Multilayer residual sparsifying transform (MARS) model for low-dose CT image reconstruction.

Area of Science:

• Medical Imaging
• Computational Imaging
• Signal Processing

Background:

• Sparse representations and deep neural networks are successful in image reconstruction.
• Deep models have recently been applied to image reconstruction tasks.
• Combining sparse and deep models offers a promising avenue for novel reconstruction approaches.

Purpose of the Study:

• Develop a novel image reconstruction approach using a multilayer model.
• Combine sparse representations and deep models in an unsupervised learning framework.
• Extend classical sparsifying transform models to a Multilayer residual sparsifying transform (MARS) model.

Main Methods:

• Propose new formulations for multilayer transform learning and image reconstruction.
• Utilize an efficient block coordinate descent algorithm for unsupervised transform learning.
• Incorporate the learned MARS model into penalized weighted least squares (PWLS) optimization for low-dose CT reconstruction.

Main Results:

• The MARS model outperforms conventional methods like filtered back-projection and edge-preserving regularized PWLS.
• MARS achieves superior performance in terms of RMSE and SSIM, with improved noise suppression.
• Compared to single-layer learned transform models, MARS better preserves subtle image details.

Conclusions:

• Present a novel data-driven regularization framework for CT image reconstruction using learned multilayer residual sparsifying transforms.
• The image model is learned in an unsupervised manner from limited image data.
• The proposed multilayer MARS scheme shows promising performance and better image quality than single-layer learned transforms and nonadaptive PWLS methods.