Deep Residual Learning for Compressed Sensing CT Reconstruction via Persistent Homology Analysis (1611.06391v2)

Abstract: Recently, compressed sensing (CS) computed tomography (CT) using sparse projection views has been extensively investigated to reduce the potential risk of radiation to patient. However, due to the insufficient number of projection views, an analytic reconstruction approach results in severe streaking artifacts and CS-based iterative approach is computationally very expensive. To address this issue, here we propose a novel deep residual learning approach for sparse view CT reconstruction. Specifically, based on a novel persistent homology analysis showing that the manifold of streaking artifacts is topologically simpler than original ones, a deep residual learning architecture that estimates the streaking artifacts is developed. Once a streaking artifact image is estimated, an artifact-free image can be obtained by subtracting the streaking artifacts from the input image. Using extensive experiments with real patient data set, we confirm that the proposed residual learning provides significantly better image reconstruction performance with several orders of magnitude faster computational speed.

• The paper introduces a novel deep residual learning approach that efficiently removes streaking artifacts in sparse-view CT reconstruction.
• It leverages persistent homology analysis to reveal a simpler topological structure in the artifact manifold, enhancing learning efficiency.
• Experimental validation shows significantly improved image quality and speed over traditional compressed sensing methods.

Deep Residual Learning for Compressed Sensing CT Reconstruction

The paper presents a novel deep residual learning approach for sparse-view computed tomography (CT) reconstruction, specifically utilizing persistent homology analysis. The primary focus is on addressing the limitations of traditional compressed sensing (CS) methods, which are computationally demanding, and the analytical reconstruction approaches, which often produce severe streaking artifacts due to insufficient projection views.

The authors propose a deep residual learning architecture tailored to estimate and subsequently remove these streaking artifacts. The foundation of this approach is a persistent homology analysis that demonstrates the streaking artifacts form a topologically simpler manifold compared to the original artifact-free images. This simplicity indicates that learning the artifact structure could be more efficient than reconstructing the images directly.

Key Methods and Findings

• Novel Architecture: The authors develop a unique deep residual learning architecture that incorporates U-net-like multi-scale deconvolution networks. This architecture excels in capturing globally distributed artifact patterns due to its enlarged receptive field.
• Persistent Homology Application: Persistent homology is employed to theoretically ground the residual learning concept. The analysis reveals that the residual manifold, comprising the streaking artifacts, is simpler in topological terms than the original image manifold.
• Performance and Efficiency: The residual network significantly outperforms traditional CS-based CT methods in both image quality and computational speed, handling sparse-view scenarios with remarkable efficiency. Specifically, the proposed method operates several orders of magnitude faster while maintaining enhanced image reconstruction performance.
• Experimental Validation: Using real patient data, the proposed method is tested and validated. The results show successful streaking artifact removal across varying levels of view sparsity, confirming the universality and robustness of the deep residual learning architecture.

Implications

This work extends the application of deep learning in medical imaging beyond diagnostic tasks to reconstruction challenges, illustrating both practical improvements in computational efficiency and theoretical advancements in understanding manifold complexities through topological methodologies. The approach paves the way for future explorations in artifact removal and image reconstruction in various imaging modalities, potentially transforming CS methods in low-dose CT applications.

Speculations and Future Directions

• Broader Applications: While primarily targeted at CT reconstruction, the proposed methodology may find applications in denoising and artifact removal in other imaging domains, such as MRI or PET, benefiting from similar streaking or globally distributed artifact patterns.
• Further Topological Exploration: Persistent homology could be further exploited to explore and understand other complex artifact structures in various medical imaging contexts, potentially leading to new insights and methods.
• Integration with Other Deep Learning Techniques: Future work could explore integration with other advanced deep learning techniques, such as generative adversarial networks (GANs) or transformers, to further enhance the reconstruction quality and model efficiency.

In conclusion, the paper offers a comprehensive exploration of deep residual learning for CT reconstruction, effectively combining computational topology with deep learning to address practical imaging challenges, and setting a clear path for future research and development in the field.