人工智能在脑肿瘤分割、治疗计划与结局预测中的性能:系统综述与荟萃分析
该荟萃分析汇总了AI模型在脑肿瘤放疗中的定量性能,覆盖分割、计划与结局预测任务。合并AUC为0.856,DSC为0.840,靶区覆盖率达0.976,但异质性和医生覆盖率高,需标准化评估与前瞻性验证。
要点
- AI在脑肿瘤放疗中判别与分割性能较强,计划任务优于结局预测。
- 靶区覆盖率高、剂量学替代指标良好,但医生覆盖率达0.258。
- 研究异质性高、报告指标不一,需标准化评估与前瞻性验证。
核心结论
- AI用于脑肿瘤放疗表现出较强的合并判别能力和稳健分割,剂量学替代指标与靶区覆盖令人鼓舞。
研究设计与样本
- 系统综述与荟萃分析,检索主要数据库,纳入报告脑肿瘤放疗AI模型定量性能的研究,采用随机效应模型逆方差加权合并。
主要发现
- 所有AI任务合并AUC为0.856,计划任务高于结局预测。
- DSC为0.840,准确率为0.842,计划0.852、结局预测0.824、转移瘤0.863、胶质瘤0.875。
- 敏感度为0.854,特异度为0.845,HD为8.51 mm。
- 剂量学适形度为0.900,靶区覆盖为0.976,医生覆盖率为0.258。
- 中位OS为19.13个月,转移瘤15.5个月,胶质母细胞瘤17.53个月。
临床怎么用
- AI辅助分割与计划可能简化放疗流程,减少手动勾画负担并维持较高空间精度与剂量学质量。
局限性
- 异质性高、报告与指标不一、医生覆盖率高,需标准化评估、前瞻性临床验证和一致报告。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
BACKGROUND: Artificial intelligence (AI) tools for neuro-oncology are rapidly entering clinical workflows for image segmentation, treatment planning, and outcome prediction, yet their real-world performance and clinical correlates remain uncertain. We conducted a meta-analysis to synthesize diagnostic/predictive accuracy, segmentation quality, dosimetric surrogates, clinician override, and patient outcomes across studies. METHODS: We searched major databases, for studies reporting quantitative performance for AI models applied to brain tumors radiotherapy. Outcomes were pooled using random-effects models with inverse-variance weighting. We conducted prespecified subgroup analyses. Heterogeneity was performed. RESULTS: Across all AI tasks, AUC was 0.856, higher for planning than outcome prediction. DSC was 0.840. Accuracy was 0.842, with planning 0.852, outcome prediction 0.824, metastases 0.863, and glioma 0.875. Sensitivity was 0.854 (planning 0.886; outcome 0.817; metastases 0.848; glioma 0.914). Specificity was 0.845 (planning 0.953; outcome 0.793; metastases 0.856). HD was 8.51 mm, 4.46 mm for metastases and 10.07 mm for glioma. Dosimetric conformity was 0.900 and 0.917 in metastases; target coverage was 0.976 (0.969 in metastases). Physician override rate was 0.258 (0.332 in metastases). OS was 19.13 months, 15.5 in metastases, and 17.53 in glioblastoma. CONCLUSIONS: AI for brain tumors radiotherapy demonstrates strong pooled discrimination and robust segmentation with encouraging dosimetric surrogates and high target coverage. AI-assisted segmentation and planning could streamline radiotherapy workflows, reducing manual contouring burden while maintaining high spatial accuracy and dosimetric quality. Nevertheless, high heterogeneity, variable reporting/metrics, and physician override rates show the need for standardized evaluation, prospective clinical validation, and consistent reporting to translate technical gains into practice.
引用
- Buccilli B, Lucke-Wold B. Performance of artificial intelligence for brain tumor segmentation, treatment planning, and outcome prediction: a comprehensive systematic review and meta-analysis. J Neurooncol. 2025. PMID: 41254426. doi:10.1007/s11060-025-05320-6
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。