Related Experiment Video
Updated: Jul 1, 2026
Tracking the Mammary Architectural Features and Detecting Breast Cancer with Magnetic Resonance Diffusion Tensor Imaging
Published on: December 15, 2014
Preoperative prediction of Ki-67 expression and prognosis in breast cancer based on contrast-enhanced cone-beam
Wenxiu Guo1, Binglin Lyu2, Yue Ma1
1Department of Radiology, Medical Artificial General Intelligence for Computation (MAGIC) Lab, Tianjin Medical University Cancer Institute & Hospital; National Clinical Research Centre for Cancer; Tianjin's Clinical Research Centre for Cancer; Key Laboratory of Cancer Prevention and Therapy, Tianjin; Key Laboratory of Breast Cancer Prevention and Therapy, Tianjin Medical University, Ministry of Education; State Key Laboratory of Druggability Evaluation and Systematic Translational Medicine, Huan-Hu-Xi Road, Ti-Yuan-Bei, Hexi District, Tianjin, 300060, China.
Background:
This study aimed to develop and validate preoperative contrast-enhanced cone-beam breast CT (CE-CBBCT)-based radiomics, deep learning, and combined models for predicting Ki-67 expression status and to preliminarily explore their prognostic value, with the goal of informing preoperative treatment decisions.
Methods:
This two-centre study enrolled 636 breast cancer patients, divided into training (n = 310), internal validation (n = 77), external test (n = 78), and prognostic (n = 171) cohorts. Two Ki-67 cut-offs (14% and 30%) were applied. Radiomics models were constructed using five machine learning classifiers, and deep learning models using four three-dimensional architectures (ResNet50, ResNet101, DenseNet121, and ShuffleNet). The best-performing models from each category were integrated with the clinical model using logistic regression to establish combined models. Prognostic value was evaluated using disease-free survival (DFS) in the prognostic cohort.
Results:
Among the radiomics models, those based on ExtraTrees and eXtreme Gradient Boosting performed best for the 14% and 30% cut-offs, respectively, with areas under the receiver operating characteristic curves (AUCs) of 0.885 and 0.831 in the external test cohort. Among the deep learning architectures, ShuffleNet performed best, with corresponding AUCs of 0.630 and 0.825. Combined models achieved the highest AUCs of 0.895 and 0.869 for the 14% and 30% cut-offs, respectively, in the external test cohort. However, they did not significantly outperform the corresponding radiomics models. In survival analysis, the radiomics score for predicting Ki-67 ≥ 30% was an independent prognostic factor for DFS (hazard ratio: 1.661; 95% confidence interval: 1.142-2.417; P = 0.008).
Conclusions:
CE-CBBCT-based radiomics models demonstrated utility as non-invasive tools for the preoperative prediction of high Ki-67 expression defined at both the 14% and 30% cut-offs. The prognostic value of the radiomics score for predicting Ki-67 ≥ 30% was preliminarily demonstrated.
