Objectives: The purpose of this study was to evaluate the diagnostic performance of three classification models that utilize logistic regression analysis (LRA), support vector machine (SVM), and convolutional neural network (CNN) in distinguishing breast tumor properties and HER-2 status using quantitative analysis of breast dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI).
Methods: A total of 354 patients who underwent breast DCE-MRI were included in this retrospective study. All enrolled patients were divided into benign and malignant groups. Patients with breast cancer were divided into positive and negative HER-2 groups. Each group was then divided into training and validation cohorts using different MRI scanners. Machine learning texture analysis was realized using LRA and SVM. Least absolute shrinkage and selection operator (LASSO) were used to filter texture features. Deep learning was implemented using CNN. The classification model performance was evaluated using receiver operating characteristic (ROC) analysis.
Results: When distinguishing breast tumor properties, three classifiers based on LASSO_LRA, LASSO_SVM, and CNN in the training cohorts achieved the areas under ROC curves (AUCs) of 0.937, 0.946, and 0.995 (accuracy = 88.75%, 89.37%, and 95.63%), respectively. Three classifiers in the corresponding validation cohorts achieved AUCs of 0.896, 0.914, and 0.965 (accuracy = 83.30%, 85.56%, and 91.11%), respectively. When identifying the HER-2 status, three classifiers in the training cohorts reached AUCs of 0.907, 0.914, and 0.954 (accuracy = 82.61%, 86.09%, and 92.17%), respectively. Three classifiers in the corresponding validation cohorts reached AUCs of 0.819, 0.869, and 0.923 (accuracy = 81.43%, 84.29%, and 87.14%), respectively.
Conclusion: The performance of the deep learning approach with CNN is superior to that of machine learning approaches employing LRA and SVM. These results suggest that the classifier based on CNN can be used as an efficient tool for preoperative evaluation of breast tumor properties and HER-2 status.