Evaluation of a Convolutional Neural Network-Based Artifact Reduction Algorithm for Zirconia Single Crown Artifacts...

Area of Science:

• Dental Imaging
• Artificial Intelligence in Medicine
• Biomedical Engineering

Background:

• Zirconia restorations in dentistry often cause artifacts in cone-beam computed tomography (CBCT) images.
• These artifacts degrade image quality and hinder accurate clinical diagnosis.

Purpose of the Study:

• To develop and evaluate CARNet, a novel convolutional neural network-based algorithm for reducing zirconia-induced artifacts in CBCT.
• To specifically address artifacts caused by zirconia single crowns.

Main Methods:

• CARNet utilizes a UNet backbone with a ResASPP module for multi-scale feature extraction.
• A dataset was created by simulating zirconia artifacts from the Zenodo repository.
• Performance was evaluated using objective metrics (NMAE, SSIM, PSNR) and subjective Likert scale assessments, compared against ADN, CycleGAN, and Linear Interpolation.

Main Results:

• CARNet significantly outperformed other methods in reducing zirconia artifacts, achieving NMAE of 0.013±0.001, SSIM of 0.959±0.007, and PSNR of 40.39±0.77 dB.
• Subjective evaluations confirmed CARNet's superior artifact suppression while preserving crucial crown details and marginal integrity.

Conclusions:

• CARNet is an effective and reliable method for mitigating zirconia restoration artifacts in CBCT imaging.
• The algorithm demonstrates satisfactory performance in restoring image quality affected by zirconia single crowns.