Low-dose CT image denoising via frequency division and encoder-dual decoder GAN

Utilization of lower dose to generate CT images (LDCT) can reduce X-ray radiation damage to human body, but the resulting noise and artifacts will hinder its applications for clinical diagnosis. In order to solve the difficulty in separating the image artifacts from the normal tissue structure texture, we proposed a GAN network based on frequency division and encoder-dual decoder for LDCT image noise reduction. First, based on the fact that the artifact or noise is mainly distributed to high frequency band, we extracted the high frequency component of the image and used a multi-scale U-Net network for preliminary noise reduction. Second, we designed an encoder-dual decoder artifact extraction sub-network in the backbone noise reduction network. One of the decoders uses the residual structure constrained by confidence loss to perform secondary noise reduction, and the other decoder is used to generate the high-frequency edges and textures of the normal tissue structure, which would be added to the secondary noise reduction results. In addition, a multi-scale inception structure is designed to improve the multi-scale feature extraction and discrimination capabilities of the discriminator. The proposed method performs independent denoising on the high-frequency channel, according to the characteristics of noise and artifacts. In addition, the structure of a double decoder was specially designed to compensate for the component of the normal tissue edge and texture information loss during the denoising process; hence, the noise and artifacts are more targetedly suppressed. The computer experimental results show that, compared with the current popular denoising algorithms, the proposed denoising network can achieve better denoising effects while preserving the important information of CT images.