Introduction

Multiparametric magnetic resonance imaging (mpMRI) has found extensive application in the evaluation of the human epidermal growth factor receptor 2(HER2) expression status in breast cancer. However, existing methods have not fully harnessed the interrelatedness and complementarity among the mpMRI data, potentially leading to information redundancy or wastage . This study aimed to explore the interrelated and complementary features in different MRI sequences of breast cancer and to develop an effective deep learning model capable of integrating mpMRI information for accurate prediction of HER2 expression status.

Methods

A total of 390 patients with breast cancer who underwent preoperative dynamic contrast-enhanced (DCE), diffusion-weighted imaging (DWI), and T2-weighted imaging were retrospectively recruited from two institutions, including 282 HER2-negative and 108 HER2-overexpressing cases. The corresponding apparent diffusion coefficient (ADC) maps were obtained from DWI. The dataset was divided into a training set with 273 cases, a validation set with 39 cases, and a test set with 78 cases.[1] To investigate the correlation across sequences, the feature of each sequence was projected into a latent space. Simultaneously, a contrastive learning algorithm was employed to maximize the similarity and minimize the distance among the feature representations of all sequences. Aligning sequence features in this manner provided access to sequence-shared features, which helped to eliminate redundant information within mpMRI features. Furthermore, sequence-specific features as a complement to sequence-shared features were introduced to avoid wastage of mpMRI information. Since sequence-specific features focused on the internal similarity within each sequence, we divided the breast cancer tumor region into several image patches and utilized a contrastive learning technique to maximize the feature similarity within these patches. Subsequently, an adaptive weighting scheme was used to fuse these two types of features for HER2 expression prediction. Our proposed method was compared with widely used deep learning methodes, including single-modal method (classified by ResNet) and multi-modal methods (fused features extracted by ResNet with concatenation, summation, or maximum operation).

Results

When using only sequence-shared features or sequence-specific features, the model achieved AUCs of 0.725 and 0.734 in distinguishing between HER2-negative and HER2-overexpressing, respectively. When these two types of features were fused by adaptive weights, the model's AUC was improved to 0.743 (Table 1). The proposed method demonstrated the best predictive performance for HER2 expression status in the test cohort, despite the lack of significant differences compared to traditional single-modal and multi-modal learning approaches (all P > 0.05 by DeLong’s test).

Conclusion

Through the fusion of sequence-shared and specific features, we effectively harnessed the interrelatedness and complementarity of mpMRI information, which allowed us to construct a predictive model for HER2 expression status in breast cancer with high performance.

Discussion

Through sequence sharing and fusion of specific features, we effectively leverage the interconnectedness and complementarity of mpMRI information to construct a high-performance predictive model of HER2 expression status in breast cancer. This study deeply explored the correlation and complementary features between different MRI sequences of breast cancer, and established a deep learning model to accurately predict HER2, which greatly improved the treatment and prognosis of breast cancer, and provided a clearer idea for clinicians to guide patients to make treatment decisions.

Acknowledgements

Thanks to doctors and scholars who contributed to this study.

References

-