Ilkyu Han

Artificial intelligence (AI) has shown promise in detecting and characterizing musculoskeletal diseases from radiographs. However, most existing models remain task-specific, annotation-dependent, and limited in their adaptability across diseases and anatomical regions. Although a comprehensive foundation model trained on large-scale musculoskeletal radiographs is clinically needed, publicly available datasets remain limited in size and lack sufficient diversity to enable training across a wide range of musculoskeletal conditions and anatomical sites. Here, we present SKELEX, a large-scale foundation model for musculoskeletal radiographs, trained using self-supervised learning on 1.2 million diverse, condition-rich images. The model was evaluated on 12 downstream diagnostic tasks and generally outperformed baselines in fracture detection, osteoarthritis grading, and bone tumor classification. Furthermore, SKELEX demonstrated unsupervised reconstruction-based anomaly localization, producing error maps that identified pathologic regions without task-specific training. Building on this capability, we developed an interpretable, region-guided bone tumor classifier. This model maintained robust performance on independent external datasets and was deployed as a publicly accessible web application, serving as a proof of concept for its potential clinical translation. Overall, SKELEX provides a scalable, label-efficient, and broadly applicable AI framework for musculoskeletal radiographs, with its robust external validity specifically demonstrated in bone tumor applications.

Background: Patients undergoing surgery for bone metastases typically have advanced disease, and postoperative survival varies substantially. Accurate survival estimation is important for surgical decision-making and patient counseling. Several prognostic models have been externally validated in East Asian populations, but these tools were originally developed in Western cohorts and do not incorporate region-specific epidemiology or treatment patterns. Questions/purposes: (1) To develop, internally evaluate, and select a machine learning–based survival prediction model for patients undergoing surgery for nonspinal bone metastases using a multinational East Asian cohort. (2) To compare the performance of the selected model with that of an established Western prognostic tool developed by the Skeletal Oncology Research Group (SORG). (3) To identify which clinical features carried the greatest importance in the new model that we developed. Methods: All patients who underwent surgery for nonspinal bone metastases at three tertiary referral centers in the Republic of Korea, Taiwan, and Japan between January 2009 and December 2022 were included. In total, 1045 patients met the inclusion criteria. The median (range) age at surgery was 64 years (19 to 96), 46% (478 of 1045) of patients were female, and the femur was the most common metastatic site (66% [690]). Data for 3-month, 6-month, 1-year, 3-year, and 5-year overall survival were available for 82% (854), 68% (709), 51% (529), 23% (243), and 15% (160) of patients, respectively. The corresponding survival proportions were 84%, 71%, 56%, 36%, and 31%. Data on routinely available clinical, functional, and laboratory variables were collected, and candidate predictors were predefined based on clinical relevance and data availability across institutions. Missing data were < 4% for all variables in each institution and were handled by multivariate imputation by chained equations. We trained four models using different machine-learning algorithms, and the performance of each model was evaluated using leave-one-site-out validation, in which models were trained on data from two institutions and tested on the remaining institution to ensure separation between training and testing data sets. Model performance was assessed using the Concordance Index (C-index; the ability of the model to correctly rank patients according to their expected survival), Brier score (overall prediction error), time-dependent area under the curve (tdAUC; how well the model distinguishes patients with different survival outcomes at specific time points), calibration slope and intercept (agreement between predicted and observed survival), and decision curve analysis (the potential clinical benefit of using the model to guide treatment decisions). The best-performing model was designated as the East Asian Survival Tool for Bone Metastasis Surgery (EAST-BMS) and was compared with the SORG model. To allow a fair comparison, the performance of the SORG model was evaluated on the same held-out test data sets in each iteration of the leave-one-site-out validation, applying the same performance metrics used to select the final model. Results: Gradient boosting survival analysis demonstrated the most favorable overall performance and was selected as the EAST-BMS. The number of outcome events used for model evaluation was 170 at 3 months and 447 at 12 months. The EAST-BMS achieved tdAUC values of 0.81 (95% confidence interval [CI] 0.78 to 0.85) at 3 months and 0.78 (95% CI 0.70 to 0.84) at 12 months, compared with 0.81 (95% CI 0.74 to 0.86) and 0.76 (95% CI 0.67 to 0.83), respectively, for the SORG model, indicating comparable ability to distinguish patients with different survival outcomes. Brier scores were 0.12 (95% CI 0.09 to 0.15) and 0.23 (95% CI 0.17 to 0.28) for EAST-BMS versus 0.14 (95% CI 0.12 to 0.16) and 0.25 (95% CI 0.15 to 0.34) for SORG, indicating lower prediction error in EAST-BMS. Calibration intercepts were -0.08 (95% CI -0.25 to 0.09) versus -1.06 (95% CI -1.26 to -0.86) at 3 months and -0.35 (95% CI -0.49 to -0.22) versus -1.23 (95% CI -1.37 to -1.08) at 12 months, indicating better agreement between predicted and observed survival in EAST-BMS. Decision curve analysis showed wider threshold probability ranges with positive net clinical benefit for EAST-BMS (0.04 to 0.96 versus 0.05 to 0.66 at 3 months; 0.17 to 0.77 versus 0.08 to 0.67 at 12 months), which means that using the EAST-BMS to guide treatment decisions may provide greater clinical benefit than the SORG model. Albumin, Karnofsky performance status, percentage of lymphocytes, and C-reactive protein level were among the most influential predictors. Conclusion: The EAST-BMS, the first multinational machine-learning survival model for patients from East Asia undergoing surgery for nonspinal bone metastases of which we are aware, demonstrated favorable predictive accuracy and clinical utility. This web-based tool may support personalized prognostic assessment and surgical decision-making. It is freely available as a web-based tool at https://bms.east-mskonco.org. Level of Evidence: Level III, therapeutic study.

Artificial intelligence (AI) has shown promise in detecting and characterizing musculoskeletal diseases from radiographs. However, most existing models remain task-specific, annotation-dependent, and limited in generalizability across diseases and anatomical regions. Although a generalizable foundation model trained on large-scale musculoskeletal radiographs is clinically needed, publicly available datasets remain limited in size and lack sufficient diversity to enable training across a wide range of musculoskeletal conditions and anatomical sites. Here, we present SKELEX, a large-scale foundation model for musculoskeletal radiographs, trained using self-supervised learning on 1.2 million diverse, condition-rich images. The model was evaluated on 12 downstream diagnostic tasks and generally outperformed baselines in fracture detection, osteoarthritis grading, and bone tumor classification. Furthermore, SKELEX demonstrated zero-shot abnormality localization, producing error maps that identified pathologic regions without task-specific training. Building on this capability, we developed an interpretable, region-guided model for predicting bone tumors, which maintained robust performance on independent external datasets and was deployed as a publicly accessible web application. Overall, SKELEX provides a scalable, label-efficient, and generalizable AI framework for musculoskeletal imaging, establishing a foundation for both clinical translation and data-efficient research in musculoskeletal radiology.

Background: Although radiographs are the first-line imaging modality, differentiating between neoplastic pathologic fractures and nonpathologic fractures on radiographs can sometimes be challenging. In this study, we aimed to develop and evaluate a deep learning model capable of distinguishing neoplastic pathologic fractures from nonpathologic fractures on hip radiographs to enhance diagnostic accuracy. Methods: This retrospective, multicenter study analyzed anteroposterior hip radiographs from patients who visited the emergency department at 4 different institutions. The deep learning model was trained on, and tested using, data from 338 patients at a single institution and externally validated on data from 488 patients across 3 additional institutions. Results: The model achieved an overall accuracy of 0.880, with a sensitivity of 0.882 and a specificity of 0.879, on the internal test set. It was then externally validated using the data of 488 patients (67 with neoplastic pathologic fracture and 421 with nonpathologic fracture) from institutions separate from where the model was developed. The model achieved an overall accuracy of 0.848, sensitivity of 0.910, and specificity of 0.786. Its performance was comparable with that of general orthopaedic surgeons. Conclusions: The developed deep learning model is a reliable and valid tool for distinguishing neoplastic pathologic fractures from nonpathologic fractures on hip radiographs. It has the potential to assist orthopaedic surgeons in resource-limited settings, where optimizing the interpretation of radiographs is critical for patient care. The model is publicly available at https://pathfxdx.org. Level of Evidence: Diagnostic Level III . See Instructions for Authors for a complete description of levels of evidence.

Clinical Outcomes and Histologic Subtype-Specific Patterns in Oligometastatic Soft Tissue Sarcoma of the Extremities and Trunk Wall: Implications for Curative Local Ablative Therapy

Kim S, et al. Clin Orthop Surg. 2025 May;17:e35. https://doi.org/10.4055/cios24336

Sarcomas, malignant tumors from mesenchymal tissues, exhibit poor prognosis despite advancements in treatment modalities such as surgery, radiotherapy, and chemotherapy, with doxorubicin being a cornerstone treatment. Resistance to doxorubicin remains a significant hurdle in therapy optimization. This study aims to dissect the molecular bases of doxorubicin resistance in sarcoma cell lines, which could guide the development of tailored therapeutic strategies. Eighteen sarcoma cell lines from 14 patients were established under ethical approvals and classified into seven subtypes. Molecular, genomic, and transcriptomic analyses included whole-exome sequencing, RNA sequencing, drug sensitivity assays, and pathway enrichment studies to elucidate the resistance mechanisms. Variability in doxorubicin sensitivity was linked to specific genetic alterations, including mutations in TP53 and variations in the copy number of genomic loci like 11q24.2. Transcriptomic profiling divided cell lines into clusters by karyotype complexity, influencing drug responses. Additionally, pathway analyses highlighted the role of signaling pathways like WNT/BETA-CATENIN and HEDGEHOG in doxorubicin-resistant lines. Comprehensive molecular profiling of sarcoma cell lines has revealed complex interplays of genetic and transcriptomic factors dictating doxorubicin resistance, underscoring the need for personalized medicine approaches in sarcoma treatment. Further investigations into these resistance mechanisms could facilitate the development of more effective, customized therapy regimens.

Aims: This study aimed to compare the performance of survival prediction models for bone metastases of the extremities (BM-E) with pathological fractures in an Asian cohort, and investigate patient characteristics associated with survival.Methods: This retrospective cohort study included 469 patients, who underwent surgery for BM-E between January 2009 and March 2022 at a tertiary hospital in South Korea. Postoperative survival was calculated using the PATHFx3.0, SPRING13, OPTIModel, SORG, and IOR models. Model performance was assessed with area under the curve (AUC), calibration curve, Brier score, and decision curve analysis. Cox regression analyses were performed to evaluate the factors contributing to survival.Results: The SORG model demonstrated the highest discriminatory accuracy with AUC (0.80 (95% confidence interval (CI) 0.76 to 0.85)) at 12 months. In calibration analysis, the PATHfx3.0 and OPTIModel models underestimated survival, while the SPRING13 and IOR models overestimated survival. The SORG model exhibited excellent calibration with intercepts of 0.10 (95% CI -0.13 to 0.33) at 12 months. The SORG model also had lower Brier scores than the null score at three and 12 months, indicating good overall performance. Decision curve analysis showed that all five survival prediction models provided greater net benefit than the default strategy of operating on either all or no patients. Rapid growth cancer and low serum albumin levels were associated with three-, six-, and 12-month survival.Conclusion: State-of-art survival prediction models for BM-E (PATHFx3.0, SPRING13, OPTIModel, SORG, and IOR models) are useful clinical tools for orthopaedic surgeons in the decision-making process for the treatment in Asian patients, with SORG models offering the best predictive performance. Rapid growth cancer and serum albumin level are independent, statistically significant factors contributing to survival following surgery of BM-E. Further refinement of survival prediction models will bring about informed and patient-specific treatment of BM-E.Cite this article: Bone Joint J 2024;106-B(2):203–211.

Background: This study aimed to investigate the characteristics and clinical outcomes in a series of patients with extremity soft tissue sarcoma (STS) who underwent amputation at a large East Asian referral center. Patients and methods: Of the 652 patients who underwent surgery for extremity STS, data of 37 consecutive patients who underwent amputation were reviewed retrospectively. The median follow-up period was 96.0 months (range, 15–216). The patients were classified in to three cohorts. The primary localized (PL) group included patients who underwent amputation as a primary surgical procedure with curative intent. The recurrent localized (RL) group included patients who underwent amputation as a revision procedure after failure of previous limb sparing surgeries. The metastatic group included patients who underwent amputation as a palliative procedure. Results: There were 22 cases of amputation in 596 STS patients and the amputation rate was 3.6% (22/596). Further, 1.8% (9/490) of patients with primary localized STS underwent amputation. Patients with localized STS who underwent amputation had a 5-year disease-specific survival (DSS) rate of 89.9% (95% Confidence Interval (CI), 87.1–92.7%), a local-recurrence-free survival (LRFS) of 84.1% (95% CI, 80.5–87.6%), and a metastasis-free survival (MFS) of 84.6%. (95% CI, 81.1–88.0%) Compared with previous studies, our results showed higher DSS and MFS rates with similar LRFS. Conclusions: The amputation rate of extremity STS in our institute in East Asia was similar but slightly lower than that reported in Western studies. The oncologic outcome of amputation reported in this study was higher than that indicated in Western studies and oncologic outcome of amputation was not statistically different from those of limb salvage surgery. However, considering the small cohort in single institute study, there is a possibility of selection bias and future multi-center study is necessary. From our results, amputation is still a feasible option for appropriately selected patients unsuitable for limb-conserving surgery.