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

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cytopathology

Background:

• Deep learning (DL) systems show potential for improving diagnostic accuracy and efficiency in cervical atypia detection.
• However, their clinical utility in cervical cytopathology requires further investigation.

Purpose of the Study:

• To evaluate the clinical utility of a DL system in cervical cytopathology for detecting cervical atypia.
• To assess the impact of DL assistance on diagnostic sensitivity, specificity, and reading time.

Main Methods:

• A multicentre, randomised crossover trial involving 1,920 women undergoing liquid-based cytology for cervical cancer screening.
• Digitized slides were assessed by non-expert cytopathologists with and without DL assistance, with roles reversed after a washout period.
• Each slide was evaluated twice in a randomly shuffled order to compare DL-assisted versus manual microscopy.

Main Results:

• DL significantly improved sensitivity (85.7% vs 71.3%, p < 0.001), exceeding the 5% superiority margin.
• Specificity was comparable (86.5% vs 85.1%, p = 0.238), with non-inferiority supported.
• Reading time was markedly reduced with DL assistance (31 seconds vs 175 seconds, p < 0.001).

Conclusions:

• DL assistance can enhance both sensitivity and efficiency in cervical cytology interpretation.
• The DL system rigorously preserves specificity, making it a valuable tool for improving cervical cancer screening.
• This study demonstrates the clinical utility of DL in augmenting cytopathologist performance.