Fourier-enhanced high-order total variation (FeHOT) iterative network for interior tomography

Area of Science:

• Medical Imaging
• Computational Imaging
• Image Reconstruction

Background:

• Traditional computed tomography (CT) methods like Filtered Back-Projection (FBP) struggle with low contrast and detail loss.
• Deep learning approaches for CT reconstruction often face challenges with data consistency and may over-smooth images.
• Interior tomography reconstruction from truncated projection data remains a significant challenge, impacting image quality and diagnostic accuracy.

Purpose of the Study:

• To develop a high-precision interior tomography reconstruction method using high-order total variation (HOT) regularization and Fourier-based frequency domain enhancement.
• To overcome limitations of existing methods, including slow convergence, over-smoothing, and loss of high-frequency details.
• To achieve accurate reconstruction from truncated projection data, enhancing both contrast and edge preservation.

Main Methods:

• Proposed a Fourier-enhanced HOT (FeHOT) network utilizing a coarse-to-fine strategy.
• Employed a HOT-based unrolled iterative network with a learned primal-dual algorithm for data consistency and high-order gradient constraints.
• Integrated a Fourier-enhanced U-Net module to selectively process frequency components, preserving edge and texture details from Filtered Back-Projection (FBP) results.

Main Results:

• FeHOT demonstrated superior performance over FBP, HOT, AG-Net, and PD-Net on AAPM and clinical medical datasets.
• Achieved high Peak Signal-to-Noise Ratio (PSNR) values (e.g., 41.17 noise-free, 39.24 noisy on medical data), significantly outperforming existing methods.
• Showcased significant improvements in edge preservation (e.g., SSIM increase from 0.9877 to 0.9976) and high-quality reconstruction within five iterations.

Conclusions:

• FeHOT represents a significant advancement in interior tomography by integrating classical HOT theory with deep learning.
• The introduction of frequency-domain operations effectively addresses limitations associated with piecewise-constant assumptions in CT images.
• FeHOT offers a computationally efficient and accurate solution for high-quality interior tomography reconstruction, suitable for low-dose imaging applications.