Machine learning method for the cellular phenotyping of nasal polyps from multicentre tissue scans

Area of Science:

• Medical Informatics
• Pathology
• Artificial Intelligence in Medicine

Background:

• Chronic rhinosinusitis with nasal polyps (CRSwNP) presents diverse phenotypes.
• Accurate identification of inflammatory cell types is crucial for CRSwNP classification and management.
• Existing diagnostic methods may lack convenience and precision.

Purpose of the Study:

• To develop and validate a convenient and accurate platform, CRSAI 1.0, for evaluating CRSwNP.
• To enable precise identification of four key inflammatory cell types within nasal polyp tissues.
• To correlate CRSwNP phenotypes with clinical outcomes like asthma and recurrence.

Main Methods:

• Utilized a semantic segmentation algorithm (Unet++ with Efficientnet-B4) for automated tissue analysis.
• Trained the CRSAI 1.0 platform using pathologist-annotated inflammatory cell data (eosinophils, neutrophils, lymphocytes, plasma cells).
• Validated the platform's performance on multicenter datasets, assessing mean average precision (mAP).

Main Results:

• CRSAI 1.0 achieved high mAP values across training, testing, and validation cohorts for all four inflammatory cell types.
• Performance metrics demonstrated consistency across internal and external datasets.
• Significant variations in CRSwNP phenotypes were observed in relation to asthma presence and disease recurrence.

Conclusions:

• CRSAI 1.0 demonstrates robust accuracy in identifying inflammatory cells in CRSwNP from multicenter data.
• The platform facilitates rapid diagnosis of CRSwNP.
• CRSAI 1.0 supports the development of personalized treatment approaches for CRSwNP.