Deep learning-assisted versus manual reading in routine cervical cytopathology: a multicentre randomised crossover trial.

Deep learning (DL) systems could improve diagnostic accuracy and efficiency in detecting cervical atypia, but their effectiveness remains insufficiently explored. This multicentre, randomised crossover trial evaluated the clinical utility of a DL system in cervical cytopathology. A total of 1,920 women aged 18 years or older undergoing liquid-based cytology for cervical cancer screening were included, and their slides were digitized and randomly assigned (1:1) to two reading sequences. Four non-expert cytopathologists with 1-3 years of experience assessed slides using DL assistance for one group and manual microscopy for the other, and then switched roles after a four-week washout period. Each slide was evaluated twice in a randomly shuffled order. DL significantly improved sensitivity (85.7% vs 71.3%, p < 0.001), with a difference of 14.3% (95% CI: 7.6% to 21.1%), exceeding the 5% superiority margin. Specificity was comparable (86.5% vs 85.1%, p = 0.238), and non-inferiority was supported, as the lower limit of the 95% CI for the difference (1.4%; 95% CI: -1.0% to 3.8%) was above the pre-specified margin of -5%. Reading time was markedly reduced with DL (175 seconds vs 31 seconds, p < 0.001). DL assistance could enhance both sensitivity and efficiency while rigorously preserving specificity in cervical cytology interpretation. Trial registration: ChiCTR2300078722.