Preliminary denoising by 3D U-Net in image domain for low dose CT images

Low dose CT (LDCT) by reducing the X-ray tube current is of huge significance during clinical scanning. However, low-dose CT images often have strong noise and artifacts, which affects the image quality and diagnostic performance. LDCT noise reduction methods based on deep learning have recently achieved good results in improving image quality. Since the reconstructed CT image itself is 3D, in this paper a LDCT denoising method based on 3D U-Net is proposed to combine the 3D spatial information by 3D convolution directly, instead of processing 2D slices from 3D volume data. Therefore, the image change continuity between the adjacent slices is guaranteed. In addition, multiple down-sampling operations in the network, which can reduce the number of parameters of the 3D network, help the network to train. The experimental results show that the proposed method can effectively preserve the structural and texture information of normal NDCT images and significantly suppress the image noise and artifacts, achieving better performance in both quantification and visualization. Compared with LDCT images without denoising, the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of the processed images were improved by 12.18 dB and 0.35 dB, respectively.