Deep Learning for Automated Prediction of Sphenoid Sinus Pneumatization in Computed Tomography

Area of Science:

• Radiology
• Artificial Intelligence
• Surgical Anatomy

Background:

• The sphenoid sinus is a critical surgical landmark for transsphenoidal procedures.
• Anatomical variations in sphenoid sinus pneumatization can increase surgical risks.
• Deep learning offers a potential solution for identifying these variations.

Purpose of the Study:

• To develop and evaluate a deep learning model for automated sphenoid sinus pneumatization pattern classification.
• To assess the model's accuracy in distinguishing between Conchal (I), presellar (II), sellar (III), and postsellar (IV) patterns.

Main Methods:

• A convolutional neural network (CNN) was developed to analyze mid-sagittal CT scans.
• Radiologists classified CT images into four pneumatization patterns.
• Data augmentation techniques were employed to enhance the training dataset.

Main Results:

• The CNN model achieved an overall diagnostic accuracy of 84% (AUC 0.84) after data augmentation.
• The model demonstrated excellent performance for type IV pneumatization (AUC 0.93, sensitivity 100%).
• High accuracy (97.33%) and specificity (99%) were observed for type I pneumatization.

Conclusions:

• The CNN model accurately identifies sphenoid sinus pneumatization patterns, particularly the high-risk type IV.
• This AI tool can assist surgeons in planning transsphenoidal surgeries, thereby enhancing patient safety.
• The model shows significant potential for clinical application in neurosurgery and ENT practices.