Comparison of Gray Values of Cone-beam Computed Tomography With Hounsfield Units of Multislice Computed Tomography Using a U-net Based Network (preprint)

ABSTRACT

Background：Dental departments generally employ cone-beam computed tomography (CBCT) instead of conventional computed tomography (CT), due to its lower price, smaller dosage, and high spatial resolution. During the corona virus disease 2019 (COVID-19) outbreak, CBCT is highly recommended to replace intraoral radiography because it greatly reduces the risk of exposure to salivary droplets. However, CBCT's inability to quantitatively measure tissue attenuation limits its application in differential diagnosis. Methods:We employed a U-Net based network to generate synthetic CT from dental CBCT. The deep neural network can be trained end-to-end to learn the complex mapping between CBCT and CT values. By the U-Net architecture, low-level and high-level features are both utilized to get fine detailed synthetic CT. We applied our method on the collected dataset contains 62 patients. Results:Experimental results on four metrics -- mean absolute error (MAE), root-mean-square error (RMSE), structural similarity index (SSIM), and peak-signal-to-noise ratio (PSNR) -- showed significant improvement of the synthetic CT compared to the original CBCT data. The MAE and RMSE improvement percentages are 64.44% and 66.44%.The MAE level of synthetic CT for most of the tissues are small enough to separate most important tissues,including dentin and cancellous bone, dentin and root canal，implants and cortical bone.Conclusions:CBCT and synthetic CT values can be used to distinguish different high-attenuation structures that are of interest to dentists. The application of CBCT assisted by this U-net based network in medical imaging of other parts of the body is promising.