Edinburgh Napier University

  Introduction
Soft-tissue sarcomas (STSs) constitute a rare group of malignancies. Diagnostic biopsies require experts and are invasive with long diagnostic intervals. Artificial intelligence (AI)-based STS models may serve as accurate detection and categorization tools; however, their diagnostic performance remains questionable.
Method
The PubMed, Scopus, and Web of Science databases were searched for related studies published until January 10, 2024. Studies that developed or used AI-based models to diagnose STS were included.
Results
Nine studies were included in this meta-analysis. The common effects model yielded an accuracy of 0.8923 [0.8831; 0.9016], and the random effects model yielded an accuracy of 0.8524 [0.8132; 0.8916]. The Tau ^2 was 0.0094 [0.0055; 0.0202] and the I^2 statistic was 93.2% [91.1%, 94.7%], suggesting a high level of heterogeneity among the studies. The most accurate model was the decision tree (DT) model used in this study, with an accuracy of 0.9900 [0.9675–1.0125]. The least accurate model was the MLP model used in this study, which had an accuracy of 0.5720 [0.5107; 0.6333]. The pooled sensitivity of the AI model was 100%. Radiomics and simple vector-machine-based models are the most sensitive.
Conclusion
AI-based models show promising results for the diagnosis of STS. However, future studies should address the value of AI in the real-world setting. To provide precise comparisons across various models, it is essential to create a uniform training dataset to mitigate any sources of bias or heterogeneity. This information will assist scientists in identifying the optimal model for diagnosing STS and implementing this model in clinical practice.