Automated classification of osteomeatal complex inflammation on computed tomography using convolutional neural...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Otolaryngology

Background:

• Convolutional Neural Networks (CNNs) excel at image classification tasks.
• CNNs have broad applications in areas like facial recognition and fraud detection.
• This study explores CNN application in diagnosing sinonasal conditions.

Purpose of the Study:

• To retrain a CNN using coronal CT images for classifying osteomeatal complex (OMC) occlusion.
• To evaluate the performance of this AI technology in analyzing rhinologic data.

Main Methods:

• Utilized Google's Inception-V3 CNN as a base model.
• Retrained the CNN using 956 coronal CT images from 239 chronic rhinosinusitis patients.
• Employed a transfer learning approach for sinonasal CT image interpretation.

Main Results:

• The retrained CNN achieved 85% accuracy in classifying OMC occlusion (95% CI: 78%-92%).
• Receiver operating characteristic (ROC) curve analysis showed good classification ability (AUC = 0.87).
• Performance was significantly better than random guessing and dominant classifiers (p < 0.0001).

Conclusions:

• State-of-the-art CNNs can learn clinically relevant information from 2D sinonasal CT images with minimal supervision.
• Future research will focus on 3D images to enhance clinical applicability.
• This AI approach shows promise for future clinical applications in rhinology.