CT radiomics-based machine learning model for differentiating between enchondroma and low-grade chondrosarcoma

Abstract

It may be difficult to distinguish between enchondroma and low-grade malignant cartilage tumors (grade 1) radiologically. This study aimed to construct machine learning models using 3D computed tomography (CT)-based radiomics analysis to differentiate low-grade chondrosarcoma from enchondroma. A total of 30 patients with enchondroma and 26 with chondrosarcoma were included in this retrospective study. Tumor volume segmentation was manually performed by 2 musculoskeletal radiologists. In total, 107 radiomic features were obtained for each patient. The intraclass correlation coefficient was used to assess interobserver reliability and estimate the absolute agreement between the 2 radiologists. Algorithm-based information gain was used as a feature reduction method, and the 5 most important features were detected. For classification, 7 machine learning models were utilized. Classification was carried out using either all features or 5 features. There was good to excellent agreement between the 2 radiologists for the 107 features of each patient. Therefore, a dataset containing 107 features was used for machine learning classification. When assessed based on area under curve (AUC) values, classification using all features revealed that naive Bayes was the best model (AUC = 0.950), while classification using 5 features revealed that random forest was the best model for differentiating chondrosarcoma from enchondroma (AUC = 0.967). In conclusion, machine learning models using CT-based radiomics analysis can be used to differentiate between low-grade chondrosarcoma and enchondroma.