Scatter correction for cone-beam CT via scatter kernel superposition-inspired convolutional neural network.

Area of Science:

• Medical Imaging
• Computational Imaging
• Radiological Physics

Background:

• X-ray scatter in Computed Tomography (CT) causes signal bias and degrades image quality.
• Conventional scatter kernel superposition (SKS) methods are fast but lack accuracy due to challenges in scatter kernel determination.

Purpose of the Study:

• To develop a novel, accurate method for real-time X-ray scatter estimation and correction in CT imaging.
• To improve upon the limitations of traditional SKS methods using deep learning.

Main Methods:

• A new method combining SKS principles with convolutional neural networks (CNNs) was developed.
• The proposed CNN generates scatter amplitude and width maps from projection images, enabling scatter field estimation via convolution.
• Physics-informed network design reduced trainable parameters compared to other deep learning methods.

Main Results:

• The SKS-inspired CNN demonstrated superior performance over conventional SKS and other deep learning methods in both numerical simulations and physical experiments.
• The method achieved significant improvements in qualitative and quantitative aspects of scatter correction.
• Fewer trainable parameters were required due to the physics-informed network architecture.

Conclusions:

• The proposed SKS-inspired CNN effectively corrects scatter-related artifacts in CT imaging.
• This approach offers a promising solution for enhancing CT image quality and reducing signal bias.