[Paper Review] Momentum-Net for Low-Dose CT Image Reconstruction

Siqi Ye, Yong Long+1 more|arXiv (Cornell University)|Feb 27, 2020

Medical Imaging Techniques and Applications23 references4 citations

TL;DR

This paper proposes Momentum-Net, a fast iterative neural network architecture for low-dose CT image reconstruction that combines model-based image reconstruction (MBIR) with a four-layer residual CNN for image refinement at each layer. The method significantly improves reconstruction accuracy over state-of-the-art noniterative denoising networks like WavResNet, achieving lower RMSE and enhanced visual quality, while spectral normalization techniques do not improve performance.

ABSTRACT

This paper applies the recent fast iterative neural network framework, Momentum-Net, using appropriate models to low-dose X-ray computed tomography (LDCT) image reconstruction. At each layer of the proposed Momentum-Net, the model-based image reconstruction module solves the majorized penalized weighted least-square problem, and the image refining module uses a four-layer convolutional neural network (CNN). Experimental results with the NIH AAPM-Mayo Clinic Low Dose CT Grand Challenge dataset show that the proposed Momentum-Net architecture significantly improves image reconstruction accuracy, compared to a state-of-the-art noniterative image denoising deep neural network (NN), WavResNet (in LDCT). We also investigated the spectral normalization technique that applies to image refining NN learning to satisfy the nonexpansive NN property; however, experimental results show that this does not improve the image reconstruction performance of Momentum-Net.

Motivation & Objective

• To address the challenge of low-dose CT image reconstruction with reduced radiation exposure while maintaining high image quality.
• To improve upon existing noniterative denoising deep neural networks like WavResNet, which suffer from limited generalization and high overfitting risks due to scarce high-dose reference data.
• To accelerate convergence of model-based image reconstruction (MBIR) by integrating a fast iterative neural network framework, Momentum-Net.
• To investigate whether spectral normalization techniques can enhance the performance of image refining neural networks in the Momentum-Net framework.

Proposed method

• Momentum-Net employs a layered architecture where each layer integrates three components: image refining via a learned residual CNN, extrapolation for acceleration, and a majorized penalized weighted least-squares (PWLS) MBIR module.
• The image refining module uses a four-layer residual CNN, SimpleCNN, defined as $\mathcal{D}_{\theta^{(l+1)}}(\mathbf{x}^{(l)}) = \mathbf{x}^{(l)} - \mathcal{R}_{\theta^{(l+1)}}(\mathbf{x}^{(l)})$, with parameters initialized from previous layers.
• The MBIR module solves a majorized penalized weighted least-squares problem using closed-form solutions, enabling faster convergence compared to inner-iterative methods.
• Spectral normalization (RSN) is applied to the residual mapping and denoiser to enforce nonexpansive properties, with two variants: SimpleCNN-RSN and Dn-RSN.
• Training uses mean-square error (MSE) loss between denoised outputs and reference images, with Adam optimization and learning rate decay.
• The framework is evaluated on the NIH AAPM-Mayo Clinic Low Dose CT Grand Challenge dataset using RMSE and visual assessment.

Experimental results

Research questions

• RQ1 Can Momentum-Net significantly improve low-dose CT reconstruction accuracy compared to state-of-the-art noniterative denoising networks like WavResNet?
• RQ2 Does applying spectral normalization to the image refining neural network enhance the performance of Momentum-Net in low-dose CT reconstruction?
• RQ3 Is the nonexpansive property enforced via spectral normalization beneficial for image reconstruction stability and accuracy in the Momentum-Net framework?
• RQ4 How does the iterative integration of MBIR and deep learning modules affect convergence speed and reconstruction quality compared to noniterative methods?
• RQ5 What is the impact of different training strategies for the residual denoising network on reconstruction performance?

Key findings

• Momentum-Net with SimpleCNN reduces mean RMSE to 3.55 HU, representing a 4.5 HU improvement over WavResNet (36.10 HU) and a 114.3 HU reduction compared to FBP (117.85 HU).
• The proposed Momentum-Net achieves lower standard deviation in RMSE (3.55 HU) than WavResNet (6.46 HU), indicating improved reconstruction stability.
• Visual results show that Momentum-Net with SimpleCNN better preserves fine details and reduces streak artifacts compared to WavResNet, particularly in zoomed regions of the images.
• Momentum-Net with SimpleCNN-RSN achieves a mean RMSE of 36.10 HU, which is higher than WavResNet (6.46 HU), indicating no performance gain from RSN-based training.
• The Dn-RSN variant, which removes skip connections and applies RSN to the full denoiser, yields the worst performance with 41.72 HU RMSE, suggesting that RSN does not improve reconstruction quality and may degrade it.
• Spectral normalization techniques do not improve reconstruction accuracy in Momentum-Net, despite their theoretical appeal for enforcing nonexpansive mappings.