Purpose: We aim to reduce complex artifacts due to both metal implants and sparse view sampling in dental CT images by utilizing prior images via convolutional neural network (CNN).
Methods: To generate sparse view dental CT images in the presence of metal implants, we inserted two circular-shaped gold implants within a region of the teeth and acquired projection data by simulating polychromatic X-ray CT scans with one-fourth of full view data set. CT images were then reconstructed with the filtered backprojection algorithm. Since the image quality is severely degraded by metal implants, we initially reduced the metal artifacts using linear metal artifact reduction (LMAR) algorithm. To reduce the streak artifacts, we trained CNN in the image domain, which is denoted by LMARCNN. Then, by using a thresholding, the prior image was acquired from the results of LMARCNN. By utilizing prior images, normalized metal artifacts reduction (NMAR) method is applied for the sparse view sampled sinogram data. Then, the residual streak artifacts are reduced again by CNN in the image domain, which is denoted by NMARCNN. For both CNNs, we employed a U-net architecture to ensure a large receptive field to capture the globally distributed streak artifacts.
Results: It is observed that NMARCNN can reduce the residual errors of NMAR and shows the best qualitative and quantitative quality of images through ROIs. The quantitative evaluation of mean squared error (MSE) and structural similarity (SSIM) also confirm our observations. Since the remaining streak artifacts in LMAR were reduced in LMARCNN, clear prior images can be acquired.
Conclusion: In this work, we proposed CNN-based NMAR approach for sparse view dental CT image reconstruction. Our results show that the combination of metal artifacts and streak artifacts due to sparse view sampling can be reduced effectively by the proposed method.
Funding Support, Disclosures, and Conflict of Interest: This research was supported by the National Research Foundation of Korea (NRF) grant funded by the Ministry of Science and ICT (NRF-2019R1A2C2084936, 2020R1A4A1016619)