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

Jaehyup Lee; Chang-Hyeon An; Seo-Young An; Eun-Kyong Kim; Young-Eun Kwon

doi:10.1016/j.identj.2026.109590

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

Int Dent J. 2026 Apr 27;76(4):109590. doi: 10.1016/j.identj.2026.109590. Online ahead of print.

Authors

Jaehyup Lee¹, Chang-Hyeon An², Seo-Young An³, Eun-Kyong Kim⁴, Young-Eun Kwon⁵

Affiliations

¹ Department of Computer Science and Engineering, Kyungpook National University, Daegu, Republic of Korea.
² Department of Oral and Maxillofacial Radiology, Kyungpook National University School of Dentistry, IHBR, Daegu, Republic of Korea.
³ Department of Oral and Maxillofacial Radiology, Kyungpook National University School of Dentistry, IHBR, ITRD, Daegu, Republic of Korea.
⁴ Department of Preventive Dentistry, School of Dentistry, Kyungpook National University, Daegu Republic of Korea.
⁵ Department of Oral and Maxillofacial Radiology, Kyungpook National University School of Dentistry, IHBR, Daegu, Republic of Korea. Electronic address: yekwon@knu.ac.kr.

Abstract

Objectives: This study aimed to develop and evaluate a confidence-aware transformer-based super-resolution framework, termed CAT-PRSR, to enhance image quality and diagnostic reliability in panoramic dental radiographs.

Methods: A total of 1078 anonymised panoramic radiographs were retrospectively collected (950 for training, 128 for testing). The CAT-PRSR framework integrating a transformer-based SR backbone with a confidence-aware training strategy was developed. The model generates a high-resolution output with pixel-wise uncertainty estimation, allowing adaptive learning focused on diagnostically relevant regions while minimising over-enhancement in noise-sensitive areas. Model performance was evaluated using 6 quantitative metrics - peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), spatial correlation coefficient (SCC), natural image quality evaluator (NIQE), learned perceptual image patch similarity (LPIPS), and Fréchet inception distance (FID)-and mean opinion score (MOS) assessment. Based on quantitative performance, 4 representative state-of-the-art SR models were selected for comparison at 4×, 6×, and 8× magnifications.

Results: CAT-PRSR demonstrated superior performance across all metrics and magnification levels. It achieved the highest peak signal-to-noise ratio (36.41 at 4 ×, 36.19 at 6 ×, and 33.73 at 8 ×) and the lowest FID (1.77, 9.29, and 2.09, respectively), outperforming all comparison models. In MOS evaluations, CAT-PRSR maintained diagnostic utility scores statistically comparable to ground truth images (P > .05), while other models showed significant degradation (P < .001).

Conclusion: The proposed CAT-PRSR framework demonstrated potential to enhance panoramic radiograph resolution by integrating pixel-level fidelity with improved diagnostic reliability.

Clinical relevance: The CAT-PRSR model may enhance the diagnostic reliability of panoramic radiographs acquired under low-resolution conditions, supporting more accurate clinical decision-making and serving as a reliable imaging resource for AI-driven dental research.

Keywords: Artificial intelligence; Deep learning; Diagnostic imaging; Image processing, Computer-assisted; Panoramic radiography.