Radiograph-based Deep Learning Algorithm for Assisting Bone Tumor Risk Assessment: A Multicenter Study

Background Accurate radiograph-based risk stratification is critical for bone tumor management. However, the multicenter performance of the Bone Reporting and Data System (Bone-RADS) and the added value of deep learning-based decision support remain unclear. Purpose To develop and evaluate a radiograph-based deep learning model for binary risk stratification-benign versus potentially malignant (ie, intermediate or malignant)-and to compare it with radiologist and Bone-RADS assessments. Materials and Methods This multicenter study retrospectively obtained radiographs of histopathologically confirmed bone tumors from 10 institutions (January 2010 to December 2024) and prospectively recruited participants from two of these centers for testing (May-August 2025). The developed image-clinical/semantic-RADS attention combined model (DL-Clinic-RADS) used a two-view cross-attention vision transformer integrating clinical and semantic features with expert consensus. The model performance was compared with that of nine musculoskeletal radiologists, Bone-RADS, and artificial intelligence (AI)-aided classification. The primary endpoint was the area under the receiver operating characteristic curve (AUC), with secondary endpoints including precision-recall curves, calibration, decision curve analysis, interreader agreement, reading time and confidence, and simulated clinical decision-making. Results The retrospective dataset comprised 1777 patients (1009 male; mean age, 36.93 years ± 21.32 [SD]), and the prospective testing dataset comprised 152 participants (79 male; mean age, 36.22 years ± 21.27). The DL-Clinic-RADS achieved an AUC of 0.96 (95% CI: 0.94, 0.98) in the external set and 0.99 (95% CI: 0.97, 1.00) in the prospective set, outperforming the other five models (P < .001), radiologist assessments (external set AUC, 0.68-0.82; P < .001), and the Bone-RADS assessment (external set AUC, 0.75-0.81; P < .001). With AI assistance, the mean reader AUC increased by 0.07 (95% CI: 0.03, 0.11), the interreader agreement improved (Fleiss κ, +0.06 to +0.23), and the median reading time decreased by 3 seconds overall. Conclusion Compared with the Bone-RADS and radiologist assessments, the DL-Clinic-RADS significantly improved the radiographic risk stratification of bone tumors, enhancing clinical decision-making and workflow efficiency. © RSNA, 2026 Clinical trial registration no. ChiCTR2500102778 Supplemental material is available for this article. See also the editorial by Tordjman and Taouli in this issue.