文献速递神经MR

人工智能影像组学在术前MRI预测胶质瘤空间复发:一项系统综述

AI radiomics predicts spatial glioma recurrence on preoperative MRI: a systematic review.
PubMed 文献 · Eur J Radiol · 2025-09-11

本系统综述纳入8项研究、共1004例高级别胶质瘤患者,评估AI模型基于术前MRI预测肿瘤局部或远处复发的能力。模型架构与输入序列异质,预测复发的敏感度40%-97%、特异度29%-98%,比值比8.13-19.48;4项研究行外部验证,其中1项为6家机构多中心队列。证据主要来自小型、单中心、回顾性队列,临床采用前需更大前瞻性多中心研究。

要点

  • AI模型可基于术前MRI预测胶质瘤局部或远处复发,报告比值比8.13-19.48。
  • 预测复发敏感度40%-97%、特异度29%-98%,各研究间变异较大。
  • 外部验证仅4项研究,其中1项为6家机构多中心队列,显示较高泛化性。

核心结论

  • AI模型已显示能够从基线MRI数据预测胶质瘤患者的局部和远处肿瘤复发;多中心研究报告的高比值比令人鼓舞,但证据主要来自小型、单中心、回顾性队列,临床采用前需更大前瞻性多中心研究。

研究设计与样本

  • 系统综述,检索PubMed、EMBASE和Cochrane Library,纳入8项使用术前MRI评估AI模型预测胶质瘤空间复发的研究,共1004例高级别胶质瘤患者,采用PROBAST+AI工具评估研究质量与方法学严谨性。

诊断标准与关键数值

  • 预测肿瘤复发的敏感度范围为40%-97%,特异度范围为29%-98%,预测肿瘤复发区域的比值比范围为8.13-19.48;外部验证在4项研究中实施,其中1项使用6家不同机构的多中心队列。

临床应用与阅片要点

  • 作者认为多中心研究报告的高比值比令人鼓舞,但证据主要来自小型、单中心、回顾性队列,临床采用前需更大前瞻性多中心研究。

局限性

  • 证据主要来自小型、单中心、回顾性队列。
  • 临床采用前需更大前瞻性多中心研究。

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

BACKGROUND: Adult-type diffuse gliomas are highly infiltrative primary brain tumors in which, after a combination of surgery and chemoradiation therapy, tumor recurrence is inevitable. Artificial Intelligence (AI) models have been found capable to predict local and distant tumor recurrence at baseline, with the potential to guide surgical margins and enable focal dose escalation in radiotherapy. This study systematically reviews the literature on the performance of AI models in predicting local or distant tumor recurrence in glioma patients using preoperative MRI data.

METHODS: A systematic literature search was conducted across PubMed, EMBASE, and the Cochrane Library. Studies evaluating AI-based models for spatial recurrence prediction in gliomas using preoperative MRI were included. Study quality and methodological rigor were assessed using the PROBAST + AI tool.

FINDINGS: Eight studies, comprising 1004 high grade glioma patients, were included. A variety of machine learning and deep learning model architectures (e.g., Random Forest classifiers, Support Vector Machines and custom Convolutional Neural Networks) were employed. Input data were a heterogeneous combination of conventional MRI (e.g., T1CE, FLAIR) and more advanced imaging modalities (e.g., diffusion-weighted imaging). Considerable variability was reported with regard to sensitivity and specificity rates (ranging between 40 %-97 % and 29 %-98 %, respectively) for predicting tumor recurrence. The odds ratios for predicting regions of tumor recurrence, however, were generally high (ranging between 8.13-19.48). External validation was performed in 4 studies with one study using a multicenter cohort of 6 different institutions, demonstrating high generalizability in predictive performance. Risk of bias analysis was performed using the recently published PROBAST + AI tool and revealed generally low to unclear concern for risk of bias and low concern for applicability.

INTERPRETATION: AI models have been shown capable of predicting local and distant tumor recurrence in glioma patients from baseline MRI data. While the high odds ratios reported from the multicenter study are encouraging, the evidence comes mainly from small, single-center, retrospective cohorts, so larger prospective multicenter studies are needed before clinical adoption.

来源:Eur J Radiol,查看原文。版权归原作者所有。

引用

  1. Bijlsma S, Maal T, Rubbert C, et al. AI radiomics predicts spatial glioma recurrence on preoperative MRI: a systematic review. Eur J Radiol. 2025. PMID: 40987214. doi:10.1016/j.ejrad.2025.112412

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。