Related Experiment Video
Updated: Jul 14, 2026
Tracking the Mammary Architectural Features and Detecting Breast Cancer with Magnetic Resonance Diffusion Tensor Imaging
Published on: December 15, 2014
An interpretable model based on unenhanced MRI radiomics and clinical features for the identification of HER2
Mengyi Shen1, Dingyi Zhang2, Xin Huang1
1Department of Radiology, Affiliated Hospital of North Sichuan Medical College, No.1 Maoyuan South Road, Shunqing District, Nanchong, Sichuan, China.
Purpose:
To develop a model combining unenhanced MRI radiomics and clinical features for identifying HER2-low and HER2-overexpressing breast cancer. SHapley Additive exPlanations (SHAP) was used to interpret the model.
Methods:
A total of 182 patients (102 HER2-low and 80 HER2-overexpressing) were retrospectively analyzed and split 7:3 into training (n = 127) and test (n = 55) sets. Radiomics features were extracted from T2-weighted imaging (T2WI), apparent diffusion coefficient (ADC) maps, and dynamic contrast-enhanced MRI (DCE-MRI), with the optimal features selected. The top-performing radiomics model was integrated with significant clinical features to establish the combined model. Model performance was assessed by area under the curve (AUC), with DeLong's test employed for AUC comparisons. SHAP was used to interpret feature contributions.
Results:
The T2WI, ADC, and DCE radiomics models achieved training/test set AUCs of 0.693/0.647, 0.675/0.633, and 0.698/0.619, respectively. The T2WI + ADC radiomics model achieved the highest performance, with AUCs of 0.794 in the training set and 0.761 in the test set. Edema demonstrated significant clinical between-group differences (P = 0.044). The combined model integrating radiomics features from the T2WI + ADC model with edema achieved peak AUC values of 0.819 and 0.806, respectively. DeLong's test confirmed the combined model's superiority over the T2WI (P = 0.007, 0.032), ADC (P = 0.002, 0.019), and DCE (P = 0.016, 0.025) models in the training and test sets. ADC-derived MaximumProbability emerged as the most influential radiomics predictor.
Conclusion:
The combined model integrating radiomics of T2WI and ADC maps with edema demonstrates potential as a non-invasive, contrast-free imaging approach for differentiating HER2-low and HER2-overexpressing statuses, identifying candidates for HER2-targeted therapies. SHAP enhances the interpretability of the model.