EDRAM-Net: Encoder-Decoder with Residual Attention Module Network for Low-dose Computed Tomography Reconstruction

The medical application of Computed Tomography (CT) is to provide detailed anatomical structures of patients without the need for invasive procedures like surgery, which is very useful for clinicians in disease diagnosis. Excessive radiation exposure can lead to the development of cancers. It is of great importance to reduce this radiation exposure by using low-dose CT (LDCT) acquisition, which is effective, but reconstructed CT images tend to be degraded, leading to the loss of vital information which is one of the most significant drawbacks of this technique. In the past few years, multiscale convolutional networks (MSCN) have been widely adopted in LDCT reconstruction to preserve vital details in reconstructed images. Based on this inspiration, we proposed an encoder-decoder network with a residual attention module (EDRAM-Net) for LDCT reconstruction. The proposed EDRAM-Net embeds the cascaded residual attention module (RAM) block into the skip connection connecting the encoder-decoder architecture. Specifically, the encoder captures and encodes details in the latent space, which is reconstructed in the decoder of the network. The RAM blocks consist of three modules: the MSCN, channel attention module (CAN), and spatial attention module (SAM). The MSCN captures features at different scales, while the CAM and SAM focus on channel and spatial details during reconstruction. The performance of EDRAM-Net evaluated on the public AAPM low-dose dataset shows that the method has improved performance in terms of estimated image quality metric compared to other comparative methods. The ablation study further revealed that using the kernel size of (7×7) for the RAM block significantly enhanced the performance of our model. It was also observed that a higher number of RAM blocks yielded improved performance but at the expense of computational complexity.