The use of radiomics and machine learning for the differentiation of chondrosarcoma from enchondroma

PURPOSE: To construct and compare machine learning models for differentiating chondrosarcoma from enchondroma using radiomic features from T1 and fat suppressed Proton density (PD) magnetic resonance imaging (MRI). METHODS: Eighty-eight patients (57 with enchondroma, 31 with chondrosarcoma) were retrospectively included. Histogram matching and N4ITK MRI bias correction filters were applied. An experienced musculoskeletal radiologist and a senior resident in radiology performed manual segmentation. Voxel sizes were resampled. Laplacian of Gaussian filter and wavelet-based features were used. One thousand eight hundred eighty-eight features were obtained for each patient, with 944 from T1 and 944 from PD images. Sixty-four unstable features were removed. Seven machine learning models were used for classification. RESULTS: Classification with all features showed neural network was the best model for both readers' datasets with area under the curve (AUC), classification accuracy (CA), and F1 score of 0.979, 0.984; 0.920, 0.932; and 0.889, 0.903, respectively. Four features, including one common to both readers, were selected using fast correlation based filter. The best performing models with selected features were gradient boosting for Fatih Erdem's dataset and neural network for Gülen Demirpolat's dataset with AUC, CA, and F1 score of 0.990, 0.979; 0.943, 0.955; 0.921, 0.933, respectively. Neural Network was the second-best model for FE's dataset based on AUC (0.984). CONCLUSION: Using pathology as a gold standard, this study defined and compared seven well-performing models to distinguish enchondromas from chondrosarcomas and provided radiomic feature stability and reproducibility among the readers.