ADMM-TransNet: ADMM-Based Sparse-View CT Reconstruction Method Combining Convolution and Transformer Network.

Area of Science:

• Medical Imaging
• Computational Imaging
• Radiology

Background:

• X-ray computed tomography (CT) is crucial for clinical diagnostics, but radiation exposure is a concern.
• Sparse-view scanning reduces radiation but poses reconstruction challenges.
• Existing methods struggle with manual tuning or data dependency and lack global correlation capture.

Purpose of the Study:

• To develop an advanced sparse-view CT reconstruction algorithm.
• To combine model-driven and data-driven techniques for improved accuracy and reduced data requirements.
• To enhance the learning of global and local image representations.

Main Methods:

• A hybrid approach integrating model-driven and data-driven methods.
• Utilizing the ADMM iterative algorithm framework to constrain deep learning models.
• Employing Convolutional Neural Network (CNN) and Transformer models for enhanced image representation.

Main Results:

• The proposed method demonstrates superior performance in sparse-view reconstruction.
• Achieved high quantitative metrics: PSNR of 42.036 dB, SSIM of 0.979, and MAE of 0.011 at 32 views.
• Outperformed current advanced reconstruction algorithms in qualitative and quantitative evaluations.

Conclusions:

• The developed algorithm is highly effective for sparse-view CT reconstruction.
• Exhibits better generalization capabilities compared to existing deep learning algorithms.
• Achieves superior reconstruction accuracy in sparse-view CT imaging.