Quantitative and Qualitative Evaluation of a Confidence-Aware Transformer-Based Super-Resolution Framework for...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Medical Image Enhancement
• Radiology and Dental Diagnostics

Background:

• Panoramic dental radiographs are crucial for diagnosis but can suffer from low resolution.
• Enhancing image quality is vital for accurate interpretation and clinical decision-making.
• Existing super-resolution methods may struggle with diagnostic reliability in complex radiographic images.

Purpose of the Study:

• To develop and evaluate a novel confidence-aware transformer-based super-resolution framework (CAT-PRSR).
• To improve image quality and diagnostic reliability of panoramic dental radiographs.
• To enable adaptive learning focused on diagnostically relevant regions while minimizing noise-related artifacts.

Main Methods:

• Developed CAT-PRSR, integrating a transformer SR backbone with a confidence-aware training strategy.
• Utilized 1078 anonymized panoramic radiographs for training and testing.
• Evaluated performance using quantitative metrics (PSNR, SSIM, FID, etc.) and Mean Opinion Score (MOS) assessments, comparing against state-of-the-art models at various magnifications.

Main Results:

• CAT-PRSR significantly outperformed comparison models across all metrics and magnification levels (4×, 6×, 8×).
• Achieved highest PSNR and lowest FID, demonstrating superior image fidelity.
• Maintained diagnostic utility comparable to ground truth images in MOS evaluations, unlike other models.

Conclusions:

• The CAT-PRSR framework effectively enhances panoramic radiograph resolution and diagnostic reliability.
• This model shows potential for improving clinical decision-making in low-resolution scenarios.
• CAT-PRSR can serve as a reliable imaging resource for AI-driven dental research and applications.