Development and Validation of a Deep Learning System to Differentiate HER2-Zero, HER2-Low, and HER2-Positive Breast...

Accurate identification of Human Epidermal Growth Factor Receptor 2 (HER2) expression levels remains a fundamental requirement for tailoring therapeutic strategies in breast cancer management, as these levels dictate the eligibility for specific targeted biological therapies. Prior research has shown that radiomic features derived from Magnetic Resonance Imaging (MRI) can provide non-invasive insights into the molecular subtypes of malignant breast lesions by analyzing texture and intensity patterns. Traditional histopathological assessments, while considered the gold standard, are inherently invasive and may not fully capture the spatial heterogeneity of the entire tumor volume during a single biopsy. Clinicians have increasingly looked toward advanced imaging modalities to supplement biopsy results and provide a more comprehensive view of tumor biology through the analysis of dynamic contrast enhancement patterns. Despite the promise of radiomics, the specific utility and reliability of deep learning algorithms in distinguishing between HER2-zero, HER2-low, and HER2-positive categories have remained largely unproven in large-scale multicenter studies. This absence of evidence motivated the current investigation into the efficacy of automated computational systems for preoperative molecular profiling using high-resolution Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) data.

This investigation seeks to construct and evaluate a robust deep learning framework utilizing Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) for the precise categorization of HER2 expression across the full clinical spectrum. The researchers focused on creating a system capable of performing both automated tumor segmentation and subsequent classification into three distinct HER2 categories to streamline the diagnostic pipeline. By leveraging a large multicenter dataset comprising over one thousand patients, the study aimed to ensure the generalizability and clinical relevance of the developed ResNetGN models in diverse medical settings. The project specifically addressed the need for a non-invasive tool that could differentiate HER2-low status from HER2-zero and HER2-positive statuses preoperatively, which is an essential distinction for modern oncology. Establishing a reliable computational method for these distinctions could significantly impact the selection of targeted therapies, particularly for patients who might benefit from novel antibody-drug conjugates. The study also intended to validate the segmentation accuracy of the automated model against manual delineations performed by experienced radiologists to ensure the spatial precision of the deep learning analysis.

The research team retrospectively analyzed data from 1,294 breast cancer patients across three different medical centers to provide a diverse training and testing environment for the deep learning algorithms. Imaging was performed using 3 Tesla (3T) scanners employing specific protocols including T1-weighted 3D fast spoiled gradient-echo, T1-weighted 3D enhanced fast gradient-echo, and T1-weighted turbo field echo sequences to capture high-resolution temporal data. The cohort was partitioned into a training set of 811 patients, an internal testing set of 204 patients, and an external testing set of 279 patients to rigorously assess model performance. An automated segmentation network processed the Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) data to isolate tumor volumes, which were then evaluated using the Dice Similarity Coefficient (DSC) to measure overlap with manual ground truth. Following segmentation, the ResNetGN deep learning architecture was trained to execute binary classification tasks for each Human Epidermal Growth Factor Receptor 2 (HER2) status against all other categories using the extracted volumetric features. Performance metrics including the Area Under the Curve (AUC) from Receiver Operating Characteristic (ROC) analysis, sensitivity, and specificity were calculated to determine the diagnostic accuracy of the system.

The ResNetGN models exhibited strong performance in identifying HER2-low status, achieving AUC values of 0.820 in the training set and 0.787 in the external validation set, indicating high discriminative power. Automated segmentation reached high precision, with Dice Similarity Coefficient (DSC) values ranging from 0.85 to 0.90 when compared to manual expert segmentation across all internal and external datasets. For the differentiation of HER2-zero from other statuses, the system yielded AUCs of 0.782 in training, 0.776 in internal testing, and 0.768 in external testing, reflecting consistent results across centers. The classification of HER2-positive tumors resulted in AUC scores of 0.792, 0.745, and 0.781 for the training, internal, and external cohorts, respectively, confirming the model's versatility. Statistical analysis confirmed that all performance metrics were significant, with P-values consistently falling below the 0.05 threshold for all primary and secondary classification tasks. The results indicated that the deep learning system maintained consistent diagnostic capability even when applied to data from independent external institutions using different imaging hardware.

The integration of deep learning with DCE-MRI provides a viable pathway for the non-invasive preoperative assessment of HER2 expression levels in breast cancer patients, potentially reducing the reliance on invasive biopsies. These findings suggest that automated systems can effectively handle the complexities of tumor segmentation and molecular classification with high accuracy, offering a scalable solution for clinical radiology departments. The ability to distinguish HER2-low from HER2-zero status is particularly relevant given the emerging therapeutic options specifically targeting low-HER2-expressing tumors that were previously grouped with HER2-negative cases. Future clinical workflows might incorporate these ResNetGN models to assist radiologists in providing more detailed prognostic information before surgical intervention or the initiation of systemic therapy. The successful external validation underscores the potential for this technology to be deployed across different healthcare facilities using various 3T imaging platforms without significant loss in diagnostic performance. The researchers conclude that this DCE-MRI-based system offers substantial therapeutic implications by refining the selection process for HER2-targeted treatments and improving personalized medicine in oncology.