included ultrasound images from 492 patients with breast cancer who were treated at Gongli Hospital (Shanghai, China) between December 2023 and January 2025. The dataset was randomly divided into a training dataset (n = 343), a validation dataset (n = 73), and an internal test dataset (n = 76) in a ratio of 7:1.5:1.5 for model development, parameter optimization, and internal evaluation. In addition, an independent external cohort consisting of 72 patients from another hospital (January 2025 to June 2025) was used as an external test dataset to evaluate the generalizability of the model. Based on a transfer learning framework, classification models were developed using six convolutional neural network backbone architectures, involving GoogLeNet, ResNet-18, ResNet-50, DenseNet-161, MobileNetV2, and EfficientNet-B0. Model performance was comprehensively evaluated using sensitivity, specificity, accuracy, positive predictive value, negative predictive value, the area under the curve (AUC), calibration curves, and decision curve analysis (DCA). The model demonstrating the best overall performance in the internal test dataset was subsequently validated using the external test cohort to further assess its generalization capability. In the internal test dataset, the AUC values of the six models ranged from 0.913 to 0.989, with DenseNet-161 (AUC = 0.989) and EfficientNet-B0 (AUC = 0.987) demonstrating the strongest discriminative capacity. Among these architectures, EfficientNet-B0 achieved the most favorable overall performance, characterized by superior accuracy and specificity. When evaluated on the external test dataset, EfficientNet-B0 maintained excellent predictive performance, yielding an AUC of 0.972. DCA further indicated that this model provided a greater net clinical benefit across a wide range of threshold probabilities, while calibration analysis demonstrated good agreement between predicted probabilities and observed outcomes. In conclusion, ultrasound-based deep transfer learning model demonstrated remarkable potential for the noninvasive prediction of HER2-positive breast cancer. Among the evaluated architectures, EfficientNet-B0 exhibited the most robust overall performance and strong generalizability, indicating noticeable promise for future clinical translation in supporting precision diagnosis and individualized management of breast cancer. As a non-invasive preoperative auxiliary screening tool, this tool promotes the preliminary identification of high-risk, HER2-positive patient cohorts prior to tissue biopsy or surgical intervention, thereby establishing a substantive radiological basis to guide clinical decision-making.