Preoperative prediction of HER2 status in breast cancer using multiparametric MRI deep learning and clinical features.
Authors
Affiliations (3)
Affiliations (3)
- Department of Nuclear Medicine, the First Affiliated Hospital of Hebei North University, 12 Changqing Road, Qiaoxi District, Zhangjiakou, Hebei 075000, China.
- School of Information Engineering, Hebei North University, Zhangjiakou, Hebei 075000, China.
- Department of Ultrasound, the First Affiliated Hospital of Hebei North University, 12 Changqing Road, Qiaoxi District, Zhangjiakou, Hebei 075000, China. Electronic address: [email protected].
Abstract
HER2 status is an important biomarker in breast cancer, but its assessment primarily relies on core needle biopsy and pathological examination. Therefore, noninvasive preoperative prediction of HER2 status has important clinical value. This study aimed to evaluate the predictive value of models integrating multiparametric MRI deep learning, radiomics, and clinical features for preoperative HER2 status prediction in breast cancer. A total of 257 patients with pathologically confirmed breast cancer were retrospectively enrolled, including 98 HER2-positive and 159 HER2-negative patients. All patients underwent preoperative T1WI, T2WI, DWI, and DCE-MRI. Patients were randomly divided at the patient level into training (n = 174) and internal test (n = 83) cohorts at a 7:3 ratio. Cli, Rad, and DL models were constructed based on clinical, radiomics, and deep learning features, respectively, along with single-sequence fusion models. Subsequently, an mpMRI model integrating radiomics and deep learning features from all four MRI sequences and an mpMRI_Cli model further incorporating statistically significant preoperative clinical features were developed. Model performance was evaluated using the AUC, accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, Brier score, decision curve analysis, calibration curves, and confusion matrices. DeLong's test was used to compare AUCs between models. In the internal test cohort, the AUCs of the T1WI, T2WI, DWI, and DCE-MRI models were 0.666, 0.689, 0.710, and 0.728, respectively. The AUCs of the Cli, Rad, and DL models were 0.571, 0.764, and 0.791, respectively. The mpMRI model achieved an AUC of 0.910 (95% CI: 0.837-0.966), with an accuracy of 0.831, sensitivity of 0.742, and specificity of 0.885. The mpMRI_Cli model achieved an AUC of 0.923 (95% CI: 0.860-0.972), with an accuracy of 0.819, sensitivity of 0.839, specificity of 0.808, and Brier score of 0.122. Although the mpMRI_Cli model had a higher AUC than the mpMRI model, the difference was not statistically significant after Holm correction (P = 0.744). Integrating multiparametric MRI radiomics and deep learning features improved preoperative prediction of HER2 status in breast cancer. Further incorporation of clinical features increased the AUC and sensitivity and reduced the Brier score, although the improvement in AUC was not statistically significant. Both models show potential for noninvasive preoperative HER2 prediction, but further validation in independent external and multicenter cohorts is warranted.