Role of Radiomics Based on Magnetic Resonance Imaging in Patients with Chondroid Bone Tumors: A Systematic Review and Meta-analysis

Purpose or Learning Objective: The reliable differentiation of cartilaginous tumors is challenging. Distinguishing between benign chondrogenic tumors and low-grade chondrosarcoma requires careful clinical, radiologic, and pathologic evaluation. Conventional magnetic resonance imaging (MRI) techniques rely on the use of T1-weighted and fat-suppressed T2-weighted images, but diagnostic accuracy depends on the expertise of the reader. This study investigated the usefulness of MRI-based radiomics in differentiating enchondromas from chondrosarcomas.

Methods or Background: This systematic review and meta-analysis included all studies between 2018 and 2022 that evaluated the diagnostic performance of MRI-based radiomics in distinguishing enchondromas versus chondrosarcomas. Two independent readers conducted the selection and data extraction, and a third reviewer resolved the disagreements.

These data were extracted from the eligible articles: title, first author, year of publication, countries, study design, end points, patient demographics, and main study findings. Similarly, data regarding the radiomics analysis, such as software, preprocessing steps, feature extraction properties, feature selection steps, prediction model algorithms, and the use of independent internal/external test data sets, were extracted. Finally, the radiomics quality score (RQS) of the studies was also evaluated.

Results or Findings: A total of nine articles were included in the final analysis. The average RQS achieved was 9.89 points, with minimum and maximum scores ranging from 5 to 15 of 36 possible points. The T1-weighted-based radiomics models slightly overperformed the T2-weighted-based ones in differentiating enchondromas from chondrosarcomas but without statistical significance. The T1-weighted and T2-weighted random effects models had a sensitivity of 0.88 (95% confidence interval, 0.80–0.93] and 0.81 (95% CI, 0.70–0.88], respectively (P = 0.4319). Similarly, no significant difference was observed in specificities (P = 0.663). The T1-weighted versus T2-weighted random effects models had a specificity of 0.84 (0.67–0.93) and 0.77 (0.66–0.86), respectively. We also evaluated the area under the receiver operating characteristic (AUROC) of the prediction models on the test data sets. The random effect models showed an AUROC of 0.88 (0.77–0.99) and 0.90 (0.81–1.00) for the T1-weighted and the T2-weighted models (P = 0.8433), respectively. Interestingly, the performance of expert radiologists in these articles showed an overall poor accuracy between 0.602 and 0.718, with a wide range of sensitivity of 0.333 to 0.604 but a balanced specificity of 0.80 to 0.84.

Conclusion: The results of the RQS assessment show that the overall quality of previously published articles is poor. The meta-analysis demonstrated that T1-weighted-based radiomics models may have better predictive performance compared with T2-weighted-based ones, and both may provide better diagnostic performance compared with expert radiologists. Due to the low number of articles, however, statistical significance cannot be proven.

Publication History

Article published online:
26 May 2023