Abstract: Objective  To establish and validate an artificial intelligence (AI) model based on deep learning to automatically identify and quantitatively count eosinophils in nasal polyp tissue sections, thereby assisting in the pathological diagnosis of nasal polyps. Methods  A total of 220 surgical specimens from patients with chronic rhinosinusitis with nasal polyps in Beijing Chaoyang Hospital, Capital Medical University were retrospectively included to construct a comparative framework between manual  and AI  counting. Manual counting were performed by two senior pathologists and two beginners, who selected four hotspots under high magnification of a traditional microscope for counting. To achieve as much field-of-view matching as possible, the AI counting automatically identified the four hotspots with the highest eosinophil density on digital full-field images and calculated their average. The consistency between methods and among observers was systematically evaluated by using Pearson's correlation coefficient and intraclass correlation coefficient.  Results  The AI model demonstrated excellent recognition performance, with its counting results highly positively correlated with manual counting (r=0.925,P<0.001), and excellent consistency (ICC = 0.892, 95% CI: 0.859-0.917). In contrast, the consistency between the four counters was only moderate (ICC = 0.750).  Conclusion  The AI model developed in this study can achieve high-precision and high-consistency quantitative analysis of eosinophils in nasal polyps, significantly reducing the inter-observer variation inherent in traditional manual methods. This model provides objective and efficient technical support for the standardized pathological diagnosis of eosinophilic nasal polyps, and it is of great value in promoting the precise classification of this disease.

Key words: chronic rhinosinusitis with nasal polyps, artificial intelligence, eosinophil counting, deep learning, full-field digital images, pathological diagnosis