MRI radiomics-based machine-learning classification of bone chondrosarcoma

• Published in: European journal of radiology 2020-07, Vol.128, p.109043-109043, Article 109043
• Main Authors: Gitto, Salvatore, Cuocolo, Renato, Albano, Domenico, Chianca, Vito, Messina, Carmelo, Gambino, Angelo, Ugga, Lorenzo, Cortese, Maria Cristina, Lazzara, Angelo, Ricci, Domenico, Spairani, Riccardo, Zanchetta, Edoardo, Luzzati, Alessandro, Brunetti, Arturo, Parafioriti, Antonina, Sconfienza, Luca Maria
• Format: Article
• Language: English
• Subjects: Adult; Artificial intelligence; Bone and Bones - diagnostic imaging; Bone and Bones - pathology; Bone Neoplasms - diagnostic imaging; Bone Neoplasms - pathology; Cartilaginous tumor; Chondrosarcoma - diagnostic imaging; Chondrosarcoma - pathology; Female; Humans; Image Interpretation, Computer-Assisted - methods; Machine Learning; Magnetic Resonance Imaging - methods; Male; Middle Aged; Neoplasm Grading; Radiomics; Reproducibility of Results; Retrospective Studies; ROC Curve; Texture analysis
• ISSN: 0720-048X; 1872-7727
• DOI: 10.1016/j.ejrad.2020.109043

•MRI radiomics-based machine learning is promising for chondrosarcoma classification.•It yielded 85.7 % and 75 % accuracy in our training and test cohorts, respectively.•Its performance was similar compared to a musculoskeletal radiologist. To evaluate the diagnostic performance of machine learning for discrimination between low-grade and high-grade cartilaginous bone tumors based on radiomic parameters extracted from unenhanced magnetic resonance imaging (MRI). We retrospectively enrolled 58 patients with histologically-proven low-grade/atypical cartilaginous tumor of the appendicular skeleton (n = 26) or higher-grade chondrosarcoma (n = 32, including 16 appendicular and 16 axial lesions). They were randomly divided into training (n = 42) and test (n = 16) groups for model tuning and testing, respectively. All tumors were manually segmented on T1-weighted and T2-weighted images by drawing bidimensional regions of interest, which were used for first order and texture feature extraction. A Random Forest wrapper was employed for feature selection. The resulting dataset was used to train a locally weighted ensemble classifier (AdaboostM1). Its performance was assessed via 10-fold cross-validation on the training data and then on the previously unseen test set. Thereafter, an experienced musculoskeletal radiologist blinded to histological and radiomic data qualitatively evaluated the cartilaginous tumors in the test group. After feature selection, the dataset was reduced to 4 features extracted from T1-weighted images. AdaboostM1 correctly classified 85.7 % and 75 % of the lesions in the training and test groups, respectively. The corresponding areas under the receiver operating characteristic curve were 0.85 and 0.78. The radiologist correctly graded 81.3 % of the lesions. There was no significant difference in performance between the radiologist and machine learning classifier (P = 0.453). Our machine learning approach showed good diagnostic performance for classification of low-to-high grade cartilaginous bone tumors and could prove a valuable aid in preoperative tumor characterization.