JMIR Publications

ABSTRACT

Background:

Human epidermal growth factor receptor 2 (HER2) status is an important biomarker for guiding anti-HER2 targeted therapy in breast cancer. However, its assessment mainly relies on invasive biopsy and immunohistochemistry, which are associated with invasiveness and interobserver variability. Therefore, developing a reliable noninvasive method for predicting HER2 status is of considerable clinical significance.

Objective:

This study aimed to develop a deep learning-based classification method using dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) to noninvasively predict HER2 status in breast cancer by integrating bias field correction, transfer learning, and an attention mechanism.

Methods:

A clinical DCE-MRI dataset was constructed using real-world data collected from the Second Affiliated Hospital of Dalian Medical University. First, N4 bias field correction was applied to reduce intensity inhomogeneity, and the corrected images were combined with the original images for model training. Subsequently, standardized image preprocessing was performed, and a deep learning classification model was developed using a pretrained ResNet50 backbone integrated with a convolutional block attention module (CBAM). The channel and spatial attention mechanisms were employed to adaptively recalibrate feature representations, thereby enhancing the extraction of imaging features associated with HER2 status. Finally, model performance was evaluated using accuracy, precision, recall, and F1-score.

Results:

Experiments were conducted on a breast cancer HER2 expression dataset constructed from real clinical data. A total of 5503 annotated DCE-MRI image samples were included, comprising 1751 HER2-positive and 3752 HER2-negative samples. The proposed method achieved an accuracy of 99.91%, a precision of 100%, a recall of 99.72%, and an F1-score of 99.86% in classifying HER2-positive and HER2-negative breast cancers, indicating strong classification performance.

Conclusions:

The proposed deep learning method, integrating N4 bias field correction, transfer learning, and channel-spatial attention mechanisms, effectively differentiated HER2-positive from HER2-negative breast cancers using DCE-MRI. These findings suggest that the proposed approach may provide a promising noninvasive tool for assessing HER2 status in breast cancer, with potential clinical applicability.