Deep learning based spectral CT imaging.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Healthcare

Background:

• Spectral computed tomography (CT) offers advantages in dose reduction and material discrimination.
• Low signal-noise ratio (SNR) in energy-specific projections presents reconstruction challenges.
• Traditional iterative methods are computationally expensive.

Purpose of the Study:

• To develop an efficient deep learning-based reconstruction method for spectral CT.
• To address challenges in low SNR and computational cost associated with spectral CT.
• To improve the quality of spectral CT images.

Main Methods:

• Developed ULTRA (U-net with Lp^p-norm, Total variation, Residual learning, and Anisotropic adaption), a deep learning reconstruction method.
• Employed multi-scale feature fusion, multichannel filtering, and a denser encoding architecture.
• Introduced a general Lp^p-loss (p>0) and incorporated cross-energy bin correlations for regularization.
• Utilized anisotropically weighted total variation for spatial-spectral domain sparsity.

Main Results:

• ULTRA demonstrated excellent performance on large-scale spectral CT datasets.
• Achieved superior results compared to competing algorithms in quantitative and qualitative evaluations.
• Validated through numerical simulations and preclinical experiments.

Conclusions:

• The proposed ULTRA network provides accurate, efficient, and robust high-quality spectral CT image reconstruction.
• Deep learning offers a computationally efficient alternative to traditional methods for spectral CT.
• The method shows significant potential for advancing spectral CT imaging applications.