Artificial intelligence-assisted cytology improves diagnostic consistency and accuracy in cervical cancer screening: A multicenter crossover study

Objectives: The effective prevention and control of cervical cancer (CC) depend on high-quality screening. However, traditional cytology interpretation faces challenges such as strong subjectivity, significant inter-observer variability, and a shortage of cytopathologists in primary care settings. The distribution of high-quality medical resources in China is highly uneven between urban and rural areas. Artificial intelligence (AI)-assisted cytology offers a new approach to enhance screening efficacy. This study evaluated the diagnostic performance of a deep learning-based AI-assisted cytology system in a multicenter clinical environment in China, comparing its utility as a standalone tool and in a collaborative model with cytopathologists. Materials and Methods: This was a multicenter, crossover study enrolling a total of 769 outpatient liquid-based cytology samples obtained from three participating hospitals. Using a crossover design, two junior-to-mid-level cytopathologists from primary hospitals performed independent diagnoses and AI-assisted diagnoses (following an AI preliminary screening combined with cytopathologist verification), which were compared with AI-alone diagnoses. Using an expert cytology panel diagnosis as the cytological reference standard, sensitivity (Se), specificity (Sp), overall accuracy, positive predictive value (PPV) negative predictive value, and false-positive rates/false-negative rates and the area under the receiver operating characteristic curve (AUC) were calculated and compared for each modality at the atypical squamous cells of undetermined significance (ASC-US+), low-grade squamous intraepithelial lesions (LSIL+), and high-grade squamous intraepithelial lesions (HSIL+) thresholds. Results: The study found that the independent diagnostic Se of the primary cytopathologists was insufficient (LSIL+: 46.0% (95% confidence interval [CI]: 0.359–0.564)–50.6% (95% CI: 0.403–0.608); HSIL+: 36.6% (95% CI: 0.236–0.519)–46.3% (95% CI: 0.321–0.613), indicating a high risk of missed diagnoses. AI-alone diagnosis showed high Se (ASC-US+: 92.9% [95% CI: 0.877–0.960]) but a relatively lower PPV (72.0% [95% CI: 0.654–0.778]). AI-assisted diagnostic model achieved a favorable balance. Across all diagnostic thresholds, the AI-assisted modality significantly outperformed cytopathologist-alone diagnosis. At the ASC-US+ threshold, AI-alone Se was 92.9% (95% CI: 0.877–0.960) with a Sp of 90.9% (95% CI: 0.883–0.929), whereas the sensitivities of the two primary cytopathologists alone were 68.4% (95% CI: 0.607–0.752) and 81.3% (95% CI: 0.744–0.866), respectively. With AI assistance, primary cytopathologist Se increased to 92.3% (95% CI: 0.870–0.955)–94.2% (95% CI: 0.893–0.969), and Sp reached 94.6% (95% CI: 0.925–0.961)–96.9% (95% CI: 0.952–0.980). At the LSIL+ and HSIL+ thresholds, where cytopathologist-alone sensitivities were low (LSIL+: 46.0% [95% CI: 0.359–0.564]–50.6% [95% CI: 0.403–0.608]; HSIL+: 36.6% [95% CI: 0.236–0.519]–46.3% [95% CI: 0.321–0.613]), AI assistance improved them to 80.5% (95% CI: 0.709–0.874) and 73.2% (95% CI: 0.581–0.843)–75.6% (95% CI: 0.607–0.862), respectively, with Sp exceeding 97.0% in both cases. Receiver operating characteristic curve analysis confirmed that the AUC for AI-assisted diagnosis (0.938 [95% CI: 0.911–0.965]–0.971 [95% CI: 0.946–0.996]) was significantly higher than that for cytopathologist-alone diagnosis (0.791 [95% CI: 0.746–0.836]–0.868 [95% CI: 0.819–0.918]). Conclusion: In this exploratory study, the AI-assisted cytology system may help reduce the experience gap for primary cytopathologists in primary hospitals, substantially improving the Se and Sp of cervical cytology interpretation. The AI-assisted diagnostic model is an effective strategy for achieving standardized, high-quality CC screening, particularly for promoting homogeneous services in regions with uneven healthcare resources.