Abstract
OBJECTIVE: To evaluate the performance of ultrasound-based neural networks in predicting HER2 status in invasive breast cancer (IBC) patients, comparing DenseNet201, ResNet50, Breast Imaging Reporting and Data System (BI-RADS), and a multilayer perceptron (MLP) model. METHODS: Between March 1 and December 30, 2019, 268 female patients with IBC underwent ultrasound-guided core needle biopsy. A total of 1127 ultrasonic images were collected, divided into a training set (70%) and an internal validation set (30%). The HER2 status was predicted using BI-RADS, MLP, ResNet50, and DenseNet201 models. The diagnostic performance of these models was evaluated using accuracy and the area under the receiver operating characteristic curve (AUC). RESULTS: BI-RADS demonstrated the weakest prognostic capability, with an AUC of 0.526, sensitivity of 74.7%, and specificity of 67.4%. The MLP model showed moderate performance with an AUC of 0.637 and accuracy of 75.1%. Among CNN models, DenseNet201 outperformed ResNet50, achieving an AUC of 0.660 and an accuracy of 73%, compared to ResNet50's AUC of 0.537 and accuracy of 67%. For distinguishing HER2-low and HER2-zero expression levels, the MLP model exhibited the highest AUC of 0.790, followed by DenseNet201 at 0.783. In external validation, DenseNet201 demonstrated a robust AUC of 0.860 (95% CI: 0.674-1.000; P < 0.05). CONCLUSIONS: Ultrasound-based DenseNet201 outperformed BI-RADS and ResNet50 for predicting HER2 status in IBC, offering a promising, non-invasive diagnostic tool for clinical application.