Noise-Generating and Imaging Mechanism Inspired Implicit Regularization Learning Network for Low Dose CT...

Area of Science:

• Medical Imaging
• Computational Imaging
• Radiology

Background:

• Low-dose computed tomography (LDCT) aims to minimize radiation exposure while preserving diagnostic image quality.
• Existing deep learning methods for LDCT reconstruction often overlook the inherent noise characteristics of projection data, limiting performance.
• Current frameworks treat LDCT reconstruction as a general inverse problem, failing to fully leverage domain-specific information.

Purpose of the Study:

• To develop a novel full-domain reconstruction model for LDCT that accounts for noise generation mechanisms.
• To integrate statistical properties of intrinsic LDCT noise and prior information from both sinogram and image domains.
• To enhance the performance and interpretability of LDCT reconstruction compared to existing methods.

Main Methods:

• A novel full-domain reconstruction model incorporating noise-generating and imaging mechanisms was proposed.
• An optimization algorithm based on proximal gradient techniques was developed to solve the model.
• The optimization algorithm was unrolled into a deep network, implicitly learning proximal operators for sinogram and image regularizers via two deep neural networks.

Main Results:

• The proposed method demonstrated significant improvements over state-of-the-art LDCT techniques.
• Achieved > 2.9 dB increase in peak signal-to-noise ratio (PSNR).
• Showcased > 1.4% promotion in structural similarity index measure (SSIM) and > 9 HU decrement in root mean square error (RMSE).

Conclusions:

• The developed full-domain model effectively addresses the noise characteristics in LDCT reconstruction.
• The unrolled deep network provides an interpretable and effective approach for LDCT image reconstruction.
• The proposed method offers superior performance in quantitative metrics, advancing the field of low-dose CT imaging.