Deep learning detection and classification of fungal and non-fungal calcifications on paranasal sinus CT imaging.

Abstract

This study aimed to develop and evaluate a deep learning algorithm for detecting and classifying intrasinus calcifications on paranasal sinus (PNS) computed tomography (CT) for the diagnosis of fungal sinusitis and differentiation of fungal and non-fungal sinusitis. A dataset of 277 PNS CT cases from Korea University Guro Hospital, supplemented by temporal and geographic external test sets, was utilized. A 3D U-Net model was employed to segment maxillary sinus regions. YOLO v5 identified calcifications, followed by classification into three patterns: normal sinus or chronic sinusitis without calcifications, dense peripheral dystrophic calcification, and central punctate fungal calcification. A separate convolutional neural network (CNN) refined the classification to ensure accurate categorization of calcification patterns. The 3D U-Net model achieved a Dice Similarity Coefficient of 0.9674, indicating accurate segmentation. YOLO v5 demonstrated precision of 79.50% and recall of 92.14% in detecting calcifications. The CNN classification model attained F1 scores of 94.73%, 90.60%, and 94.01%, and overall accuracies of 97.48%, 86.87%, and 94.01% for internal, temporal, and geographic test sets, respectively. This study demonstrated the capability of deep learning algorithms to accurately detect and classify fungal sinusitis-related calcifications on PNS CT scans. The developed framework achieved high accuracy in segmentation of sinus area and detection/classification of intrasinus calcifications. The framework also demonstrated its potential for broader application to radiographic imaging.