CT Image Denoising Using Patch-Based Reinforcement Learning with Frequency-Selective Actions
Abstract
Purpose
The patch-based network training is widely used for CT denoising due to computation memory and time constraints. However, conventional techniques are limited to learncontextual information and preserve fine structures. In this study, we proposed a patch-based reinforcement learning CT denoising framework using frequency selective actions for performance improvement. Also, the performance of the proposed framework was evaluated with different training patch sizes.
Methods
The proposed framework was based on an asynchronous advantage actor-critic (A3C) algorithm, and the main network consisted of feature extraction, policy and value sub-networks. The each sub-network had convolution and dilated convolution layers, and the features with a size of 32, 64 and 128 were extracted in the feature extraction sub-network. The extracted features were shared in the two sub-networks for determining appropriate actions and estimating reward values. An action set included pixel-wise intensity updates, spatial-domain filters and two frequency-selective filters for balancing between noise suppression and structural preservation. A total of 11 actions was defined, and the determined action was implemented by multiple agents to adjust current states. The performance of the proposed framework was compared to the DnCNN, UNet, FFDNet and SwinIR.
Results
For all the patch sizes, the proposed framework showed the highest accuracy compared to the other models. The average PSNR and SSIM of the proposed framework with the various patch size were 32.73 and 0.82, respectively, while the maximum PSNR and SSIM of the other models were 29.69 and 0.68, respectively. The visual inspection showed that the edges of structures and fine textures were better preserved by the proposed framework with the small patches.
Conclusion
The proposed framework with the frequency-selective actions enables the spatially adaptive CT denoising and mitigates the degradation of spatial resolution. Thus, the proposed framework can be considered as an alternative techniques for improving CT image quality.