人工智能模型与放射科医生在mpMRI检测临床显著前列腺癌中的比较:一项荟萃分析
该荟萃分析纳入10项非随机对照研究共2586例患者,比较人工智能系统与放射科医生在mpMRI检测临床显著前列腺癌(Gleason分级组≥2)中的诊断效能。AI汇总敏感度0.90、特异度0.69,放射科医生敏感度0.89、特异度0.60,AUC分别为0.88与0.85,置信区间有重叠。结论认为AI表现与放射科医生相当,特异度可能略优,但高异质性和回顾性设计限制可靠性。
要点
- AI与放射科医生检测csPCa的敏感度相近(0.90 vs 0.89),特异度AI可能略高(0.69 vs 0.60),但置信区间重叠。
- AI的DOR为17.54,放射科医生为12.35,SROC的AUC分别为0.88与0.85,诊断准确性相似。
- 所有纳入研究均为回顾性,异质性高,AI作为辅助工具的价值需前瞻性验证。
核心结论
- AI系统在mpMRI检测临床显著前列腺癌方面与放射科医生表现相当,特异度可能略有优势,但置信区间重叠;高异质性和回顾性设计限制可靠性,需前瞻性验证;AI可作为辅助手段,有望提高精度并减少不必要活检。
研究设计与样本
- 系统综述与荟萃分析,检索PubMed、Embase和Cochrane数据库至2025年5月,共855项研究,最终纳入10项英文非随机对照研究、2586例患者,仅纳入比较基于CNN或深度学习AI模型与放射科医生的研究,采用双变量随机效应模型计算汇总敏感度、特异度、DOR和AUC。
诊断标准与关键数值
- AI系统检测csPCa的汇总敏感度为0.90(95%CI 0.84-0.94),特异度为0.69(95%CI 0.45-0.85)。
- 放射科医生检测csPCa的敏感度为0.89(95%CI 0.82-0.94),特异度为0.60(95%CI 0.43-0.75)。
- AI的DOR为17.54(95%CI 9.34-32.94),放射科医生的DOR为12.35(95%CI 4.96-30.76)。
- SROC曲线显示AI的AUC为0.88,放射科医生为0.85。
临床应用与阅片要点
- AI可作为前列腺癌诊断的辅助手段,可能提高诊断精度并减少不必要的活检,但需进一步模型优化。
- 由于置信区间重叠且异质性高,AI与放射科医生的诊断效能差异尚未确定,阅片时需结合临床综合判断。
局限性
- 所有纳入研究均为回顾性研究,限制可靠性。
- 存在高异质性,限制可靠性。
- 置信区间重叠,需前瞻性验证。
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
OBJECTIVE: Multiparametric magnetic resonance imaging (mpMRI) detects clinically significant prostate cancer (csPCa, Gleason Grade Group ≥ 2) with high sensitivity but limited specificity and inter-reader variability. Artificial intelligence (AI), particularly convolutional neural networks (CNNs) and deep learning, may improve diagnostic consistency and accuracy. This meta-analysis compares AI systems and experienced radiologists in detecting csPCa using mpMRI.
METHODS: We performed a systematic review and meta-analysis of English-language non-RCT studies. PubMed, Embase, and Cochrane databases were searched up to May 2025, yielding 855 studies. Only studies comparing CNN-based or deep-learning AI models to radiologists were included. Pooled sensitivity, specificity, diagnostic odds ratio (DOR), and area under the receiver operating curve (AUC) were calculated using a bivariate random-effects model.
RESULTS: Ten studies with 2586 patients were analyzed. AI systems showed pooled sensitivity of 0.90 (95% CI 0.84-0.94) and specificity of 0.69 (95% CI 0.45-0.85). Radiologists had a sensitivity of 0.89 (95% CI 0.82-0.94) and a specificity of 0.60 (95% CI 0.43-0.75). DOR was 17.54 (95% CI 9.34-32.94) for AI and 12.35 (95% CI 4.96-30.76) for radiologists. Summary receiver operating characteristic (SROC) curves indicated similar diagnostic accuracy, with AI slightly outperforming radiologists (AUC 0.88 vs. 0.85).
CONCLUSION: AI systems perform comparably to radiologists in detecting csPCa on mpMRI, with a potential edge in specificity, though confidence intervals overlapped. High heterogeneity and the retrospective nature of all included studies limit reliability, necessitating prospective validation. AI could serve as an adjunct in prostate cancer diagnosis, potentially improving precision and reducing unnecessary biopsies with further model refinement.
KEY POINTS: Question Interpretation of mpMRI for clinically significant prostate cancer varies among radiologists, affecting diagnostic consistency. Findings Meta-analysis shows AI has comparable sensitivity and potentially superior specificity to radiologists in detecting significant prostate cancer on mpMRI, though confidence intervals overlapped. Clinical relevance AI has the potential to enhance diagnostic accuracy, reduce unnecessary biopsies, and improve consistency in prostate cancer detection, thereby supporting more reliable and standardized imaging assessments across centers.
引用
- Andrade MA, Rodrigues H, Colhado CH, et al. Artificial-intelligence models vs. radiologists in the detection of clinically significant prostate cancer on mpMRI: a meta-analysis. Eur Radiol. 2026. PMID: 41844851. doi:10.1007/s00330-026-12465-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。