Low-dose CT reconstruction using cross-domain deep learning with domain transfer module.

Area of Science:

• Medical Imaging
• Computational Science

Background:

• Low-dose X-ray computed tomography (CT) is crucial for reducing radiation exposure.
• Traditional analytic reconstruction methods suffer from noise artifacts with low photon counts.
• Deep learning (DL) methods, including uni-domain and dual-domain networks, have shown promise in noise reduction.

Purpose of the Study:

• To develop a more computationally efficient deep learning approach for low-dose X-ray CT reconstruction.
• To overcome the redundancy in computational resources required by existing dual-domain networks.
• To improve reconstruction quality while minimizing radiation dose.

Main Methods:

• Proposed a cross-domain deep learning (DL) approach utilizing analytical domain transfer functions.
• Implemented a network with a projection-domain encoder and an image-domain decoder.
• Leveraged transfer functions to reduce redundant encoder/decoder computations, optimizing resource efficiency.

Main Results:

• The proposed cross-domain network achieved comparable results to dual-domain networks with half the trainable parameters.
• Demonstrated superior reconstruction quality compared to conventional iterative reconstruction and existing DL methods.
• Effectively reduced image noise and Poisson noise inherent in low-dose CT.

Conclusions:

• The cross-domain DL approach offers a significant advancement in low-dose X-ray CT reconstruction.
• This method achieves high-quality imaging with reduced radiation doses and computational demands.
• The use of domain transfer functions enables efficient feature transfer across domains, enhancing performance.