多供应商AI生态的延迟、报告周转与采纳情况
一项历时4.5年、覆盖20个门诊影像中心的多供应商AI实施项目,通过回顾性技术队列、报告周转时间队列和放射科医生调查评估基础延迟、工作流及医生态度。结果显示总体AI主动使用率达91.4%,中位总延迟2.06分钟,其中72%来自数据路由;AI可用性与创伤X线和膝关节MRI报告周转时间显著缩短相关。
要点
- AI结果到达晚于报告完成的比例总体为7.2%,其中胸部CT最高为13.2%,可能影响实际阅片采用。
- AI缩短报告周转时间主要体现在创伤X线(-26%)和膝关节MRI(-18%),脑容积MRI未见显著改善。
- 基础设施数据路由占中位总延迟的72%,部署AI时应优先优化PACS流程而非算法速度。
核心结论
- 大规模多供应商AI在高流量检查中与可测量的报告周转时间缩短相关。
- 基础数据路由延迟而非算法速度是限制临床效用的主要障碍。
研究设计与样本
- 研究类型为多中心回顾性观察性研究,包含三个回顾性队列。
- 技术队列纳入96874例检查,时间为2023年9月至2025年9月。
- 报告周转时间队列纳入20909例检查,时间为2025年3月至2025年8月。
- 调查队列为58名放射科医生,2025年进行两轮调查。
诊断标准与关键数值
- 总体AI主动使用率为91.4%(53/58),其中66%(35/53)报告经常使用。
- 中位总延迟为2.06分钟,四分位距为1.74至3.05分钟,其中72%归因于数据路由。
- 「过迟」率范围从膝关节MRI的3.0%到胸部CT的13.2%。
- 创伤X线AI可用性与中位报告周转时间降低26%相关(P<0.001),膝关节MRI降低18%(P<0.001),脑容积MRI变化+9.2%(P=0.33)。
临床应用与阅片要点
- 创伤X线和膝关节MRI的AI可用性与报告周转时间显著缩短相关,适合优先推广至高流量部位。
- 胸部CT的AI结果13.2%在报告完成后才到达,阅片时不应依赖其作为首轮筛查工具。
- 净推荐值在胸部CT从+38降至-3、主动脉CT从+22降至-25,提示医生对部分AI工具的实际体验不佳。
局限性
- 研究为回顾性观察性设计,无法排除混杂因素对报告周转时间的影响。
- 净推荐值探索性分析未进行多重比较校正,结果仅为名义显著。
- 研究基于门诊影像网络环境,结果可能不适用于住院或急诊工作流。
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: To evaluate infrastructure latency, workflow, and radiologist sentiment across a 4.5-year, multi-vendor AI implementation program in a 20-center outpatient radiology network.
METHODS: Three retrospective cohorts: a technical cohort (96,874 examinations; Sep 2023-Sep 2025) for PACS-to-PACS latency and temporal alignment of 10 AI tools from 7 vendors; a report turnaround time (TAT) analysis cohort (20,909 examinations; Mar-Aug 2025) comparing TAT between AI-available and concurrent non-AI workflows by Mann-Whitney U; and a survey cohort (58 radiologists; two waves 2025) for adoption, Likert-scale perceptions, and Net Promoter Score (NPS).
RESULTS: Active AI adoption was 91.4% (53/58), with 66% (35/53) reporting regular use. Median total latency was 2.06 minutes [interquartile range 1.74-3.05], 72% of which attributable to data routing. The "Too Late" rate (AI result arriving after report finalization) was 7.2% overall, ranging from 3.0% for knee MRI to 13.2% for chest CT. After adjustment for radiologist (linear mixed-effects models), AI availability was associated with lower median TAT for trauma radiography (-26%) and knee MRI (-18%; both p<0.001); brain volumetry MRI showed no significant change (+9.2%; p=0.33). In exploratory analysis, NPS declined for chest CT (+38 to -3) and aorta CT (+22 to -25), nominally significant before multiple-comparison correction.
DISCUSSION: Multi-vendor AI at scale was associated with measurable TAT gains in high-volume modalities. Infrastructure latency, not algorithm speed, was the primary barrier to clinical utility. At ∼23-25% annualized cost of one radiologist full-time equivalent (FTE) salary, the program generated 0.69 FTE of capacity through trauma radiography alone (0.46 FTE after radiologist-adjusted sensitivity analysis).
引用
- Morozov, Heracleous, Korka, et al. AI Latency, Report Turnaround Time, and Adoption in a Multi-Vendor AI Ecosystem: A Multi-Site Observational Study. J Am Coll Radiol. 2026. PMID: 42810607. doi:10.1016/j.jacr.2026.09.026
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。