文献速递乳腺腹部技术与安全乳腺X线CTMR

AI用于医疗筛查的承诺与挑战:多模态筛查应用综述

AI for screening in healthcare: promise and challenges.
PubMed 文献 · Abdom Radiol (NY) · 2026-01-12

本文为一篇叙述性综述,系统回顾了人工智能在乳腺X线、糖尿病视网膜病变、结肠镜及多癌早筛中的当前应用。AI在多项前瞻性试验中显示出非劣效检出能力或工作流改善,但面临低患病率背景下假阳性放大、泛化性不足及落地障碍。文章提出基础模型、联邦学习及不确定性驱动的决策转诊等未来方向。

要点

  • 乳腺X线AI前瞻性试验显示癌症检出非劣效并显著减少放射科工作量。
  • 实时结肠镜AI提高腺瘤检出,但未增加非肿瘤组织切除。
  • 低患病率人群中AI假阳性负担被放大,需警惕过度诊断与不必要检查。

核心结论

  • AI在多种筛查路径中已验证为可增强医师能力、扩大可及性并提高效率的工具,但持久公共卫生效益取决于公平设计、严格评估与持续人工监督。
  • 当前从实验室性能到人群获益的转化仍然有限,需要前瞻性试验追踪间期癌、分期迁移和成本效果。

研究设计与样本

  • 叙述性综述,描述当前AI在主要筛查路径中的应用现状。
  • 纳入乳腺X线前瞻性试验、糖尿病视网膜病变FDA授权自主系统研究、结肠镜实时计算机视觉研究及多癌早筛多模态系统相关文献。

关键影像表现/征象

  • 乳腺X线AI在影像判读中表现出与医师相当的癌症检出能力,同时大幅降低阅片工作量。
  • 糖尿病视网膜病变AI自主系统可在初级诊疗中提供专科级筛查,并改善筛查参与率。
  • 结肠镜实时AI可识别并提示可疑腺瘤性息肉,提高腺瘤检出且不增加非肿瘤组织切除。

诊断标准与关键数值

  • 本文未报告具体敏感度、特异度、AUC、阈值或分级标准的原始数值。
  • 文中提及AI系统在乳腺X线筛查中达到「非劣效」癌症检出,但未给出具体统计量或P值。
  • 文中提及AI提高腺瘤检出但未增加非肿瘤组织切除,未附具体检出率数值或置信区间。

临床应用与阅片要点

  • 影像科可将AI作为第二阅读者整合入乳腺X线筛查流程,以减少阅片负担的同时维持检出水平。
  • 结肠镜报告中需区分AI提示的腺瘤性息肉与非肿瘤性病变,避免不必要切除。
  • 在低患病率筛查场景中解读AI阳性提示时应结合临床背景,避免假阳性导致的过度检查。
  • AI系统跨机构部署前需评估数据分布差异与模型泛化能力,避免偏倚。

局限性

  • 原文述及存在基础率问题:低患病率队列中假阳性负担被放大。
  • 原文述及跨机构和人群的泛化性有限,存在潜在偏倚。
  • 原文述及工作流、监管和信任方面的实用落地障碍。

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

Artificial intelligence (AI) is reshaping population screening, yet the translation from laboratory performance to population benefit remains limited. This narrative review describes current uses of AI across major screening pathways. Prospective trials in mammography demonstrate non‑inferior cancer detection with large reductions in radiologist workload. In diabetic retinopathy, the first FDA‑authorized autonomous system extends specialist‑level screening into primary care and improves uptake. During colonoscopy, real‑time computer vision improves adenoma detection without increasing removal of non‑neoplastic tissue. Emerging multimodal approaches, including transformer‑based and large language model-enabled systems, integrate images, clinical variables, and molecular signals and underpin multi‑cancer early detection tests. Despite these gains, three constraints currently limit impact: the base‑rate problem in low‑prevalence cohorts, which magnifies the burden of false positives; limited generalizability and potential bias across institutions and populations; and practical barriers in workflow, regulation, and trust. Opportunities ahead include foundation models pre‑trained on diverse data, uncertainty‑aware "decision referral," federated learning, larger representative datasets, and prospective trials that track interval cancers, stage shift, and cost‑effectiveness. The overarching conclusion is cautious optimism: when validated and invisibly integrated, AI augments physicians, expands access, and improves efficiency; realizing durable public‑health benefits will depend on equity‑focused design, rigorous evaluation, and sustained human oversight.

来源:Abdom Radiol (NY),查看原文。版权归原作者所有。

引用

  1. Saito, Walston, Takita, et al. AI for screening in healthcare: promise and challenges. Abdom Radiol (NY). 2026. PMID: 41524988. doi:10.1007/s00261-025-05370-4

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。