Can Radiomics and Machine Learning Predict Diagnosis in Musculoskeletal Myxoid Soft Tissue Tumors?

Abstract

Introduction: Differentiating benign and malignant musculoskeletal myxoid soft tissue tumors is challenging due to their shared clinical, imaging, and histologic features. We assess if radiomics and machine learning can assist in differentiating benign and malignant musculoskeletal myxoid tumors. Methods: Forty patients with a pre-treatment MRI and a histologically confirmed myxoid soft tissue tumor (20 myxomas and 20 myxofibrosarcomas) on final resection pathology were reviewed. Baseline clinical features consisted of patient age, sex, tumor size, tumor depth, tumor location, pain, and tumor as an incidental finding. Manual image segmentation of the tumor was performed on T1 and T2 images by attending clinicians. A LASSO model was used for MRI feature reduction. The performance of five machine learning models were implemented using cross validation and compared with classify benign and malignant myxoid soft tissue tumors. Results: The five models using T1 and T2 radiomics features + clinical features achieved mean area under the ROC curves (AUC) of 0.729 to 0.780 and accuracy of 0.696 to 0.757. The highest AUC achieved was 0.780 using a support vector machine model. All combined radiomics + clinical feature models outperformed the classification models using clinical features alone. Conclusion: Classification models using T1 and T2 radiomics features + clinical features perform better in differentiating benign and malignant myxoid soft tissue tumors than models using clinical features alone. Future studies include validation of these preliminary findings in a larger and separate data set, and comparison of the diagnostic performance of radiomics to manual review of images by clinicians.