Enhanced CT and MRI Focal Bone Tumor Classification with Machine Learning-based Stratification: A Multicenter Retrospective Study.

Abstract

Background Standardized bone tumor reporting is crucial for consistent, risk-aligned patient management. Current systems are based on expert consensus and/or lack multicenter validation. Purpose To evaluate a machine learning-based approach for differentiating between benign and malignant focal bone lesions and to propose a Bone Tumor Imaging Reporting and Data System (BTI-RADS) 2.0 for further risk stratification. Materials and Methods This retrospective multicenter trial included patients with solitary bone tumors undergoing <i>(a)</i> radiography or CT and <i>(b)</i> MRI at 10 centers from November 2009 to March 2022. Patients were divided into training and test datasets. Predefined radioclinical features were extracted. The training dataset was considered for bootstrapped χ² feature selection, and extreme gradient boosting (XGBoost) classifiers were optimized using nested cross-validation. Continuous classifier outputs were thresholded to stratify patients into seven malignancy risk classes (BTI-RADS 2.0), and malignancy rates were assessed for the test set. XGBoost and human expert performances were compared using the Wilcoxon signed-rank significance test with a significance level of .05. Results In total, 1113 patients (mean age, 39 years ± 22 [SD]; 623 men) were included: 298 in the training and 815 in the test datasets. Twenty-seven of 80 (34%) multimodal features were selected based on χ² analysis. Best classification performances were achieved by an XGBoost model trained on 27 features, with an F1 score of 0.81 (95% CI: 0.78, 0.84). This model performed slightly inferior to 28 experienced radiologists, who demonstrated an F1 score of 0.83 (95% CI: 0.80, 0.85; <i>P</i> < .001). BTI-RADS 2.0 risk grades II-V were associated with malignancy rates of 0% (0 of 102; 95% CI: 0, 0), 8.3% (14 of 168; 95% CI: 4, 13), 45% (121 of 271; 95% CI: 39, 50), and 92% (252 of 274; 95% CI: 89, 95), respectively, identifying malignant lesions with a sensitivity of 96% (373 of 387; 95% CI: 94, 98). Conclusion A machine learning algorithm and risk stratification system achieved accurate and standardized bone tumor malignancy grading. Clinical trial registration no. NCT04884048 © RSNA, 2025 <i>Supplemental material is available for this article.</i> See also the editorial by Tordjman and Murphey in this issue.