文献速递技术与安全CTMR

人工智能用于CT和MRI检查方案分配:传统机器学习、BERT和大语言模型的Meta分析

Artificial Intelligence for CT and MRI Protocoling: A Meta-Analysis of Traditional Machine Learning, BERT, and Large Language Models.
PubMed 文献 · AJR Am J Roentgenol · 2026-02-25

该Meta分析纳入23项研究、共1,196,259份影像检查申请单,评估传统机器学习、BERT模型和大语言模型自动分配CT和MRI检查方案的表现。总体汇总准确率为85%(95%CI 83-87%),三种方法两两比较无显著差异(均P > .05),表现最佳的模型以微调BERT模型为主(最高为BioBERT,准确率93%)。

要点

  • 总体汇总方案分配准确率为85%(95%CI 83-87%),传统机器学习83%、BERT 87%、大语言模型86%,两两比较均无显著差异(均P > .05)。
  • 表现前10的模型以微调BERT模型为主(6个),最佳为BioBERT,准确率93%。
  • 常见错误来源包括申请单文本歧义、数据不平衡导致低使用量方案分配错误、同一申请存在多个临床合理方案。

核心结论

  • 用于自动CT和MRI方案分配的表现最佳人工智能模型以微调BERT模型为主。
  • 人工智能工具在帮助简化放射科医生工作流程方面显示出较大潜力,可能通过人工智能-放射科医生混合模式实现。
  • 微调大语言模型值得进一步探索。

研究设计与样本

  • 系统检索MEDLINE、Embase、Scopus、Web of Science、IEEE Xplore和Google Scholar截至2025年7月的研究,纳入报告人工智能技术为CT或MRI申请单分配方案性能的研究。
  • 最终分析纳入23项研究,共1,196,259份影像检查申请单。
  • 申请单亚专科包括体部影像(n = 4)、肌肉骨骼影像(n = 3)、神经放射学(n = 6)、胸部影像(n = 1)和多个亚专科(n = 9)。
  • 16项研究评估传统机器学习模型,8项评估BERT模型,5项评估大语言模型;任务特异性模型微调在传统机器学习中3项研究进行、BERT模型中全部研究进行、大语言模型中1项研究进行。
  • 各研究测试的所有模型准确率分别提取,采用随机效应Meta分析汇总,人工智能方法间比较采用Welch t检验,常见错误来源进行定性总结。

诊断标准与关键数值

  • 总体汇总方案分配准确率为85%(95%CI 83-87%)。
  • 传统机器学习模型汇总准确率为83%(95%CI 80-85%),BERT模型为87%(95%CI 85-89%),大语言模型为86%(95%CI 83-89%)。
  • 三种人工智能方法任意两两组合的汇总准确率均无显著差异(均P > .05)。
  • 30个不同模型(14个传统机器学习模型、9个BERT模型、7个大语言模型)中,表现前10的模型包括2个传统机器学习模型、6个BERT模型和2个大语言模型。
  • 表现最佳的模型为BioBERT(生物医学领域BERT),准确率93%。

临床应用与阅片要点

  • 常见错误来源包括申请单文本歧义、数据不平衡导致低使用量方案分配错误、同一申请存在多个临床合理方案、以及难以处理包含与不同方案强相关术语的申请单。
  • 人工智能工具可能通过人工智能-放射科医生混合模式帮助简化放射科医生工作流程。

局限性

  • 微调大语言模型值得进一步探索。

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. Examination protocoling is a resource-intensive task. Various artificial intelligence (AI) approaches have been investigated to automate this process. OBJECTIVE. The purpose of this study was to evaluate performance of traditional machine learning (ML) models, bidirectional encoder representations from transformers (BERT) models, and large language models (LLMs) for automated CT and MRI protocoling. EVIDENCE ACQUISITION. MEDLINE, Embase, Scopus, Web of Science, IEEE Xplore, and Google Scholar databases were searched through July 2025 for studies reporting the performance of an AI-based technique in assigning protocols for CT or MRI requisitions. Accuracy results were separately extracted for all models tested in each study and pooled using a random-effects meta-analysis. AI approaches were compared using Welch t tests. Common sources of error were qualitatively summarized. EVIDENCE SYNTHESIS. The final analysis included 23 studies, comprising 1,196,259 imaging requisitions. Requisition subspecialties included body imaging (n = 4), musculoskeletal imaging (n = 3), neuroradiology (n = 6), thoracic imaging (n = 1), and multiple subspecialties (n = 9). Sixteen studies evaluated traditional ML models, eight evaluated BERT models, and five evaluated LLMs. Task-specific model fine-tuning was performed in three studies for traditional ML models, all studies for BERT models, and one study for LLMs. The overall pooled protocoling accuracy was 85% (95% CI, 83-87%). The pooled accuracy was 83% (95% CI, 80-85%) for traditional ML models, 87% (95% CI, 85-89%) for BERT models, and 86% (95% CI, 83-89%) for LLMs; these pooled accuracies were not significantly different between any pairwise combination of the three AI approaches (all p > .05). Among 30 distinct models (14 traditional ML models, nine BERT models, seven LLMs), the top-10 performing models comprised two traditional ML models, six BERT models (including the top performing model [BioBERT, a biomedical-domain BERT; accuracy, 93%]), and two LLMs. Common sources of error included ambiguous requisition text, data imbalance yielding incorrect protocol assignments for low-volume protocols, the presence of multiple clinically reasonable protocols for given requisitions, and difficulty handling requisitions containing terms strongly associated with disparate protocols. CONCLUSION. The top-performing AI models for automated CT and MRI protocoling included predominantly fine-tuned BERT models. CLINICAL IMPACT. AI tools show strong potential to help streamline radiologist workflows, possibly through hybrid AI-radiologist approaches. Fine-tuned LLMs warrant further exploration. TRIAL REGISTRATION. PROSPERO identifier CRD420251088671.

来源:AJR Am J Roentgenol,查看原文。版权归原作者所有。

引用

  1. Sacoransky E, Chauhan Y, Adams SJ, et al. Artificial Intelligence for CT and MRI Protocoling: A Meta-Analysis of Traditional Machine Learning, BERT, and Large Language Models. AJR Am J Roentgenol. 2026. PMID: 41159787. doi:10.2214/AJR.25.33759

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。