Computed Tomography Reconstruction Using Deep Image Prior and Learned Reconstruction Methods (2003.04989v2)

Abstract: In this work, we investigate the application of deep learning methods for computed tomography in the context of having a low-data regime. As motivation, we review some of the existing approaches and obtain quantitative results after training them with different amounts of data. We find that the learned primal-dual has an outstanding performance in terms of reconstruction quality and data efficiency. However, in general, end-to-end learned methods have two issues: a) lack of classical guarantees in inverse problems and b) lack of generalization when not trained with enough data. To overcome these issues, we bring in the deep image prior approach in combination with classical regularization. The proposed methods improve the state-of-the-art results in the low data-regime.

• The paper compares learned primal-dual and Deep Image Prior (DIP) methods for CT reconstruction, finding learned primal-dual efficient with less data and DIP combined with TV effective in data-free settings.
• A novel hybrid approach integrating DIP with initial learned reconstructions is proposed to enhance data consistency while preserving image quality in low-data regimes.
• These findings offer practical implications for reducing scan doses and improving image quality in medical imaging by leveraging both data-driven and structural image priors.

Computed Tomography Reconstruction Using Deep Image Prior and Learned Reconstruction Methods

This paper presents a comprehensive paper on the reconstruction methods employed in computed tomography (CT), specifically targeting low-data regimes. The researchers explore a variety of deep learning techniques, focusing on both existing approaches and innovative combinations thereof. In particular, the work scrutinizes the efficacy of the learned primal-dual method and the Deep Image Prior (DIP) across different dataset sizes with varying levels of noise.

The primary motivation stems from a recognized challenge in medical imaging, where large sets of labeled data are often not available, making traditional data-driven models less viable. This is crucial for modalities like CT, where high-dose scans—necessary for generating ground truth—pose health risks to patients. Consequently, the paper introduces novel techniques that need fewer ground truth samples and attempt to remedy the inherent issues related to model generalization and classical guarantees in inverse problems.

The paper showcases the remarkable efficiency of the learned primal-dual method in terms of data usage, outperforming traditional approaches like Total Variation (TV) regularization with significantly less data. It also highlights the DIP combined with TV, which stands out as the best data-free approach, surpassing simple TV regularization in image quality metrics such as PSNR.

To address data consistency concerns in learned methods, the authors propose a hybrid approach that integrates DIP with initial reconstructions obtained via learned methods. This novel technique refines initial reconstructions by enforcing consistency with the observed data whilst maintaining image naturalness through the DIP framework. These reconstructions benefited from adjustments, showing improved data consistency without introducing artifacts—a common concern when chasing data fidelity too aggressively in ill-posed problems like CT reconstruction.

The implications of these findings are multifold. Practically, they offer promising avenues for reducing the need for high-dose scans while maintaining or improving image quality. However, future developments may focus on better understanding the interaction between network architectures and structural biases, optimizing future models for even less data dependence, or enhanced integration with classical regularization methods to extend these promising results beyond CT to other imaging modalities such as MRI or PET.

The paper contributes substantially to the understanding of sophisticated model interactions in CT reconstruction, setting a strong precedent for future exploration. By combining diverse techniques—leveraging both data-driven processes and structural image biases—the authors pave the way for safer medical imaging practices that yield reliable reconstructions under restrictive data conditions. For experienced researchers in medical imaging, this document provides a solid basis for developing new methods or expanding the scope of existing models.