• Hexiang Wang, Haisong Chen, Shaofeng Duan, Dapeng Hao, and Jihua Liu.
• Department of Radiology, Affiliated Hospital of Qingdao University, Qingdao, Shandong, China.
• J Magn Reson Imaging. 2020 Mar 1; 51 (3): 791-797.

BackgroundPreoperative prediction of the grade of soft tissue sarcomas (STSs) is important because of its effect on treatment planning.PurposeTo assess the value of radiomics features in distinguishing histological grades of STSs.Study TypeRetrospective.PopulationIn all, 113 patients with pathology-confirmed low-grade (grade I), intermediate-grade (grade II), or high-grade (grade III) soft tissue sarcoma were collected.Field Strength/SequenceThe 3.0T axial T1 -weighted imaging (T1 WI) with 550 msec repetition time (TR); 18 msec echo time (TE), 312 × 312 matrix, fat-suppressed fast spin-echo T2 WI with 4291 msec TR, 85 msec TE, 312 × 312 matrix.AssessmentMultiple machine-learning methods were trained to establish classification models for predicting STS grades. Eighty STS patients (18 low-grade [grade I]; 62 high-grade [grades II-III]) were enrolled in the primary set and we tested the model with a validation set with 33 patients (7 low-grade, 26 high-grade).Statistical Tests1) Student's t-tests were applied for continuous variables and the χ2 test were applied for categorical variables between low-grade STS and high-grade STS groups. 2) For feature subset selection, either no subset selection or recursive feature elimination was performed. This technology was combined with random forest and support vector machine-learning methods. Finally, to overcome the disparity in the frequencies of the STS grades, each machine-learning model was trained i) without subsampling, ii) with the synthetic minority oversampling technique, and iii) with random oversampling examples, for a total of 12 combinations of machine-learning algorithms that were assessed, trained, and tested in the validation cohort.ResultsThe best classification model for the prediction of STS grade was a combination of features selected by recursive feature elimination and random forest classification algorithms with a synthetic minority oversampling technique, which had an area under the curve of 0.9615 (95% confidence interval 0.8944-1.0) in the validation set.Data ConclusionRadiomics feature-based machine-learning methods are useful for distinguishing STS grades.Level Of Evidence3 Technical Efficacy: Stage 2 J. Magn. Reson. Imaging 2020;51:791-797.© 2019 International Society for Magnetic Resonance in Medicine.

12 keywords...

hide…