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

Area of Science:

• Dentistry
• Medical Imaging
• Artificial Intelligence

Background:

• Insufficient data and rapid advancements in machine learning (ML) necessitate novel approaches for dental radiograph enhancement.
• Super Resolution (SR) theory offers a promising avenue for improving the diagnostic quality of dental images.

Purpose of the Study:

• To investigate a deep learning approach for enhancing the resolution of dental Bite-wing (BW) and Peri-Apical (PA) radiographs (Rg).
• To evaluate the effectiveness of different deep learning-based SR methods for dental imaging.

Main Methods:

• Collected 1000 dental radiographs (BW and PA), with 750 for training and 250 for testing.
• Downscaled High Resolution (HR) images to create Low Resolution (LR) images using 4*4 average pooling.
• Selected and trained the most efficient deep learning SR model: down sampled skip-connection/Multi-scale (DSC/MS).

Main Results:

• The DSC/MS model achieved superior performance across six evaluation metrics (R², RSME, MSE, MAE, SSIM, PSNR) compared to conventional methods.
• Key metrics included R² of 0.90 ± 0.0006, RSME of 0.039 ± 0.001, and SSIM of 0.85 ± 0.003.
• The model demonstrated high accuracy and reliability in super-resolving dental radiographs.

Conclusions:

• The developed Super Resolution (SR) model demonstrated significant effectiveness in enhancing dental radiograph resolution.
• The DSC/MS deep learning approach yielded noticeably superior results compared to linear, cubic, or nearest neighbor interpolations.
• This deep learning technique holds potential for improving diagnostic accuracy in digital dentistry.