Predicting breast cancer response to neoadjuvant therapy by integrating radiomic and deep-learning features from
Yabin Zhang1,2, Jiumei Cai3, Chunxiao Cui4
1College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, 110169, China.
Background:
Non-invasive prediction of pathological complete response (pCR) in breast cancer patients undergoing neoadjuvant therapy (NAT) is crucial for adjusting surgical strategies and optimizing treatment plans. This study aims to develop a predictive model that integrates traditional radiomics and 3D deep learning features from early and peak phases of dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) to predict pCR after NAT.
Methods:
This retrospective study included 234 breast cancer patients from two hospitals who received neoadjuvant therapy. Dataset 1 (n = 204) was used for model development, and Dataset 2 (n = 30) was used for external validation. Traditional radiomics and 3D deep learning features were extracted from both the whole DCE-MRI image and the tumor region of interest (ROI). Features from different sources were integrated, followed by feature selection using independent sample t-tests and least absolute shrinkage and selection operator (LASSO) regression, and the top ten discriminative features were selected for model training. Logistic regression was used to build predictive models, and their performance was evaluated using receiver operating characteristic curves and area under curve (AUC). The DeLong test was used to assess differences between the AUC values of different models, and SHAP (SHapley Additive exPlanations) analysis was employed to examine the relationship between the features of the models and pCR.
Results:
For models using only traditional radiomics features, the combined model integrating early and peak phases of DCE-MRI provided the best pCR prediction. The performance of this combined model was further enhanced by adding 3D deep learning features. The optimal model (RD_EP), which integrated radiomics and deep learning features from early and peak phases of DCE-MRI, achieved AUC values of 0.892 (95% CI: 0.853-0.922) on Dataset 1 and 0.825 (95% CI: 0.713-0.886) on Dataset 2. The DeLong test showed that RD_EP had statistically significant differences compared to other prediction models (p < 0.05). SHAP analysis demonstrated that two radiomics texture features contributed the most to the model.
Conclusion:
Integrating traditional radiomics and 3D deep learning features from different phases of DCE-MRI can accurately predict pCR to NAT in breast cancer accurately. Multi-phase imaging and diverse features are important to improve predictive accuracy and the constructed model may help guide personalized treatment strategies.
