Abstract
OBJECTIVE: This study aimed to develop a robust, automated framework for predicting HER2 expression in breast cancer by integrating multi-sequence breast MRI with deep learning-based feature extraction and clinical data. The goal was to improve prediction accuracy for HER2 status, which is crucial for guiding targeted therapies. MATERIALS AND METHODS: A retrospective analysis was conducted on 6,438 breast cancer patients (2006-2024), with 2,400 cases (1,286 HER2-positive, 1,114 HER2-negative) selected based on complete imaging and molecular data. Patients underwent 3T MRI scans with T1, T2, and contrast-enhanced (DCE) sequences. Imaging data from four medical centers were standardized through preprocessing steps, including intensity normalization, registration, and motion correction. Deep learning feature extraction utilized ResNet50, VGG16, EfficientNet-B0, and ViT-Small, followed by ICC filtering (≥0.9) and LASSO regression for feature selection. Nomogram construction, ROC analysis, and DCA evaluation were performed to assess model performance. Statistical analyses were conducted using Python and R, with significance set at p < 0.05. RESULTS: In this study, we developed an integrated predictive model for HER2 status in breast cancer by combining deep learning-based MRI features and clinical data. The model achieved an AUC of 0.94, outperforming traditional methods. Analysis revealed significant differences between HER2-positive and HER2-negative groups in tumor size, lymph node involvement, and microcalcifications. Imaging features, such as washout enhancement and peritumoral edema, were indicative of HER2 positivity. After applying ICC filtering and LASSO regression, the selected features were used to construct a nomogram, which demonstrated strong predictive accuracy and calibration. The DCA confirmed the model's clinical utility, offering enhanced decision-making for personalized treatment. CONCLUSIONS: This study demonstrates that integrating deep learning with multi-sequence breast MRI and clinical data provides a highly effective and reliable tool for predicting HER2 expression in breast cancer. The model's performance, validated through rigorous evaluation, offers significant potential for clinical implementation in personalized oncology, improving decision-making and treatment planning for breast cancer patients.