Explainable deep learning for predicting HER-2 expression in breast cancer: a multicenter study

Abstract

Background

Human epidermal growth factor receptor 2 (HER-2) is a key biomarker in breast cancer, guiding therapeutic decisions and prognosis. Conventional assessment relies on tissue biopsy, an invasive procedure that may impose both physical and financial burdens on patients.

Purpose

To develop an interpretable deep learning-based imaging framework capable of non-invasively predicting preoperative HER-2 expression.

Material and Methods

We retrospectively analyzed magnetic resonance imaging data and clinical records from 450 patients with pathologically confirmed HER-2 status across four medical centers. Several conventional machine learning algorithms were compared with a deep neural network model. A ResNet-based architecture was used to generate a probability score (D-score) reflecting the likelihood of HER-2 positivity. Independent clinical predictors were identified through logistic regression and integrated with the D-score to construct a combined predictive framework. Model performance was evaluated using receiver operating characteristic analysis, and interpretability techniques were applied to visualize the contribution of individual features.

Results

The combined deep learning model achieved an area under the curve of 0.809 in the external validation cohort, outperforming the clinical model. Interpretability analysis identified the D-score, rim enhancement, and diameter of the largest axillary lymph node as the most influential predictors, consistent with established clinical knowledge.

Conclusion

The proposed model enables accurate, non-invasive, and interpretable prediction of HER-2 expression in breast cancer. It may serve as a preoperative stratification tool, support individualized treatment planning, and reduce reliance on invasive diagnostic procedures.