Comparative Evaluation of Super-Resolution Processed Image Quality of Lingual Mucosal Images by Generative Adversarial Network

Abstract

To evaluate new image processing methods using deep learning and to obtain basic data for improving the quality of oral mucosa images, we processed lingual mucosa images (dorsum of tongue, lateral border of tongue, and sublingual surface) obtained our previously developed observation device (oral mucoscope) using three types of Generative Adversarial Network (GAN), TecoGAN, Real-ESRGAN, and Real-CUGAN network. The images were evaluated objectively by calculated values, and sensory evaluation by dental clinicians. The TecoGAN-processed image had slight contour correction and appeared to be the most natural image. The Real-ESRGAN processed image was intermediate in quality, between that of TecoGAN and Real-CUGAN, and showed outline details of the lingual papillae. Real-CUGAN improved image graininess after processing but appeared to have poor morphological reproducibility of contour details. The objective evaluation value, LPIPS, showed a significant decrease in value in the order of TecoGAN, Real-ESRGAN, and Real-CUGAN, and a similar trend to the subjective evaluation value (MOS) by dentists, but in PSNR and SSIM, no significant differences were observed between the TecoGAN, Real-ESRGAN, and Real-CUGAN in PSNR and SSIM. In a comparison by observation aspect, the lingual dorsum was lower than the other observation aspects in terms of objective rating values, but there was no significant difference between the three observation surfaces in terms of subjective assessment values by dentists. The site of the mucosa may be affected by a deep “feature” extracted from the image. Further study is needed to evaluate images after super-resolution processing when considering the morphological characteristics of the mucosal surface.