人工智能与机器学习驱动的脑动静脉畸形分割与量化模型:一项系统综述
本系统综述纳入13项研究、共3,010例个体,评估人工智能与机器学习用于脑动静脉畸形自动分割、量化及治疗规划的表现。常用模态为TOF-MRA和MRI,模型包括U-Net、Dense U-Net、YOLO、SVM及模糊C均值聚类,平均Dice相似度评分为0.758。作者认为这些模型在出血风险评估、皮质脊髓束受累、畸形团弥散度预测、畸形团分割及立体定向放射外科规划中显示出潜在价值,但模型泛化性与临床实施仍有限。
要点
- 13项研究共3,010例个体,常用模态为TOF-MRA和MRI,平均Dice相似度评分为0.758。
- 模型包括U-Net、Dense U-Net、YOLO、SVM及模糊C均值聚类,用于畸形团分割与量化。
- 模型在出血风险评估、皮质脊髓束受累、畸形团弥散度预测及立体定向放射外科规划中显示有用性。
核心结论
- 人工智能与机器学习在脑动静脉畸形评估中显示出潜力,早期研究提示可能支持诊断与治疗规划的效率和标准化。
- 尽管结果令人鼓舞,模型的泛化性与临床实施仍然有限。
- 未来研究应关注前瞻性验证、多模态影像整合及治疗后分割,以促进临床转化。
研究设计与样本
- 遵循PRISMA的系统综述,检索PubMed、ScienceDirect、Scopus和Web of Science,使用特定关键词与布尔运算符。
- 纳入标准为使用人工智能或机器学习模型对动静脉畸形进行基于影像分析的研究。
- 共纳入13项研究,合计3,010例个体;最常用的模态为TOF-MRA和MRI。
诊断标准与关键数值
- 全部实验中平均Dice相似度评分为0.758。
- 在放射规划与出血风险相关任务中,多个模型提供了更好或相当的预测准确度,并与手动分割显示良好一致性。
临床应用与阅片要点
- 系统综述纳入的研究显示,模型可用于出血风险评估、皮质脊髓束受累评估、动静脉畸形弥散度预测、畸形团分割及立体定向放射外科规划。
- 在放射规划与出血风险任务中,多个模型与手动分割显示良好一致性,预测准确度更好或相当。
局限性
- 模型的泛化性与临床实施仍然有限。
- 未来研究应关注前瞻性验证、多模态影像整合及治疗后分割,以促进临床转化。
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)。
英文原文(摘要)· 点击展开
PURPOSE: Our systematic review aims to evaluate the application of artificial intelligence (AI) and machine learning (ML) techniques for the automatic segmentation, quantification, and treatment planning of brain arteriovenous malformations (AVMs).
METHODS: A Preferred Reporting Items for systematic reviews and Meta-Analysis (PRISMA) guided systematic review was conducted using specific keywords and Boolean operators across PubMed, ScienceDirect, Scopus, and Web of Science. Studies were included based on the use of AI or machine learning (ML) models for imaging-based analysis of arteriovenous malformations (AVMs).
RESULTS: There were thirteen studies with 3,010 individuals. The most popular modalities were TOF-MRA and MRI. U-Net, Dense U-Net, YOLO, SVM, and fuzzy c-means clustering were among the models. Across all experiments, the average Dice similarity score was 0.758. The models showed usefulness in bleeding risk assessment, corticospinal tract involvement, AVM diffuseness prediction, nidus segmentation, and stereotactic radiosurgery (SRS) planning. In tasks involving radiation planning and hemorrhagic risk, a number of models provided better or comparable predicted accuracy and showed good agreement with manual segmentations.
CONCLUSION: AI and ML show potential for AVM evaluation, with early studies suggesting they may support efficiency and standardization in diagnosis and treatment planning. Despite encouraging findings, model generalizability and clinical implementation remain limited. Future studies should focus on prospective validation, integration of multimodal imaging, and post-treatment segmentation to enhance clinical translation.
引用
- Yurtluk MD, Balasubramanian K, Blackwell MP, et al. Artificial intelligence and machine learning driven segmentation and quantification models for brain arteriovenous malformations: A systematic review. Neuroradiology. 2025. PMID: 41240100. doi:10.1007/s00234-025-03839-z
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。