Artificial Intelligence‐Based Pathological Subtype Diagnosis of Nasal Polyps: A Multidimensional and Micro‐Visualization Study

BACKGROUND: Nasal polyps (NP) are common upper respiratory conditions with diverse inflammatory subtypes influencing clinical features and prognosis. Manual counting of inflammatory cells in microscopic images (MI) is laborious and subjective, limiting diagnostic precision and treatment decisions. METHODS: A total of 2457 slides from 20 hospitals were used to develop an AI-based NP subtype diagnosis system (NPSS). NPSS-MI was built using 1047 slides (15,705 MIs) annotated by pathologists. NPSS-WSI was trained on 1410 slides (21,150 images) combining PA-P2PNet for cell detection and U-KAN for region segmentation. Three-dimensional reconstruction (3DNP) using registration and point cloud analysis enabled spatial quantification of inflammatory cells. Twelve pathologists evaluated NPSS accuracy and efficiency on 200 slides, and recurrence prediction models were developed using logistic regression in a 131-patient cohort. RESULTS: NPSS achieved performance with a weighted average F1-score of 0.809 for cell detection and an intersection over union (IoU) of 0.827 for region segmentation with an internal dataset. External dataset performance showed an F1-score of 0.792 and an IoU of 0.815. Forty randomly accumulated MIs were approximated WSI results. NPSS-MI and NPSS-WSI reached accuracies of 90% and 91%, reducing diagnostic time from 193 to 8 s and from 10,450 to 250 s, respectively. Junior pathologists using NPSS improved accuracy from 50% to 89%. Inflammatory cells showed distinct spatial patterns in 3DNP. NPSS-WSI prognostic model outperformed the MI-based model (AUC 86.64% vs. 79.81%, p = 0.039). CONCLUSIONS: NPSS integrates MI, WSI, and 3DNP to enable accurate and efficient NP subtype diagnosis and prognosis prediction, greatly enhancing diagnostic precision and clinical utility.