Half-scan artifact correction using generative adversarial network for dental CT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Dental CT imaging often suffers from motion artifacts due to slow scanning.
• Existing artifact correction methods like Parker weighting may not fully resolve severe half-scan artifacts.
• Compromised image quality impacts the diagnostic value of dental CT, particularly for surface-rendered bone and tooth structures.

Purpose of the Study:

• To develop and evaluate a novel generative adversarial network (GAN) for correcting half-scan artifacts in dental CT images.
• To assess the performance of the proposed U-WGAN against traditional methods and other GAN architectures.
• To quantify improvements in image quality and processing efficiency.

Main Methods:

• A modified GAN, termed U-WGAN, utilizing a five-stage U-net generative network was employed.
• The network was trained using the Wasserstein loss function on dental CT images from 40 patients.
• Performance was evaluated quantitatively using image quality metrics and qualitatively through surface-rendered images, compared against Parker weighting, SRGAN, and m-WGAN.

Main Results:

• The proposed U-WGAN demonstrated superior performance over Parker weighting and other GANs across all evaluated image quality metrics.
• U-WGAN significantly reduced residual half-scan artifacts, leading to clearer surface-rendered bone and tooth images.
• Processing time for 3D images was drastically reduced to approximately 3 seconds, compared to 50-54 seconds for other GANs.

Conclusions:

• U-WGAN is a highly effective deep learning approach for correcting half-scan artifacts in dental CT.
• The method offers significant improvements in image fidelity and computational efficiency for dental imaging applications.
• This advanced artifact correction holds promise for enhancing diagnostic accuracy and workflow in dentistry.