Evaluation of a convolutional neural network-based artifact reduction algorithm for zirconia single crown artifacts in cone-beam CT.

OBJECTIVES

Zirconia-induced artifacts frequently degrade the image quality in dental cone-beam CT (CBCT) and complicate clinical diagnosis. This study aimed to develop and evaluate a novel image-domain artifact reduction algorithm based on convolutional neural networks, named Crown Artifact Reduction Net (CARNet), specifically for mitigating artifacts caused by zirconia single crowns.

METHODS

The CARNet model was architectured on a UNet backbone, enhanced with a dedicated Residual Atrous Spatial Pyramid Pooling (ResASPP) module designed for multi-scale feature extraction and fusion via residual and short-cut connections. A dataset was constructed from the public Zenodo clinical repository by simulating zirconia artifacts. Algorithm performance was assessed using both objective metrics-normalized mean absolute error (NMAE), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR)-and subjective evaluation via a Likert scale. CARNet's efficacy was compared against established image-domain algorithms, including Artifact Disentanglement Network (ADN) and Cycle-Consistent Adversarial Network (CycleGAN), as well as traditional linear interpolation (LI).

RESULTS

CARNet processing yielded significantly superior results for zirconia artifacts (axial/occlusal surfaces) than other methods, with NMAE: 0.013 ± 0.001/0.013 ± 0.001; SSIM: 0.959 ± 0.007/0.956 ± 0.008; PSNR: 40.39 ± 0.77/41.41 ± 0.80 dB. Subjective assessments confirmed that CARNet effectively suppressed artifacts while better preserving crown details and marginal integrity.

CONCLUSIONS

The CARNet model proves to be an effective and reliable method for reducing zirconia restoration artifacts in CBCT, demonstrating satisfactory performance in restoring images affected by zirconia single crowns.