Scholarlink

Background and purpose <p>Bitewing radiographs have high diagnostic value in detecting interproximal caries and assessing alveolar bone levels. However, image compression and sensor limitations may compromise image quality and diagnostic value. This study investigates whether two advanced deep-learning methods—a Real-Enhanced Super-Resolution Generative Adversarial Network (Real-ESRGAN) and a Swin Transformer–based image restoration network (SwinIR)—can effectively reconstruct clinically practical details from degraded bitewing images.</p> Materials and methods <p>A curated dataset of 4,004 high-quality bitewing radiographs was downsampled to produce paired low- and high-resolution images. Both deep-learning methods were fine-tuned using identical training protocols. Image quality was evaluated using peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). Additionally, a blinded panel of twelve dentists assessed anatomic clarity, diagnostic accuracy, and perceptual realism using a descriptive rating scale.</p> Results <p>The SwinIR achieved superior quantitative results, with a PSNR of 35.56 dB and a SSIM of 0.9287, outperforming Real-ESRGAN, which recorded 31.93 dB and 0.8227, respectively. Clinicians also favored SwinIR for its ability to preserve diagnostic structures such as tooth margins and trabecular patterns, whereas Real-ESRGAN was favored for its perceptual realism and smoother texture rendering.</p> Conclusion <p>For bitewing radiographs, the Swin Transformer–based super-resolution model demonstrated more faithful preservation of clinically important structures than Real-ESRGAN, making it the preferred choice when diagnostic precision is prioritized.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing diagnostic quality in dental bitewings using transformer and GAN-Based image restoration

  • 期刊论文

  • 发表日期：

    2026年2月4日

查看全文
  • Mostafa Abtahi,
  • Sara Majidinia,
  • Meisam Mansourzadeh,
  • Seyyed Hossein Hosseini Zarch,
  • Ghazal Asadi,
  • Yasin Talafi Noghani,
  • Mohammad Reza Shakiba,
  • Amirhossein Saeedi,
  • Shayan Yousefi

摘要

Background and purpose

Bitewing radiographs have high diagnostic value in detecting interproximal caries and assessing alveolar bone levels. However, image compression and sensor limitations may compromise image quality and diagnostic value. This study investigates whether two advanced deep-learning methods—a Real-Enhanced Super-Resolution Generative Adversarial Network (Real-ESRGAN) and a Swin Transformer–based image restoration network (SwinIR)—can effectively reconstruct clinically practical details from degraded bitewing images.

Materials and methods

A curated dataset of 4,004 high-quality bitewing radiographs was downsampled to produce paired low- and high-resolution images. Both deep-learning methods were fine-tuned using identical training protocols. Image quality was evaluated using peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). Additionally, a blinded panel of twelve dentists assessed anatomic clarity, diagnostic accuracy, and perceptual realism using a descriptive rating scale.

Results

The SwinIR achieved superior quantitative results, with a PSNR of 35.56 dB and a SSIM of 0.9287, outperforming Real-ESRGAN, which recorded 31.93 dB and 0.8227, respectively. Clinicians also favored SwinIR for its ability to preserve diagnostic structures such as tooth margins and trabecular patterns, whereas Real-ESRGAN was favored for its perceptual realism and smoother texture rendering.

Conclusion

For bitewing radiographs, the Swin Transformer–based super-resolution model demonstrated more faithful preservation of clinically important structures than Real-ESRGAN, making it the preferred choice when diagnostic precision is prioritized.