深度学习辅助系统提升宫颈细胞病理诊断效果:不同阅片模式的对比研究
该研究采用随机对照四交叉设计,纳入2021至2022年回顾性收集的1620张宫颈细胞涂片,由108名不同经验的细胞病理医生分别以无辅助、同步、二次及分诊四种模式阅片。与无辅助模式相比,所有深度学习辅助模式均提升诊断性能,其中二次模式灵敏度提升最高,分诊模式特异度提升最高,同步模式效率最优;问卷显示超过90%医生认可辅助模式提升诊断信心并偏好同步模式。
要点
- 二次模式灵敏度提升最显著,分诊模式特异度提升最显著,同步模式阅片时间最短。
- 超过90%细胞病理医生认为深度学习辅助模式有助于提升诊断信心,且偏好同步模式。
- 分诊模式在模式评分中表现最高,二次模式以高灵敏度见长。
核心结论
- 与无辅助模式相比,所有深度学习辅助阅片模式均提升细胞病理医生的诊断性能。
- 二次模式灵敏度提升突出,分诊模式特异度提升突出,同步模式效率最优。
研究设计与样本
- 随机对照四交叉设计,评估四种宫颈细胞病理阅片模式。
- 回顾性纳入2021至2022年1620张宫颈涂片,由108名不同经验的认证细胞病理医生阅片。
主要发现
- 无辅助模式灵敏度66.2%、特异度72.5%。
- 二次、同步和分诊模式灵敏度分别提升20.2%、17.3%和16.7%,特异度分别提升13.2%、10.8%和22.3%。
- 每张涂片中位阅片时间在二次模式从200秒延长至235秒,分诊模式降至130秒,同步模式降至53秒(P < .0001 for all)。
- 超过90%细胞病理医生同意深度学习辅助模式有助于提升诊断信心,偏好同步模式,分诊模式模式评分最高。
临床怎么用
- 可根据临床需求选择辅助模式:追求高灵敏度选二次模式,追求高特异度选分诊模式,追求效率选同步模式。
- 深度学习辅助模式可提升细胞病理医生诊断信心,同步模式接受度较高。
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
Deep learning (DL) has significantly improved the diagnostic accuracy and efficiency of cytopathologists. However, current DL-assisted reading modes have yet to be fully evaluated, and there is limited evidence regarding cytopathologists' preferences and experiences. This study employs a randomized, controlled, 4-way crossover design to assess the effectiveness of 4 distinct cytopatholog reading modes. This study included retrospectively collected 1620 cervical slides between 2021 and 2022. These slides were read by 108 certified cytopathologists with varying expertise using the 4 reading modes: unassisted, concurrent, second, and triage mode. A questionnaire survey was conducted to gather the cytopathologists' adoption of each mode, including mode score and their confidence and preferences. Compared with unassisted, all DL-assisted modes improved the cytopathologists' diagnostic performance. The unassisted mode had a sensitivity of 66.2% and a specificity of 72.5%. The second, concurrent, and triage modes all improved on these metrics: sensitivity increased by 20.2%, 17.3%, and 16.7%, respectively, whereas specificity increased by 13.2%, 10.8%, and 22.3%, respectively. The median reading time per slide was prolonged in second mode from 200 to 235 seconds but substantially reduced to 130 seconds in triage mode, and 53 seconds in concurrent mode (P < .0001 for all). Moreover, the questionnaire survey showed that >90% of cytopathologists agreed that the DL-assisted mode was useful in improving diagnostic confidence, and they preferred the concurrent mode, whereas the triage mode achieved the highest performance in the mode score. The second mode stands out for its heightened sensitivity, the triage mode boasts superior specificity, and the concurrent mode leads in efficiency when assisting cytopathologists. These findings provide useful information for choosing appropriate DL-assisted modes in cytopathological practice.
引用
- Ye Z, Zhang P, Wei R, et al. Deep Learning-Assisted System Improves Practical Effects in Cervical Cytopathology Diagnosis: A Comparative Study of Reading Modes. Mod Pathol. 2025. PMID: 40914500. doi:10.1016/j.modpat.2025.100882
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。