Noninvasive Identification of Breast Cancer HER2 Status by Deep Learning on Multiparametric MRI Images

Motivation: Multiparametric magnetic resonance imaging (mpMRI) offers valuable insights for predicting HER2 expression. However, when fusing mpMRI features, redundancy or wastage of information may impact model performance. Goal(s): Our aim was to construct an effective deep learning model by incorporating the interrelated and complementary features of different MRI sequences. Approach: Leveraging a contrastive learning approach, we aligned features across sequences and within each sequence separately to obtain sequence-shared and sequence-specific features. Subsequently, these two features were fused by utilizing an adaptive weighting scheme. Results: When compared to widely used deep learning approaches, our method achieved the best AUC of 0.743. Impact: The method explored the interrelated and complementary features of different MRI sequences, which outperformed widely used deep learning methods in terms of performance. This method was expected to have a positive impact on the accurate prediction of HER2 expression status.