Preoperatively predicting human epidermal growth factor receptor 2-low expression in breast cancer using neural...

Preoperative prediction of human epidermal growth factor receptor 2 (HER2)-low expression using magnetic resonance imaging (MRI) can enhance the selection of clinical treatment strategies and enhance patient outcomes. Herein, we investigated the value of a neural network model constructed with multiparametric MRI in diagnosing HER2-low breast cancer.

This retrospective study involved two different centers. A total of 895 breast cancer patients (903 lesions) were enrolled from the Second Hospital of Shandong University (known as "Center 1") between January 2015 to December 2022. They were allocated to the training set (626 cases/632 lesions) and the internal validation set (269 cases/271 lesions). The external validation set included 100 patients (100 lesions) from the Qilu Hospital of Shandong University (referred to as "Center 2") between June 2021 to December 2022. All patients were subgrouped into HER2-low and HER2-0 expression groups. We used t-tests, Wilcoxon rank sum tests, and Chi-squared tests or Fisher's exact test to compare the dynamic contrast-enhanced MRI features (morphological/hemodynamic features), and the apparent diffusion coefficient (ADC) values. A neural network model was constructed using the Neuralnet package in R, with the architecture specified as c(5,2) for the hidden layers. Bootstrapping was used for internal validation. The diagnostic performance in the training set was analyzed using receiver operating characteristic (ROC) curves. The clinical effectiveness of the model was validated using a decision curve analysis (DCA).

HER2-low breast cancer lesions had irregular morphology, high early enhancement rate, and low ADC value compared to HER2-0 expressed lesions. The differences were significant (P<0.05). We then constructed a neural network model using these significant variables. ROC analysis showed that the area under the ROC curve of the model for diagnosing HER2-low breast cancer in the training, internal validation, and external validation sets was 0.757 [95% confidence interval (CI): 0.712-0.802], 0.728 (95% CI: 0.658-0.798), and 0.791 (95% CI: 0.693-0.890), respectively. The DCA demonstrated that the net benefit of the model was significantly greater than zero at a predicted probability of 0.764.

The neural network model based on MRI features is an effective tool in predicting HER2-low breast cancer, which may facilitate clinical treatment decision-making.