Supervised Learning for CT Denoising and Deconvolution Without High-Resolution Reference Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence

Background:

• Convolutional neural networks (CNNs) for CT super-resolution typically require high-resolution reference data.
• Conventional deconvolution methods for CT image enhancement often amplify noise, degrading image quality.

Approach:

• Developed a dual CNN model comprising a noise reduction U-Net and a deconvolution CNN.
• The deconvolution CNN utilizes a fixed decoder representing the point spread function and a trained encoder to prevent noise amplification.
• A difference of gradients loss function term was implemented to control ringing artifacts.

Key Points:

• The CNN model successfully mitigates noise amplification during deconvolution.
• Achieved visually sharper images with reduced noise compared to conventional methods.
• Demonstrated improved visual delineation of components in ex vivo kidney stones.

Conclusions:

• The developed CNN approach enhances spatial resolution and reduces noise in CT images.
• This technique eliminates the need for high-resolution reference images in CT super-resolution.
• The method shows promise for improving diagnostic accuracy in various clinical applications.