Texture-preserving low dose CT image denoising using Pearson divergence.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Mean squared error (MSE) loss is common for image denoising but causes over-smoothed edges in low-dose computed tomography (LDCT).
• Deep learning models using MSE for LDCT denoising struggle with texture degradation due to the regression-to-the-mean problem.

Purpose of the Study:

• To develop an improved loss function for LDCT image denoising that enhances texture preservation.
• To address the limitations of MSE loss in maintaining image texture and edge quality.

Main Methods:

• Proposed a novel loss function combining MSE loss with Pearson divergence loss.
• Pearson divergence loss was computed in image space to measure intensity differences between denoised LDCT and normal-dose CT images.
• Employed multi-metric quantitative analysis using relative texture feature distance for evaluation.

Main Results:

• The proposed Pearson divergence loss significantly improved image texture compared to conventional MSE loss and generative adversarial networks (GANs).
• Both qualitative and quantitative evaluations demonstrated superior texture preservation with the new method.
• The approach effectively balances noise reduction and texture preservation, a challenge for GAN-type methods.

Conclusions:

• The Pearson divergence loss offers a more effective approach for LDCT image denoising, preserving crucial image textures.
• This method facilitates the generation of high-quality CT images, aiding clinical diagnosis and AI model development.
• The proposed loss function provides a straightforward way to balance noise minimization and texture preservation.