Scholarlink

Background <p>Radiomics-based artificial intelligence (AI) models have emerged as promising tools for brain tumor diagnosis, grading, and classification by extracting quantitative features from imaging data. Yet, clinical translation remains limited due to variability in model design, validation, and interpretability.</p> Objective <p>To systematically evaluate the diagnostic performance—defined as the ability of radiomics-based AI models to correctly classify or distinguish brain tumors, measured by sensitivity, specificity, accuracy, and area under the curve (AUC)—and to identify the impact of model type, imaging modality, and tumor subtype on predictive accuracy.</p> Methods <p>A systematic review and meta-analysis were conducted in accordance with PRISMA 2020 guidelines. PubMed, Scopus, and Web of Science were searched from January 2010 to June 2025. A total of 107 studies met inclusion criteria. Pooled estimates of AUC, sensitivity, and specificity were calculated. Subgroup analyses were performed by model architecture, imaging modality, and tumor type.</p> Results <p>The pooled AUC was 0.88 (95% CI: 0.86–0.90). Deep learning models achieved the highest diagnostic performance (AUC 0.91), followed by hybrid (0.89) and conventional machine learning models (0.86). Multimodal imaging outperformed MRI or CT alone. Glioblastoma and brain metastases demonstrated the highest accuracy (AUC &gt; 0.90). Only 31% of studies employed external validation, and methodological heterogeneity was frequent.</p> Conclusion <p>Radiomics-based AI models demonstrate strong diagnostic performance in the diagnosis, grading, and classification of brain tumors, particularly when using deep learning and multimodal imaging. However, limited validation and inconsistent reporting practices remain barriers. Standardized pipelines, explainable algorithms, and multicenter prospective studies are needed to advance clinical implementation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Radiomics-based artificial intelligence models in brain tumors: A systematic review and meta-analysis of diagnostic performance

  • 期刊论文

  • 发表日期：

    2026年4月27日

查看全文
  • Jheremy S. Reyes,
  • Timoteo Almeida,
  • Alexandros Bouras,
  • Joseph Mettenburg,
  • Ajay Niranjan,
  • L. Dade Lunsford,
  • Constantinos G. Hadjipanayis

摘要

Background

Radiomics-based artificial intelligence (AI) models have emerged as promising tools for brain tumor diagnosis, grading, and classification by extracting quantitative features from imaging data. Yet, clinical translation remains limited due to variability in model design, validation, and interpretability.

Objective

To systematically evaluate the diagnostic performance—defined as the ability of radiomics-based AI models to correctly classify or distinguish brain tumors, measured by sensitivity, specificity, accuracy, and area under the curve (AUC)—and to identify the impact of model type, imaging modality, and tumor subtype on predictive accuracy.

Methods

A systematic review and meta-analysis were conducted in accordance with PRISMA 2020 guidelines. PubMed, Scopus, and Web of Science were searched from January 2010 to June 2025. A total of 107 studies met inclusion criteria. Pooled estimates of AUC, sensitivity, and specificity were calculated. Subgroup analyses were performed by model architecture, imaging modality, and tumor type.

Results

The pooled AUC was 0.88 (95% CI: 0.86–0.90). Deep learning models achieved the highest diagnostic performance (AUC 0.91), followed by hybrid (0.89) and conventional machine learning models (0.86). Multimodal imaging outperformed MRI or CT alone. Glioblastoma and brain metastases demonstrated the highest accuracy (AUC > 0.90). Only 31% of studies employed external validation, and methodological heterogeneity was frequent.

Conclusion

Radiomics-based AI models demonstrate strong diagnostic performance in the diagnosis, grading, and classification of brain tumors, particularly when using deep learning and multimodal imaging. However, limited validation and inconsistent reporting practices remain barriers. Standardized pipelines, explainable algorithms, and multicenter prospective studies are needed to advance clinical implementation.