Image-spectral decomposition extended-learning assisted by sparsity for multi-energy computed tomography...

Area of Science:

• Medical Imaging
• Computational Imaging
• Image Reconstruction

Background:

• Multi-energy computed tomography (CT) generates multiple images for material identification but suffers from noise, reducing signal-to-noise ratio (SNR).
• Multi-energy CT images exhibit inherent properties like local sparsity, nonlocal self-similarity, and spectral correlation, which can be exploited for improved reconstruction.

Purpose of the Study:

• To introduce an advanced image-spectral decomposition extended-learning assisted by sparsity (IDEAS) method for multi-energy CT image reconstruction.
• To effectively utilize intrinsic image priors for noise reduction and enhanced image quality in multi-energy CT.

Main Methods:

• The IDEAS method incorporates nonlocal low-rank Tucker decomposition (TD) to leverage spectral correlation and spatial self-similarity.
• It employs multi-task tensor dictionary learning (TDL) for adaptive spatial and spectral dictionary training during reconstruction.
• A weighted total variation (TV) regularization is used to promote local sparsity.

Main Results:

• IDEAS demonstrated superior image reconstruction quality in simulations, achieving lower Root Mean Square Error (RMSE) and higher Structural Similarity (SSIM) compared to other methods.
• Material decomposition, specifically for bone components, showed improved accuracy with IDEAS (RMSE as low as 0.0152).
• The algorithm exhibited efficient computational performance, with a significantly lower iteration time (98.8 s) compared to competing tensor decomposition methods.

Conclusions:

• The proposed IDEAS method effectively integrates multiple prior regularizations to enhance multi-energy CT image reconstruction.
• Both qualitative and quantitative evaluations confirm the algorithm's outstanding performance against state-of-the-art techniques.
• IDEAS offers a promising approach for high-quality multi-energy CT imaging and material decomposition.