Low-dose spectral CT reconstruction based on structural prior network.

Area of Science:

• Medical Imaging
• Computer Vision
• Signal Processing

Background:

• Spectral CT imaging suffers from statistical noise due to insufficient photon counting, impacting image quality and material decomposition accuracy.
• Deep learning offers a promising approach for noise reduction in medical imaging, but its effectiveness can be limited by noisy training data.

Purpose of the Study:

• To develop a novel deep learning method for effective denoising of low-dose spectral CT images, specifically addressing the challenge of noisy training labels.
• To improve the robustness of deep learning models in spectral CT image processing by mitigating the adverse effects of noise in reference images.

Main Methods:

• A structural prior network (SP-Net) was proposed, integrating structural information from prior images into the network's loss function.
• The SP-Net utilizes a compressed sensing-inspired framework, guiding network training with both label supervision and prior structural information.
• This dual guidance mechanism aims to reduce the impact of noisy labels and enhance overall image quality.

Main Results:

• The proposed SP-Net demonstrated successful denoising in both simulated and experimental spectral CT data.
• The method effectively eliminated the detrimental effects of noisy labels, significantly reducing image noise.
• Crucially, the SP-Net preserved essential image structures, maintaining diagnostic information.

Conclusions:

• The developed SP-Net offers a robust solution for training deep learning models with noisy spectral CT labels, overcoming a significant limitation in current methods.
• This research provides valuable insights for future deep learning applications in spectral CT denoising.
• The method holds substantial potential for enhancing image quality and diagnostic accuracy in clinical medical imaging.