INTRODUCTION: Chondrosarcoma is the second most common malignant bone tumor in adults, with prognosis driven by histological grade and tumor location. Differentiating enchondroma, atypical cartilaginous tumor (ACT), and high-grade disease remains difficult with imaging and histopathology. Artificial intelligence (AI) offers a non-invasive approach for grading and risk stratification by analyzing radiological and clinical features. MATERIAL AND METHODS: This systematic review adhered to Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines and focused on AI applications for non-invasive grading and prognostic modelling in chondrosarcoma. A comprehensive search of five major databases identified eligible original studies, and data were extracted on populations, input modalities, AI techniques, validation methods, and model performance. RESULTS: Across non-invasive grading and aggressiveness tasks, AI models achieved area under the curve (AUC) values 0.64–0.97 across computed tomography (CT), magnetic resonance imaging (MRI), and X-ray, with the weakest performance at the ACT high-grade boundary and the strongest in MRI-based multiclass grading. Prognostic models for survival and recurrence reported C-indices 0.76–0.84 and time- or endpoint-specific AUCs 0.84–0.99, with highest values observed in distant metastasis risk models. Magnetic resonance imaging radiomics performed best for complex grading, whereas CT models improved when radiomics was combined with clinical variables or deep learning (DL) features. Calibration and decision-curve analyses were infrequently reported. CONCLUSIONS: AI shows promise in supporting non-invasive grading and risk prediction in chondrosarcoma. However, the current evidence base is constrained by predominantly retrospective study designs, heterogeneous methodologies, limited external validation, and a generally high risk of bias. Multicenter standardization, multimodal data integration, transparent reporting, and prospective validation are essential before AI tools can reliably be implemented in clinical decision making.
Żerdziński et al. (Sun,) studied this question.