Performance comparison and future perspectives of deep learning and classical machine learning in bone tumor applications: a systematic review (2019-2025).

backgroundThe diagnosis and prognostic assessment of bone tumors represent a complex and clinically significant challenge. In recent years, the rise of artificial intelligence (AI), particularly deep learning (DL) and classical machine learning (ML), has emerged as a promising tool in this field. This study systematically reviews the applications of AI in bone tumor diagnosis, prognosis, segmentation, and treatment response, with a focus on model performance, emerging trends, and current limitations.

methodsThis systematic review follows to the PRISMA guidelines and conducted a comprehensive search of four major databases (PubMed, Web of Science, Scopus, and Cochrane Library) to identify studies published between January 2019 and May 2025 on the application of AI in bone tumors. Relevant original articles were identified based on predefined inclusion and exclusion criteria, and research data such as basic information, algorithms, models, performance metrics, and clinical tasks, were systematically extracted and analyzed. And the performance of DL and ML methods in bone tumors was comparatively analyzed.

resultsThe review included 70 studies involving 53,149 cases, of which 45.83% were malignant bone tumors. DL was used in 77.63% of the studies and classical ML in 22.37%. Diagnostic tasks dominated the research focus (81.94%), followed by survival prediction (11.11%) and treatment response evaluation (6.94%). Performance metrics indicated that DL models exhibited higher weighted averages in accuracy (0.87), AUC (0.89), sensitivity (0.84), specificity (0.88), precision (0.81), and F-score (0.84), while classical ML models achieved the highest precision (0.90). Although DL demonstrated a performance advantage in image-based tasks, classical ML maintained greater stability in structured datasets. No significant performance differences were observed between large-sample and small-sample studies, reflecting the robustness of both model types. Additionally, a recent shift in research focus was observed, from diagnostic applications toward disease prediction.

conclusionArtificial intelligence has demonstrated strong performance and potential in bone tumor research. DL often demonstrates more balanced performance in image-based bone tumor tasks, while classical ML remains competitive and may hold advantages in structured, small-sample datasets, precision-prioritized settings. However, we did not observe statistically significant differences, so these findings should be interpreted as performance tendencies in specific contexts rather than universally validated superiority. Future research should focus on optimizing DL and classical ML models, developing fusion algorithms in bone tumors can improve the generalization performance, accuracy, and ability to adapt to complex data scenarios. At the same time, fostering interdisciplinary and multicenter collaborations between computer scientists and clinicians, improving data-sharing frameworks, and addressing ethical and privacy concerns will be essential to fully harness the significant potential of AI in bone tumor research and clinical applications.