High-Frequency Sensitive Generative Adversarial Network for Low-Dose CT Image Denoising

Abstract

Low-dose computed tomography (LDCT) imaging has attracted tremendous attention because it reduces the potential cancer risk for patients by decreasing the radiation dose. However, reducing the radiation dose may cause image quality degradation due to the introduction of noise and artifacts. The details of pathological information mainly exist in the high-frequency domain of LDCT image. Therefore, some useful details may be lost or destroyed while removing the noise and artifacts. To address this problem, we propose a high-frequency sensitive generative adversarial network (HFSGAN). The new generator includes two sub-networks. One is the high-frequency domain U-Net, which is specially designed to deal with the high-frequency components decomposed from LDCT image. The other is image space U-Net, which is used to process information from the whole image of LDCT. In addition, the discriminator in HFSGAN adopts an inception module to increase the receptive field and width of network, and to extract the multi-scale features of the true and false images. The experiments show that the proposed network preserves more texture details of denoised image while removing noise and artifacts. Compared with the state-of-the-art networks, the proposed denoising method achieves better performance both quantitatively and visually. INDEX TERMS Low-dose CT, image denoising, GAN, U-Net, inception module.

A review on Deep Learning approaches for low-dose Computed Tomography restoration

“…Apart from the noise, the LDCT images are degraded by blurring [ 13 , 60 , 73 ] and streaking artifacts [ 28 , 34 , 50 , 71 , 75 , 81 , 91 ]. Lack of X-ray photons during the CT scanning and patient motion cause blurring.…”

“…However, after publishing the Pix-to-Pix GAN by Isola et al [ 31 ], there were several LDCT restoration applications published based on it. The main reason for that is the generator of the Pix-to-Pix GAN followed the U-net architecture, and it accepts an image as the input instead of the noise distribution in the latent space [ 75 , 79 ]. The deeper U-net published in [ 79 ] permits to retain of the small details of the processed LDCT images.…”

“…This basic structure of GAN is known as vanilla GAN. Moreover, GAN is flexible to implement different generator models based on various CNN architectures, such as the encoder–decoder [ 58 , 67 ], U-Net [ 6 , 53 , 75 , 79 ], and ResNet [ 12 , 28 , 46 ]. Also, the discriminator mostly acts as a binary classifier to distinguish the synthetic and NDCT images apart.…”

“…4 encoder and 4 deconvolution layers Three conveying paths between the corresponding encoder and decoder layers No CNN with 6 convolutional layers L5, L8 2D-to-3D transfer learning method is efficient and stable [ 72 ] Loss of texture and tissue details [ 21 ] SMGAN, You et al [ 83 ] W-GAN Voxels of size 80 × 80 × 11 3D CNN with 8 convolution layers No No CNN with 6 convolutional layers L4, L6, L7 It successfully preserves the critical features to be diagnosed Edge blurring and structural dissimilarities appeared in some cases Kang et al [ 35 ] Cycle GAN Patches of size 56 × 56 CNN with 20 convolution layers 6 bypass connections in 6 convolution blocks Yes Patch-GAN structure of [ 31 ] (Patch size is 70 × 70) L5, L10, L11 Resistance to mode collapse and robust to cardiac motion and contrast change Lack of texture preservation due to the low stability of the DL model [ 75 ]. Altering the local image details [ 80 ] m-WGAN, Hu et al [ 28 ] W-GAN Patches of size 64 × 64 ResNet with 16 Residual blocks (38 Convolution layers) 16 Bypass connection in 16 residual blocks Yes CNN with 8 convolution layers L3, L6 Minimizing the over-smoothing The noise and artifacts remain in the restored images Park et al [ 53 ] Vanilla GAN Patches of size 128 × 128 U-net like deep convolutional framelet …”

“…However, repeated CT scans may reveal that the patient may be exposed to radiation enormously. Overexpose to the radiation would cause the development of metabolic abnormalities, radiation-induced cancer, and other genetic disorders that fall the patients’ quality of life rapidly [ 75 ]. Therefore, low-dose CT (LDCT) scanning protocols have been proposed to minimize patients’ exposure to radiation while maintaining adequate diagnostic accuracy.…”