Deep learning for predicting the human epidermal growth factor receptor 2 status of breast cancer liver metastases based on contrast-enhanced computed tomography: a development and validation study

BACKGROUND: This study investigated the value of a deep learning (DL) model based on computed tomography (CT) enhancement for predicting human epidermal growth factor receptor 2 (HER2) expression in patients with liver metastasis from breast cancer. METHODS: Data were collected for 151 female patients with liver metastasis from breast cancer who underwent abdominal enhanced CT examination in the Department of Radiology at the Affiliated Hospital of Hebei University between January 2017 and March 2022. Liver metastases were confirmed in all patients by pathology. The HER2 status of the liver metastases was assessed and enhanced CT examinations were performed before treatment. Of the 151 patients, 93 were HER2 negative and 58 were HER2 positive. Liver metastases were manually labeled with rectangular frames, layer by layer, and the labeled data were processed. Five basic networks (ResNet34, ResNet50, ResNet101, ResNeXt50, and Swim Transformer) were used for training and optimization, and the model’s performance was tested. Receiver operating characteristic (ROC) curves were used to analyze the area under the curve (AUC), as well as the accuracy, sensitivity, and specificity of the networks in predicting HER2 expression in breast cancer liver metastases. RESULTS: Overall, ResNet34 demonstrated the best prediction efficiency. The accuracy of the validation and test set models in predicting HER2 expression in liver metastases was 87.4% and 80.5%, respectively. The AUC, sensitivity, and specificity of the test set model in predicting HER2 expression in liver metastases were 0.778, 77.0%, and 84.0%, respectively. CONCLUSIONS: Our DL model based on CT enhancement has good stability and diagnostic efficacy, and is a potential non-invasive method for identifying HER2 expression in liver metastases from breast cancer.