Radiomics based on machine learning in predicting the long-term prognosis for triple-negative breast cancer after neoadjuvant chemotherapy / 中华放射学杂志

ABSTRACT

Objective:To explore the value of different radiomics models based on machine learning in predicting the risk of distant recurrence and metastasis of triple-negative breast cancer after neoadjuvant therapy.Methods:The clinical and imaging data of 150 patients with triple-negative breast cancer (TNBC) confirmed by histopathology were retrospectively analyzed. All patients underwent neoadjuvant chemotherapy and surgical resection from August 2011 to May 2017 in Fudan University Shanghai Cancer Center and Ruijin Hospital, Shanghai Jiao Tong University School of Medicine. One hundred and nine patients from Shanghai Fudan University Shanghai Cancer Center were used as the training group, and 41 patients from Ruijin Hospital, Shanghai Jiao Tong University School of Medicine were used as the validation group. The features were extracted from dynamic contrast-enhanced MRI (DCE-MRI) before treatment and were added with time domain features innovatively. Least absolute shrinkage and selection operator cross validation and recursive feature elimination were applied to select features. Six different supervised machine learning algorithms (logistic regression, linear discriminant analysis, k-nearest neighbor, naive bayesian, decision tree, support vector machine) were used to predict the prognosis. ROC curve, accuracy and F1 measure were used to evaluate the performance of the six algorithms, and also verified by the validation group.Results:The support vector machine algorithm had the best predictive effect in the recurrence and metastasis model based on 15 features, with the highest area under curve (training group was 0.917, validation group was 0.859), and the highest accuracy rate (training group was 87.5%, validation group was 82.9%) and the highest F1 measure (training group was 0.800, validation group was 0.741). In addition, of the 15 imaging features, 12 were the time domain features and 3 were spatial features.Conclusion:With the help of the time domain features and machine learning algorithms, radiomics signatures based on preoperative DCE-MRI can help predict the distant prognosis for TNBC after neoadjuvant chemotherapy and provide support for clinical decision making and follow-up management.