[Diagnostic efficacy and teaching value of artificial intelligence-assisted diagnostic systems in cervical...

Objective: To systematically evaluate the diagnostic performance of an artificial intelligence (AI)-assisted diagnostic system in identifying squamous intraepithelial lesions of different grades, glandular cell lesions, and common pathogenic microorganisms in ThinPrep cytologic test of cervical samples, and to explore its potential value in improving the diagnostic capabilities of pathologists. Methods: A total of 1 096 cervical specimens were collected from patients undergoing ThinPrep cytologic test at Beijing Anzhen Hospital, Capital Medical University from October to December 2025. The "AI-assisted senior pathologist diagnosis" served as an initial gold standard. For uncertain cases, a consensus diagnosis reached by two senior cytopathologists after joint review was used as the final gold standard. The sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and concordance rate of the AI system for each diagnostic category were calculated. Cases of misdiagnosis and missed diagnosis were further reviewed, and the underlying morphological factors contributing to diagnostic errors were analyzed. Additionally, the application effect of AI as a continuous diagnostic quality improvement tool was evaluated. Results: The AI system showed significant variation in recognition performance across different diagnostic categories. For lesion diagnosis, the sensitivities were as follows: high-grade squamous intraepithelial lesion (HSIL), 100% (6/6);atypical squamous cells of undetermined significance (ASCUS), 90.5% (19/21);low-grade squamous intraepithelial lesion (LSIL), 76.5% (26/34), and atypical glandular cells (AGC), 75.0% (9/12). ASCUS had the lowest PPV (9.7%, 19/196) and the highest false-positive rate (16.1%, 177/1 096). For microorganism identification, clue cells showed the best performance [sensitivity 93.9% (46/49); PPV 95.8% (46/48)], followed by Candida [sensitivity 88.0% (44/50); PPV 72.1% (44/61)]. The sensitivities for Trichomonas and Actinomyces were both 100% (2/2; 2/2), but the PPVs were relatively low [11.1% (2/18); 22.2% (2/9), respectively]. Due to the small size of positive samples, these sensitivity estimates should be interpreted with caution. The system overall demonstrated high NPV (all >99%) and high specificity (all >98%). The main causes of misdiagnosis included the misinterpretion histiocytes in an atrophic background as AGC; the misclassification of pseudokoilocytes or reactive changes as LSIL; and the misidentification of multinucleated cells, inflammatory debris, or clusters of Lactobacillus for Herpes Simplex Virus (HSV), Trichomonas, or Actinomyces, respectively. Missed Candida diagnosis primarily occurred in cases with scant hyphae hidden within cell layers. Furthermore, the study found that by frequently presenting AI-flagged positive cases (e.g., ASCUS, AGC, and microorganisms), the system provided diagnostic pathologists with continuous, high-quality opportunities for case review and learning. Conclusions: In cervical cytology screening, the AI-assisted diagnostic system demonstrates high sensitivity and specificity for high-grade lesions and certain microorganisms (e.g., Candida and clue cells), along with an extremely high negative predictive value. It can effectively assist in primary screening and reduce the risk of missed diagnoses. Functioning as an intelligent pathology training adjunct, this system facilitates pathologists in reinforcing and refining their proficiency in lesion identification and differential diagnosis, thereby fostering iterative clinical advancement underpinned by human-AI synergism.