Scholarlink

Background and aim <p>Due to the insufficient data and the ongoing development of machine learning (ML), this study was conducted to examine a deep learning approach for enhancing the resolution of dental Bite-wing (BW) and Peri-Apical (PA) Radiographs (Rg) based on Super Resolution (SR) theory.</p> Methods and materials <p>1000 BW, and PA Rg were collected: 750 images for training while 250 images for test. At first step, we downscaled all High Resolution (HR) images to create Low Resolution (LR) ones using 4*4 average pooling without overlap. Thereafter, we incorporated three deep learning-based super-resolution approaches and the most efficient one (down sampled skip-connection/ Multi-scale (DSC/MS)) was chosen. After training, our ML algorithm was tested by the 250 LR images incorporated six evaluation metrics.</p> Results <p>After five-time repletion of our model, the mean ± S.D of R<sup>2</sup>, RSME, MSE, MAE, SSIM, and PSNR was 0.90 ± 0.0006, 0.039 ± 0.001, 0.0017 ± 0.00015, 0.026 ± 0.001, 0.85 ± 0.003, 28.45 ± 0.30. All these metrics was superior comparing to conventional methods.</p> Conclusion <p>Our SR model demonstrated significant effectiveness and the DSC/MS showed noticeably superior results comparing to linear, cubic, or nearest neighbor interpolations.</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

Super-resolution of periapical and bitewing digital radiographs using convolutional neural network

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月18日

查看全文
  • Hossein Pourahmadiyan-Nadiki,
  • Nafiseh Alemohammad,
  • Mahshid Mohammadi-bassir,
  • Faeze Hamze

摘要

Background and aim

Due to the insufficient data and the ongoing development of machine learning (ML), this study was conducted to examine a deep learning approach for enhancing the resolution of dental Bite-wing (BW) and Peri-Apical (PA) Radiographs (Rg) based on Super Resolution (SR) theory.

Methods and materials

1000 BW, and PA Rg were collected: 750 images for training while 250 images for test. At first step, we downscaled all High Resolution (HR) images to create Low Resolution (LR) ones using 4*4 average pooling without overlap. Thereafter, we incorporated three deep learning-based super-resolution approaches and the most efficient one (down sampled skip-connection/ Multi-scale (DSC/MS)) was chosen. After training, our ML algorithm was tested by the 250 LR images incorporated six evaluation metrics.

Results

After five-time repletion of our model, the mean ± S.D of R2, RSME, MSE, MAE, SSIM, and PSNR was 0.90 ± 0.0006, 0.039 ± 0.001, 0.0017 ± 0.00015, 0.026 ± 0.001, 0.85 ± 0.003, 28.45 ± 0.30. All these metrics was superior comparing to conventional methods.

Conclusion

Our SR model demonstrated significant effectiveness and the DSC/MS showed noticeably superior results comparing to linear, cubic, or nearest neighbor interpolations.