Deep Learning Assistance vs Manual Reading in Cervical Cytopathology

Key Takeaways

• Assisted reading was associated with higher sensitivity while specificity remained broadly similar to manual reading.
• Slide review was markedly faster when the deep learning system guided reader attention.

Deep learning-assisted review increased sensitivity to 85.7% from 71.3% during routine cervical cytopathology in a randomized crossover trial involving liquid-based cytology at four centers in China. Four non-expert cytopathologists evaluated the same satisfactory slides with and without assistance in alternating reading sessions. Specificity stayed similar, and review time was shorter when assistance was available.

The study included 1,920 women aged 18 years or older who underwent routine liquid-based cytology for cervical cancer screening at four pathology centers in China. Slides were assigned 1:1 to two reading sequences, and baseline characteristics were similar across groups. Four non-expert cytopathologists with 1-3 years of experience read each slide twice, once with web-based deep learning assistance and once by manual microscopy, separated by a four-week washout. Sensitivity and specificity were the primary outcomes, while AUC, predictive values, and reading efficiency were secondary endpoints, using expert consensus as the reference standard.

In the main comparison, assisted reading maintained the higher sensitivity described above, while specificity was 86.5% with assistance and 85.1% without it, with p = 0.238. The sensitivity difference was 14.3% with a 95% confidence interval of 7.6% to 21.1%, exceeding the prespecified 5.0% superiority margin. The specificity difference was 1.4%, with a 95% confidence interval from -1.0% to 3.8%, which remained above the -5.0% noninferiority margin. Average AUC was 0.861 with assistance versus 0.782 without assistance, and median review time per slide fell from 175 seconds to 30 seconds.

The standalone deep learning system performance showed an overall AUC of 0.879, sensitivity of 91.8%, and specificity of 84.0%. Sensitivity by abnormal cell type ranged from 86.1% to 100.0%, and HSIL+ detection reached 100.0% sensitivity with 85.0% specificity. Assisted reading showed higher sensitivity in several subgroup analyses, including menopausal status, ASC-US detection, the natural sediment method, both scanners, and two study sites.