ERS3044. Refinement of Image Quality in Panoramic Radiography Using a Generative Adversarial Network
Authors * Denotes Presenting Author
  1. Hak-Sun Kim *; Yonsei University College of Dentistry
  2. Eun-Gyu Ha; Yonsei University College of Engineering
  3. Ari Lee; Yonsei University College of Dentistry
  4. Yoon Joo Choi; Yonsei University College of Dentistry
  5. Kug Jin Jeon; Yonsei University College of Dentistry
  6. Sang-Sun Han; Yonsei University College of Dentistry
  7. Chena Lee; Yonsei University College of Dentistry
The aim of this study was to develop a generative adversarial network (GAN) model for quality improvement of dental panoramic radiography and assess its clinical usefulness.

Materials and Methods:
Panoramic radiography of the patients who visited Yonsei University Dental Hospital on June 1st to 3rd, 2021 were reviewed and 100 images were randomly selected for the total dataset. This study was approved by the Institutional Review Board (IRB) of Yonsei University Dental Hospital (IRB No. 2-2022-0021) and the requirement of the informed consent was waived on account of the retrospective nature of the collection of images. Datasets with low quality images (n =400) were prepared by processing the selected images in 4 different methods: (1) Blur, (2) Noise, (3) Blur with noise and (4) Blur at anterior teeth region. The degraded images were paired with their original images and were distributed to the train and the test dataset in 9:1 ratio for each degraded images group. The Pix2Pix, GAN model, was trained using pairs of the original and the degraded image datasets for 100 epochs. For the test dataset, peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) were measured in MATLAB R2022a (MathWorks, Natick, MA, USA) and two oral and maxillofacial radiologists assessed clinical image quality scores from a previous study. The PSNR and SSIM, were compared by Analysis of Variance. The qualitative clinical image evaluation score differences were assessed by Kruskal-Wallis test and post-hoc Mann-Whitney test.

Among the 4 types of degraded images, the GAN model improved blur at anterior teeth region for the most, and the blur with noise group for the least (PSNR, 36.27 > 32.74; SSIM, 0.90 > 0.82). The mean PSNR value of each group was: Blur, 36.08±1.03; Noise, 32.82±0.69; Blur with noise, 32.74±0.70 and Blur at anterior region, 36.27±0.95. The mean SSIM value of each group was: Blur, 0.88±0.02; Noise, 0.82±0.03; Blur with noise, 0.82±0.03 and Blur at anterior region, 0.90±0.02. The model showed significantly higher performance for Blur and Blur at anterior region groups than Noise and Blur with noise groups (p < 0.05). While the mean score of clinical image quality of the original images was 44.6 points, the predicted image from the ‘Blur’ group showed the highest (=45.22 points). and the noise group showed the lowest (=36.00 points) score. The mean score of each group was: Original, 44.6±2.1; Blur, 45.2±1.2; Noise, 36.0±5.7; Blur with noise, 36.6±4.1 and Blur at anterior teeth, 44.5±2.1. The Noise, and the Blur with noise groups exhibited significantly lower scores to the original images (p < 0.01).

The developed Pix2Pix GAN model showed the potential to improvement of blurred and noisy panoramic radiographs, and it has shown better quantitative performance and clinical usefulness in refining blurred images. For more universal and steady function of the model, regardless of the degradation types, not only further exploration and refinement of numerous GAN models, but also large and heterogeneous dataset are anticipated.