A Deep Convolutional Gated Recurrent Unit for CT Image Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Filtered backprojection (FBP) is a standard but limited CT reconstruction technique, especially at low X-ray doses due to increased stochasticity.
• Iterative reconstruction (IR) offers improved performance by incorporating explicit models of the CT scan and prior knowledge, but constructing these priors is complex.

Purpose of the Study:

• To propose a novel neural network for computed tomography (CT) image reconstruction that addresses limitations of existing methods, particularly for low-dose imaging.
• To introduce a new deep learning approach based on iterative reconstruction principles and recurrent neural networks.

Main Methods:

• Developed a recurrent neural network (RNN) based on the gated recurrent unit (GRU), termed "GRU reconstruction," for CT image reconstruction.
• Implemented concurrent dual-domain learning, utilizing both sinogram and image data, unlike traditional single-domain deep learning methods.
• Introduced a novel RNN backpropagation algorithm, backpropagation through stage (BPTS), optimized for iterative reconstruction.

Main Results:

• The proposed GRU reconstruction method demonstrated superior performance compared to conventional model-based methods, single-domain deep learning, and state-of-the-art techniques.
• Quantitative improvements were observed in root mean squared error (RMSE), peak signal-to-noise ratio (PSNR), and structure similarity (SSIM).
• Visual assessment confirmed the enhanced image quality achieved by the novel dual-domain learning approach.

Conclusions:

• The GRU reconstruction method represents a significant advancement in low-dose CT image reconstruction.
• Dual-domain learning and the BPTS algorithm offer a powerful framework for improving CT image quality using deep learning.
• This novel approach holds promise for enhancing diagnostic accuracy in low-dose CT applications.