Axial Consistent Memory GAN With Interslice Consistency Loss for Low Dose Computed Tomography Image Denoising
- Resource Type
- Periodical
- Authors
- Bera, S.; Biswas, P.K.
- Source
- IEEE Transactions on Radiation and Plasma Medical Sciences IEEE Trans. Radiat. Plasma Med. Sci. Radiation and Plasma Medical Sciences, IEEE Transactions on. 8(2):173-183 Feb, 2024
- Subject
- Nuclear Engineering
Engineered Materials, Dielectrics and Plasmas
Bioengineering
Computing and Processing
Fields, Waves and Electromagnetics
Computed tomography
Three-dimensional displays
Image restoration
Noise reduction
Convolution
Logic gates
Generators
Axial consistent discriminator
inter slice similarity in computed tomography (CT)
low dose CT (LDCT) denoising
memory network
- Language
- ISSN
- 2469-7311
2469-7303
Degradation of the diagnostic quality of Low Dose CT (LDCT) due to noise is a major bottleneck that hinders the widespread application of LDCT imaging as an alternative to standard-dose computed tomography (CT) imaging. Recent deep learning-based methods for denoising LDCT images utilizing interslice similarity of CT slices have shown promising results; however high computational cost of those methods is a significant bottleneck for practical deployment. This study introduces an alternative approach for utilizing the interslice similarity among the CT slices for LDCT denoising. First, we propose a novel memory network that remembers the aggregate information about the previous slices and uses it to denoise the contemporary slice. Next, we proposed a novel axial consistent discriminator network and a novel interslice consistency loss to assist the memory network in learning the interslice information flow. Our proposed discriminator network works both as a spatial discriminator and a volume discriminator. The proposed loss appeals to the network to generate an output consistent with the previous axial slice, consequently helping to suppress the artifacts present in the current slice. The extensive experiment on two publicly available datasets validates that our method performs favorably against the existing state-of-the-art methods.