Explainable hybrid CNN-transformer with self-supervised learning for structural analysis of paranasal sinus CT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Anatomy

Background:

• Accurate paranasal sinus evaluation aids in diagnosing ear, nose, and throat (ENT) conditions.
• Deep learning methods struggle with complex sinus structures due to limited data and interpretability.

Purpose of the Study:

• To develop an explainable hybrid AI framework for precise structural analysis of paranasal sinus CT scans.
• To enhance the diagnosis and treatment of ENT conditions through improved anatomical assessment.

Main Methods:

• Implemented an explainable hybrid CNN-Transformer framework with a self-supervised 3D convolutional autoencoder.
• Utilized the multi-institutional CT-SCOPE dataset for evaluating diverse paranasal sinus CT volumes.
• Combined anatomical segmentation with residual-based structural representation learning for anomaly detection.

Main Results:

• Achieved high anatomical fidelity with Dice similarity coefficients >0.83 across all four paranasal sinus regions.
• Integrated convolutional and Transformer-based modeling for detailed structural and contextual analysis.
• Generated reconstruction residual maps highlighting structural deviations and confirmed clinical relevance via Grad-CAM.

Conclusions:

• The developed framework offers a comprehensive approach to paranasal sinus CT image analysis.
• Establishes a foundation for AI-assisted diagnosis and future pathology-aware clinical modeling.