Diagnostic performance of artefact-reduced cone-beam CT images using a generative adversarial neural network

Generative adversarial neural networks (GAN) have been demonstrating efficacy in increasing the quality of tomographic images. However, there is a lack of information about its effectiveness in cone-beam CT (CBCT). This study evaluated the performance of GAN in reducing CBCT artefacts. A phantom was custom-made with a mandible covered with Mix-D. Forty teeth were endodontically instrumented and inserted in mandibular sockets. CBCT scans were obtained with a field-of-view of 5 × 5 cm, 90kVp, 3 mA, 0.08 mm voxel size and 9 s. All scans were repeated after enabling metal artefact reduction (MAR). Two conditions with high-density materials were simulated and additional scans were performed. GAN was tailored to use the acquired raw images for validation, training and testing. Structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), mean absolute error (MAE) and root mean square error (RMSE) were calculated. Then, all the images were reconstructed and noise and signal-to-noise ratio (SNR) were measured in the cervical, middle and apical third of the second left premolar root. Also, assessments of the diagnostic confidence, artefact interference and dental and bone preservation were performed by four observers. All the conditions were compared using ANOVA two-way (α = 0,05) and Wilcoxon signed rank test. The corrected images showed significantly higher SSIM and PSNR, lower MAE and RMSE, lower noise, higher SNR, higher diagnostic confidence, lower artefact interference and high dental and bone preservation. The CBCT images with MAR and higher presence of high-density materials showed significantly lower noise values. The developed neural network has promising performance to increase the image quality of CBCT raw and reconstructed images.