Learned Tensor Neural Network Texture Prior for Photon-Counting CT Reconstruction.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Photon-counting computed tomography (PCCT) generates correlated multi-channel images.
• PCCT reconstruction is challenged by photon count scarcity, leading to noise and loss of detail.

Purpose of the Study:

• To develop a novel tensor neural network (TNN) architecture for PCCT reconstruction.
• To utilize multi-channel correlations for noise suppression and texture enhancement in PCCT images.

Main Methods:

• A tensor neural network (TNN) was designed to learn multi-channel texture priors.
• Spatial texture priors were learned per channel, then merged to capture spectral correlations.
• Low-rank representation was incorporated to model global channel correlations.
• The TNN and low-rank priors were integrated into a Bayesian reconstruction framework.

Main Results:

• The proposed TNN prior Bayesian reconstruction method demonstrated superior performance.
• It effectively preserved texture features across multiple energy channels.
• Significant noise suppression was achieved in each reconstructed channel image.
• Evaluated on simulated, dual-energy CT, and custom PCCT system data.

Conclusions:

• The TNN-based approach enhances PCCT image quality by exploiting inter-channel correlations.
• This method offers a robust solution for noise reduction and detail preservation in PCCT.
• The findings suggest a promising direction for advanced medical image reconstruction.