Tensor decomposition and non-local means based spectral CT image denoising.

Area of Science:

• Medical Imaging
• Image Processing
• Computational Science

Background:

• Photon counting detectors in spectral computed tomography (CT) generate multichannel projections.
• Low photon counts in each energy channel lead to severe noise in Filtered Back Projection (FBP) reconstructed images.
• Existing spectral CT reconstruction algorithms can be computationally expensive.

Purpose of the Study:

• To propose a novel spectral CT image denoising method.
• To enhance the quality of low-dose spectral CT images.
• To improve the efficiency of spectral CT image reconstruction.

Main Methods:

• A tensor decomposition and non-local means (TDNLM) based denoising approach is developed.
• Leverages spatial self-similarity and inter-channel correlation in spectral CT images.
• Utilizes a high signal-to-noise ratio image synthesized from all energy channels to guide denoising and preserve signal fidelity.

Main Results:

• The TDNLM method significantly reduced noise in spectral CT images, improving Root Mean Square Error (RMSE) from 0.225 cm-1 to 0.0217 cm-1 and Structural Similarity (SSIM) from 0.633 to 0.987.
• Demonstrated effectiveness in both numerical simulations and preclinical applications.
• Achieved comparable denoising performance to state-of-the-art iterative methods with over 50x speedup in computational cost.

Conclusions:

• The proposed TDNLM method offers outstanding denoising performance for spectral CT.
• The technique is computationally efficient and practical for clinical applications.
• Adaptive parameter selection ensures optimal image quality and material decomposition.