MR T1WI based radiomics and machine learning model for predicting the histopathological grades of soft tissue sarcomas

Objective To explore the value of T1WI based optimal radiomics and machine learning model in predicting histological grades of soft tissue sarcoma. Methods The preoperative MR T1WI data of 113 patients with soft tissue sarcoma in Affiliated Hospital of Qingdao University from May 2009 to November 2018 was analyzed retrospectively. The patients were divided into training set (n=80) and validation set (n=33) using randomly stratified sampling mothed. According to the French Federation Nationale des Centres de Lutte Contre le Cancer (FNCLCC) system, the soft tissue sarcomas were divided into 3 pathological levels (grade Ⅰ-Ⅲ). Grade Ⅰ was defined as low grade, grade Ⅱ and Ⅲ were defined as high grade. In the training set, there were 18 cases with low-grade lesions, 62 cases with high-grade lesions. In the validation set, there were 7 cases with low-grade lesions and 26 cases with high-grade lesions. After a normalizationapproach applied on the image, the radiomics features were extracted in the regions of interest using A.K software. Based on different feature selection methods [with or without recursive feature elimination (RFE)], machine learning algorithm [random forest (RF) or support vector machine (SVM)] and sampling technology [without subsampling, with the synthetic minority oversampling technique (SMOTE) or with random oversampling examples], a total of 12 models were built and each machine-learning combination model was trained using leave-one-out cross validation. The receiver operating characteristic (ROC) curves were used to evaluate the efficacy of the model in predicting the pathological grade of soft tissue sarcoma. Results Among the 12 different machine learning models, the optimal classification model for the prediction of soft tissue sarcoma pathological grade was a combination of RF, RFE and SMOTE, with an area under the curve of 0.909 (95% confidence interval, 0.808-1.000) in the validation set, and the accuracy, sensitivity, and specificity were 84.85%, 86.21%, and 75.00%, respectively. Conclusion The radiomics based machine learning model can be used as an attractive application approach for predicting histological grades of soft tissue sarcoma. Key words: Magnetic resonance imaging; Soft tissue neoplasms; Sarcoma; Radiomics; Artificial intelligence