Nomogram based on deep learning of mammography for the prediction of HER2 expression in breast cancer

Breast cancer is the most common malignancy in women, with its incidence increasing each year. Subtypes defined by human epidermal growth factor receptor 2 (HER2) status exhibit distinct molecular and clinical characteristics, influencing treatment strategies and prognostic outcomes. This study aimed to develop two separate binary nomograms to identify the three HER2 breast cancer subtypes by integrating mammographic imaging without manual mass segmentation and clinical parameters.

This study comprised a total of 543 individuals diagnosed with breast cancer, with 406 patients from the main campus of Jining Medical University Affiliated Hospital and 137 patients from Jining Medical University Affiliated Hospital (Taibai Lake Campus). Patients from the main campus of Jining Medical University Affiliated Hospital were randomly assigned to a training set and an internal test set at a 7:3 ratio, while the 137 patients from Jining Medical University Affiliated Hospital (Taibai Lake Campus) served as an independent external test set. A pre-trained ResNet-50 model was used to extract deep features from mammography images. After feature selection using the Pearson correlation coefficient and the least absolute shrinkage and selection operator (LASSO) regression, the Light Gradient Boosting Machine (LightGBM) algorithm was used to construct the prediction model based on the extracted deep learning (DL) features. The Rad-score was calculated as a linear combination of the features selected by LASSO. Clinical variables were analyzed using univariate and multivariate logistic regressions, with maximum tumor diameter and carcinoembryonic antigen (CEA) levels identified as significant predictors. Subsequently, two visualized nomograms were constructed by integrating imaging features (Rad-score) with key clinical variables, including maximum tumor diameter and CEA.

The two nomogram models demonstrated superior predictive performance compared with models based on single features in both internal and external validation cohorts. The nomogram distinguishing HER2-negative from HER2-positive tumors achieved areas under the receiver operating characteristic curves (AUCs) of 0.913 and 0.875 in the internal and external validation cohorts, respectively. Similarly, the nomogram differentiating HER2-zero from HER2-low tumors achieved AUCs of 0.905 and 0.882 in the internal and external validation cohorts, respectively. In both cases, the nomogram models outperformed the corresponding single-feature models. Decision curve analysis (DCA) indicated that the nomogram provided substantial net benefits across both datasets. Calibration curves further demonstrated a high level of agreement between predicted and observed outcomes.

This study proposes an interpretable and reproducible DL algorithm combined with clinical variables to construct a nomogram for the preoperative noninvasive prediction of HER2 expression status in breast cancer. The nomogram demonstrated robust stability and reliability in distinguishing among different HER2 expression subtypes.