Development of an AI-based algorithm for the identification, segmentation, and classification of chronic Rhinosinusitis (CRS)

Introduction The aim of this study was to develop an AI-based algorithm for the identification, segmentation, and classification of chronic rhinosinusitis (CRS) with and without nasal polyposis using the Lund-McKay Score.

Materials and methods We developed Convolutional Neural Networks (CNNs) trained to detect and classify chronic rhinosinusitis. The database used comprised CT scans of patients with chronic rhinosinusitis±nasal polyposis and control CTs of individuals without related symptoms. The Lund-McKay Score was employed for classification. GradCAM techniques were applied to the CNNs to understand which image sections were used for predictions. By comparing this information with control CT images, we implemented a CNN capable of contouring diseased volumes, enabling a fully automatic quantitative analysis of the images.

Results Our findings were based on a cohort of 60 different CRS-CT scans. The images used for developing the AI-based algorithm were diverse in terms of age, gender, and acquisition modality. The system achieved an accuracy between 80 and 90% in identifying sinonasal lesions on this benchmark dataset.

Discussion The Lund-McKay score was precisely calculated using independent data sets, allowing the CNN system to be applied to all cohort types.

Publication History

Article published online:
19 April 2024