Radiomics and Machine Learning With Multiparametric Preoperative MRI May Accurately Predict the Histopathological Grades of Soft Tissue Sarcomas

Abstract

Background: Preoperative prediction of the grade of soft tissue sarcomas (STSs) is important because of its effect on treatment planning.

Purpose: To assess the value of radiomics features in distinguishing histological grades of STSs.

Study type: Retrospective.

Population: In 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/sequence: The 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.

Assessment: Multiple 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 tests: 1) 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.

Results: The 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 conclusion: Radiomics feature-based machine-learning methods are useful for distinguishing STS grades.

Level of evidence: 3 Technical Efficacy: Stage 2 J. Magn. Reson. Imaging 2020;51:791-797.

Keywords: diagnosis; machine learning; magnetic resonance imaging; soft tissue sarcoma.