Scholarlink

Background <p>Breast cancer (BC) is the most common malignant tumor in women globally. Ki-67, a vital marker for prognosis, is currently detected invasively. Non-invasive magnetic resonance imaging (MRI) prediction faces challenges due to intratumoral heterogeneity.</p> Materials and methods <p>This retrospective study included 254 breast cancer patients from two centers, divided into training set (142 patients), internal validation set (60 patients), and external test set (52 patients). T2-weighted imaging (T2WI) and dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) were analyzed. Traditional radiomics features were extracted from intratumoral, habitat subregions, 5/10-mm peritumoral rings, image fusion. A pre-trained ResNet-50 model extracted 2.5D deep learning features. Feature selection used intraclass correlation coefficient (ICC), Z-score normalization, T-tests, Pearson correlations, and the least absolute shrinkage and selection operator (LASSO). A baseline clinical model was constructed using clinical and qualitative MRI semantic features. Models were built using Support Vector Machine (SVM), Random Forest (RF), and Extra-Trees (ET). Model performance was evaluated via the area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F-1 score. Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to the final convolutional layer of ResNet50 to spatially localize the decision-critical regions. Shapley Additive Explanations (SHAP) analysis enhanced interpretability.</p> Results <p>The best clinical model achieved an AUC of 0.666 in the validation set. The best-performing traditional radiomics model achieved an AUC of 0.825 in the internal validation set. The optimal deep learning model obtained an AUC of 0.804 in the internal validation set. The combined model, utilizing the best features from both traditional radiomics and deep learning, demonstrated superior performance with an AUC of 0.885 in the internal validation set and 0.839 in the external test set.</p> Conclusion <p>The integrated model combining traditional radiomics and deep learning from MRI significantly predicts Ki-67 expression in breast cancer, enhancing preoperative prediction accuracy and interpretability for personalized treatment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multiparametric MRI-based habitat analysis integrating deep learning and radiomics for predicting preoperative Ki-67 expression level in breast cancer

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月16日

查看全文
  • Yuqian Wang,
  • Yue Zhang,
  • Zaiyi Liu,
  • Yiming Xiong,
  • Mifang Li,
  • Lingyan Zhang,
  • Zhenwei Shi

摘要

Background

Breast cancer (BC) is the most common malignant tumor in women globally. Ki-67, a vital marker for prognosis, is currently detected invasively. Non-invasive magnetic resonance imaging (MRI) prediction faces challenges due to intratumoral heterogeneity.

Materials and methods

This retrospective study included 254 breast cancer patients from two centers, divided into training set (142 patients), internal validation set (60 patients), and external test set (52 patients). T2-weighted imaging (T2WI) and dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) were analyzed. Traditional radiomics features were extracted from intratumoral, habitat subregions, 5/10-mm peritumoral rings, image fusion. A pre-trained ResNet-50 model extracted 2.5D deep learning features. Feature selection used intraclass correlation coefficient (ICC), Z-score normalization, T-tests, Pearson correlations, and the least absolute shrinkage and selection operator (LASSO). A baseline clinical model was constructed using clinical and qualitative MRI semantic features. Models were built using Support Vector Machine (SVM), Random Forest (RF), and Extra-Trees (ET). Model performance was evaluated via the area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F-1 score. Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to the final convolutional layer of ResNet50 to spatially localize the decision-critical regions. Shapley Additive Explanations (SHAP) analysis enhanced interpretability.

Results

The best clinical model achieved an AUC of 0.666 in the validation set. The best-performing traditional radiomics model achieved an AUC of 0.825 in the internal validation set. The optimal deep learning model obtained an AUC of 0.804 in the internal validation set. The combined model, utilizing the best features from both traditional radiomics and deep learning, demonstrated superior performance with an AUC of 0.885 in the internal validation set and 0.839 in the external test set.

Conclusion

The integrated model combining traditional radiomics and deep learning from MRI significantly predicts Ki-67 expression in breast cancer, enhancing preoperative prediction accuracy and interpretability for personalized treatment.