Automated detection of maxillary sinus opacifications compatible with sinusitis from CT images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Otolaryngology

Background:

• Sinusitis presents a significant healthcare burden, with frequent misdiagnosis between inflammatory sinusitis and cystic formations in the maxillary sinus.
• Accurate differentiation is crucial for effective clinical treatment of maxillary sinus opacifications.

Purpose of the Study:

• To develop and assess a deep learning model for improved diagnostic accuracy in differentiating maxillary sinusitis, retention cysts, and normal sinuses.
• Investigate the feasibility of an automated system for maxillary sinus lesion detection.

Main Methods:

• A You Only Look Once (YOLO) v8 nano object detection model was trained on 1080 coronal-view CT images (2158 maxillary sinuses).
• The dataset included normal sinuses (1138), cysts (366), and sinusitis (654), divided into training, validation, and testing sets.
• Transfer learning and data augmentation techniques were employed to enhance model robustness and address data limitations.

Main Results:

• The model achieved an overall precision of 97.1% on the test set (normal: 96.9%, cyst: 95.2%, sinusitis: 99.2%).
• An average F1-score of 95.4% was recorded, with the highest scores for normal sinuses, followed by sinusitis, and then cysts.
• Precision decreased to 92.4% on a challenging subset of the test data.

Conclusions:

• The developed deep learning model demonstrates feasibility for assisting clinicians in the screening of maxillary sinusitis lesions.
• This AI-powered approach has the potential to enhance diagnostic accuracy and streamline clinical workflows for maxillary sinus conditions.