Liquid-Based Oral Brush Cytology: Evaluation of Two Artificial Intelligence Models in Papanicolaou and Silver-Stained...

Oral squamous cell carcinoma (OSCC) is the most prevalent oral malignant neoplasm. Cytopathology may represent an important tool in the screening of OSCC, and liquid-based oral brush cytology (LBOBC) has been widely studied because of its clearer cell sample results. These cytopathological analyses could be more efficient with the aid of artificial intelligence. The objective of this study was to analyze the effectiveness of two AI models (Papanicolaou and AgNOR Slide Image Examiners) in LBOBC analyses.

Two human evaluators and the AI models performed cell maturation pattern analysis and mean nucleolar organizer region (NOR) per nucleus count in Papanicolaou and silver-stained nucleolar organizer region (AgNOR) oral cytopathological samples of 20 individuals, respectively. Inter-evaluator agreement was evaluated by kappa and intraclass correlation coefficient. Chi-square and Wilcoxon matched-pairs/Friedman tests analyzed differences between the conventional and LBOBC methods and among evaluators.

Kappa between the Papanicolaou AI model and each human researcher was substantial (k = 0.69) for the conventional method and moderate for the LBOBC (k = 0.55-0.53. There were statistical differences in the cellular type analysis between cytology methods and among evaluators (p < 0.001). The automated AgNOR model showed an excellent/highly good agreement with human evaluators for NOR count in both cytology methods, with and without bounding boxes. There was no statistical difference in the NOR count between methods (p > 0.05). In the conventional method, there were differences among evaluators (p < 0.05); in the LBOBC, there were not (p > 0.05).

The AgNOR automated model is reliable when assessing NOR count in oral samples processed by different cytological methods, when compared to the human analysis. The Papanicolaou model still needs more training with LBOBC samples.