Effect of network architectures in multi-agent reinforcement learning for denoising digital tomosynthesis images

A supervised-learning strategy has showed outstanding performance in various fields. However, the application of the supervised learning for restoring digital tomosynthesis (DT) image is limited due to the necessity of a large amount of training data, the high computational cost and the occurrence of structural distortion in an output image. In this study, a multi-agent reinforcement learning (RL) network was developed to provide a denoising model for the DT image, and the effect of the network architecture was evaluated in terms of noise property, quantitative accuracy and training time. The multi-agent RL network consisted of shared, value and policy sub-networks, and the each sub-network was connected with the others in the form of a fully convolutional network (FCN). Five types of the shared sub-network were designed using general convolution layers, dilated convolution layers and residual blocks to modify the network architecture. The network training was implemented to extract image features, output a self-calculated reward and determine an optimal policy for the shared, value and policy sub-networks, respectively. The results showed that the proposed denoising models not only preserved the high quantitative accuracy but also improved the noise property of the output DT image by an average factor of 6.73. The applications of the dilated convolution layer and residual block enhanced the noise reduction capability of the proposed denoising models, and the training efficiency was dependent on the architecture type of the shared sub-network. In conclusion, the proposed model based on the multi-agent RL enables the DT image denoising, and the performance of the denoising model can be optimized by modifying the architecture of the shared sub-network.