人工智能辅助前列腺MRI诊断性能的系统评价与荟萃分析
该系统评价与荟萃分析纳入29项研究、7398例患者,比较AI与放射科医生通过MRI诊断临床显著前列腺癌(csPCa)的性能。结果显示,AI作为助手时灵敏度、特异度、PPV和NPV均显著优于单独人类阅片者,且不同经验水平的阅片者均可获益。
要点
- AI辅助人类阅片诊断csPCa的灵敏度86.5% vs 82.6%,特异度57.8% vs 50.0%。
- AI作为独立阅片者时特异度更高(58.7% vs 48.7%),但灵敏度降低(87.2% vs 90.1%)。
- 不同经验水平的阅片者均可通过AI辅助提升诊断性能。
核心结论
- 将AI作为助手整合到csPCa诊断流程中可提高准确性,尤其对经验较少的阅片者。
主要发现
- AI辅助人类阅片者诊断csPCa的灵敏度86.5% vs 82.6%(P=0.001),特异度57.8% vs 50.0%(P=0.028),PPV 64.3% vs 58.9%(P=0.001),NPV 82.9% vs 76.5%(P=0.001),CDR 40.5% vs 38.6%(P=0.093)。
- AI作为独立阅片者时特异度58.7% vs 48.7%(P=0.026),灵敏度87.2% vs 90.1%(P=0.017)。
- 亚组分析显示不同经验水平的阅片者均可通过AI辅助改善诊断性能。
临床怎么用
- 在前列腺MRI诊断流程中可将AI作为辅助工具,帮助提升csPCa检出准确性,尤其适用于经验较少的阅片者。
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: AI is increasingly integrated within prostate cancer diagnosis pathway. PURPOSE: To provide estimates of diagnostic accuracy of AI assistance for clinically significant prostate cancer (csPCa) via MRI. MATERIALS AND METHODS: A systematic search of PubMed, Embase, Cochrane, Scopus and Web of Science from January 2017 to October 2024 was performed for studies on the diagnostic utility of AI for prostate MRI. Diagnostic performance metrics were synthesized through hierarchical summary receiver operating characteristic modeling with random-effects assumptions. Specially, to test inferiority and potential superiority of AI, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), cancer detection rate (CDR), and accuracy was pairwisely compared between AI and radiologists in study level using odds ratios (ORs) with Z-statistics. RESULTS: 7398 patients from 29 studies with AI-vs-human pairwise comparison were included. When acting as an assistant to human readers, AI demonstrated superior performance compared to stand-alone human readers in diagnosing csPCa via MRI, specifically with higher sensitivity (86.5% vs 82.6%, P = 0.001), specificity (57.8% vs 50.0%, P = 0.028), PPV (64.3% vs 58.9%, P = 0.001), and NPV (82.9% vs 76.5%, P = 0.001) while maintaining comparable CDR (40.5% vs 38.6%, P = 0.093). When used as standalone readers, AI exhibited higher specificity (58.7% vs 48.7%, P = 0.026) but at the cost of reduced sensitivity (87.2% vs 90.1%, P = 0.017). Subgroup analysis indicated that readers of varying experience levels could all improve their diagnostic performance with AI assistance. CONCLUSION: Integrating AI as an assistant in csPCa diagnostic workflows could enhance accuracy, particularly for less experienced readers. CLINICAL TRIAL REGISTRATION INFORMATION: Trial Name: The efficiency comparison of radiologists with or without assistance of artificial intelligence in prostate cancer diagnosis: a meta-analysis. Registration date: April 17, 2024. REGISTRATION NUMBER: CRD42024533016. Registration information available at: https://www.crd.york.ac.uk/PROSPERO/ .
引用
- Xie XR, Hou Y, Shan S, et al. AI-aided diagnostic performance for prostate MRI: systematic review and meta-analysis. Prostate Cancer Prostatic Dis. 2025. PMID: 41272268. doi:10.1038/s41391-025-01053-9
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。