Interpretable model based on MRI radiomics to predict the expression of Ki-67 in breast cancer

This study aimed to develop an interpretable machine learning model that accurately predicts Ki-67 expression in breast cancer (BC) patients using a combination of dynamic-contrast enhanced magnetic resonance imaging (DCE-MRI) radiomics and clinical-imaging features. A total of 195 BC patients, including 201 lesions, were enrolled retrospectively. These lesions were randomized into training and testing set (7:3). The correlation between clinical-imaging features and Ki-67 expression was analyzed via univariate analysis and binary logistic regression, leading to the development of a Clinical-imaging model. Radiomics features were extracted based on the early and delayed phases of DCE-MRI. These features were screened by Pearson correlation coefficient and recursive feature elimination (RFE). The logistic regression classifier was used to develop the Radiomics model. The clinical imaging and radiomics features were combined to form a Combined model. The Shapley Additive Explanation (SHAP) algorithm was employed to explain the optimal model, and the AUC was used to assess the model's performance. Ki-67 expression was markedly different from the internal enhancement pattern and necrosis among the imaging features. Compared to the Clinical-imaging model (AUC = 0.682), the AUCs of the Radiomics and the Combined models in the training set were 0.797 and 0.821, respectively. Clinical-imaging, Radiomics, and Combined models had AUCs of 0.666, 0.796, and 0.802 in the test set. Based on the IDI results, the combined model outperformed the Clinical-imaging and Radiomics models in the training set by 11.8% and 2.1%, respectively. They increased by 11% and 1.74% in the test set. SHAP analysis showed that ph2-original-shape-surface volume ratio was the most important feature of the model. Based on clinical-imaging features and DCE-MRI radiomics, the interpretable machine learning model can accurately predict the expression of Ki-67 in BC. Combining the SHAP algorithm with the model improves its interpretability, which may assist clinicians in formulating more accurate treatment strategies.