[Paper Review] Deep Neural Network Assisted Iterative Reconstruction Method for Low Dose CT

Shabab Bazrafkan, Vincent Van Nieuwenhove+3 more|arXiv (Cornell University)|Jun 3, 2019

Medical Imaging Techniques and Applications4 citations

TL;DR

This paper proposes a hybrid deep learning and iterative reconstruction method for low-dose CT that integrates a Deep Neural Network (DNN) as a regularization term within the SIRT (Singular Value Decomposition-based Iterative Reconstruction) framework. By applying ten Mixed-Scale Dense Convolutional Networks sequentially after each SIRT step, the method improves image quality and enhances fidelity to the original sinogram data, outperforming state-of-the-art methods in PSNR, MSE, and SSIM metrics.

ABSTRACT

Low Dose Computed Tomography suffers from a high amount of noise and/or undersampling artefacts in the reconstructed image. In the current article, a Deep Learning technique is exploited as a regularization term for the iterative reconstruction method SIRT. While SIRT minimizes the error in the sinogram space, the proposed regularization model additionally steers intermediate SIRT reconstructions towards the desired output. Extensive evaluations demonstrate the superior outcomes of the proposed method compared to the state of the art techniques. Comparing the forward projection of the reconstructed image with the original signal shows a higher fidelity to the sinogram space for the current approach amongst other learning based methods.

Motivation & Objective

• To address noise and streaking artifacts in low-dose CT caused by reduced radiation or few projections.
• To improve image reconstruction fidelity by combining the data consistency of SIRT with the structural prior learning of DNNs.
• To enhance reconstruction accuracy in both image space and sinogram space, overcoming limitations of purely learning-based methods.
• To investigate the cooperative behavior between iterative reconstruction and deep learning in multi-stage processing.
• To demonstrate superior performance over state-of-the-art methods in quantitative metrics like PSNR, MSE, and SSIM.

Proposed method

• The method combines SIRT, an iterative reconstruction algorithm that minimizes error in sinogram space, with a sequence of Deep Neural Networks (DNNs) for image space regularization.
• Ten Mixed-Scale Dense Convolutional Neural Networks are trained consecutively, one after each SIRT iteration, to refine intermediate reconstructions.
• Transfer learning is employed: each subsequent DNN is initialized using the best-performing model from the prior step, improving training stability and convergence.
• The DNNs are trained to map noisy or low-quality SIRT outputs to higher-quality images, learning structural priors without direct supervision on sinogram data.
• The final output is obtained after a final SIRT step applied to the DNN output, ensuring data consistency in sinogram space.
• The framework is evaluated using PSNR, MSE, and SSIM on test data, with intermediate results analyzed after each SIRT and DNN block.

Experimental results

Research questions

• RQ1 Can a deep neural network effectively regularize iterative reconstruction to reduce noise and streaking in low-dose CT?
• RQ2 How does the sequential integration of DNNs after each SIRT step affect image quality and sinogram fidelity compared to end-to-end learning?
• RQ3 To what extent do early-stage DNNs contribute to the overall performance improvement in the reconstruction pipeline?
• RQ4 Does the proposed method achieve better reconstruction fidelity in sinogram space than existing learning-based approaches that operate solely in image space?
• RQ5 How does the method perform under varying noise levels and background intensities, especially in high-intensity regions?

Key findings

• The proposed SIRT+DNN method achieves superior PSNR, MSE, and SSIM values compared to state-of-the-art learning-based and model-based reconstruction techniques.
• The method maintains higher fidelity to the original sinogram data after forward projection, outperforming other learning-based methods that operate only in image space.
• Intermediate results show that the first few DNN blocks have the most significant impact on image quality, with marginal improvements in later stages.
• The method produces more detailed reconstructions, particularly in regions with background intensities above 10,000, where it outperforms competing models.
• The cooperative behavior between SIRT and DNN is evident: each SIRT step improves data consistency, and each DNN step enhances structural and pixel-level accuracy.
• Despite being slower than some alternatives, the method's performance and robustness suggest it is viable with modern GPU acceleration, especially with future training on noisy data for improved generalization.