[论文解读] Deep Residual Learning for Compressed Sensing CT Reconstruction via Persistent Homology Analysis

Yoseob Han, Jaejun Yoo+1 more|arXiv (Cornell University)|Nov 19, 2016

Medical Imaging Techniques and Applications参考文献 19被引用 173

一句话总结

The paper proposes a deep residual U-net architecture to remove streaking artifacts in sparse-view CT, showing artifacts reside on a simpler manifold and achieving faster, higher-quality reconstructions than traditional compressed sensing methods.

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.

研究动机与目标

• Motivate reducing radiation dose in CT by using sparse projection views.
• Introduce a residual-learning CNN that estimates streaking artifacts rather than full images.
• Demonstrate, via persistent homology, that the artifact manifold is topologically simpler, aiding learning.
• Show that a multi-scale U-net architecture effectively removes streaks while preserving details.
• Compare performance with traditional CS-CT methods in quality and speed.

提出的方法

• Formulate sparse-view CT reconstruction as artifact removal via residual learning (artifact = input minus full-view image).
• Use a multi-scale U-net–style architecture with contracting path and concatenation to enlarge receptive field.
• Apply residual learning to approximate streaking artifacts rather than the artifact-free image.
• Leverage persistent homology to compare topology of artifact manifold versus full-view image manifold.
• Train with SGD on patch-based training data, using 256x256 patches and a small L2 regularization term.
• Evaluate across varying view counts (48, 64, 96, 192) and compare PSNR to TV-based CS-CT.

实验结果

研究问题

• RQ1 Is the residual (artifact) manifold simpler topologically than the full-image manifold, facilitating learning?
• RQ2 Does multi-scale residual learning outperform single-scale and image-learning approaches for sparse-view CT reconstruction?
• RQ3 Can the proposed method generalize across a range of projection-view counts (e.g., 48–192) with a single trained model?
• RQ4 Does the method offer faster reconstruction while improving image quality compared to TV-based compressed sensing CT?
• RQ5 What is the impact of training data diversity on universal reconstruction performance across view counts?

主要发现

• Persistent homology analysis shows the residual artifact manifold is topologically simpler than the original image manifold.
• Multi-scale residual learning (U-net style) yields better reconstruction than single-scale residual or image-learning baselines.
• The proposed method achieves higher PSNR than alternatives across 48–192 view settings, with significantly faster runtimes.
• Reconstruction speed is about 123 ms per slice, ~30x faster than TV-based CS-CT methods.
• The network trained on mixed sparse-view data generalizes across 48 and 96 views, providing robust performance across a range of sparse-view conditions.
• Quantitative results (PSNR) indicate the proposed method outperforms other architectures at each view count.