What question did this study set out to answer?

The aim is to create and validate radiomics models that predict breast cancer molecular subtypes using ultrasound and DCE MRI.

April 16, 2026Open Access

The radiomics models based on ultrasound and DCE MRI for predicting molecular subtypes of breast cancer

Key Points

The aim is to create and validate radiomics models that predict breast cancer molecular subtypes using ultrasound and DCE MRI.
Collected data from 655 breast cancer patients from two centers.
Divided 551 patients from one center into training and internal testing sets.
Employed eight machine learning algorithms to develop ultrasound and MRI radiomics models.
Tested models for performance in both internal and external validation sets.
The combined model (R MRI+US) achieved the highest predictive performance with AUC of 0.973 in the training set.
In the internal testing set, the AUC was 0.872, outperforming single-modality models.
The external validation set showed an AUC of 0.835 for R MRI+US, indicating strong predictive capability.

Abstract

Abstract Background Preoperative distinguish between luminal and non-luminal cancers is crucial for treatment decisions making in patients with breast cancers, as neoadjuvant treatment is recommended for non-luminal tumors instead of luminal tumors. Nevertheless, existing methods have several limitations and there is no effective non-invasive approach to predict molecular subtypes of breast cancer. Purpose To construct and validate an radiomics model based on ultrasound and dynamic contrast enhanced magnetic resonance imaging (DCE MRI) for preoperative non-invasive prediction of breast cancer molecular subtypes. Methods A total of 655 breast cancer patients from two centers were retrospectively collected, with 551 patients from center 1 allocated in a 3:1 ratio into a training set ( n = 413) and an internal testing set ( n = 138), and 104 cases from center 2 serving as an independent external validation set. Eight machine learning algorithms were employed to construct ultrasound radiomics model (R US ), MRI radiomics model (R MRI ), and combined model (R MRI+US ) based on both ultrasound and MR images. The models were tested on the internal testing set, and the algorithm with the best predictive performance was selected for further model performance validation on the external validation set. Results The R MRI+US constructed by support vector machine showed the best predictive performance, with AUC of 0.973 (95%CI: 0.958, 0.988) and 0.872 (95%CI: 0.805, 0.939) in the training and internal testing sets, respectively, significantly higher than those of the R US and R MRI (all P < 0.05). The AUC of the R MRI+US was 0.835 (95%CI: 0.750, 0.921) in the external validation set outperformed single-modality radiomics models (R US , 0.684 95%CI: 0.560, 0.808, P = 0.066; R MRI , 0.740 95%CI: 0.620, 0.860, P = 0.113), although there was no significant statistical difference. Conclusion The radiomics combined model based on pretherapeutic ultrasound and DCE MRI demonstrates potential value for differentiating luminal and non-luminal breast cancers, and may serve as a complementary tool to support clinical decision-making.

Bookmark

View Full Paper