Efficient 4D Non-local Tensor Total-Variation for Low-Dose CT Perfusion Deconvolution

Ruogu Fang, Ming Ni, Junzhou Huang, Qianmu Li, Tao LiMCV @ MICCAI 2015 · DOI

TL;DR

This paper introduces a novel "non-local tensor total variation" (NL-TTV) deconvolution method to improve the accuracy of low-dose CT perfusion (CTP) imaging. By considering long-range anatomical similarities in the brain, NL-TTV significantly enhances the preservation of delicate structures, leading to more reliable diagnosis of cerebrovascular diseases compared to existing methods.

What Problem Does This Paper Address?

Cerebrovascular diseases like stroke are a major cause of disability, making timely and accurate diagnosis crucial. CT perfusion (CTP) is a vital imaging technique for this, but high radiation doses pose health risks. Reducing radiation dose in CTP leads to noisy images, making it difficult to accurately estimate hemodynamic parameters. While "tensor total variation" (TTV) deconvolution has shown promise in improving low-dose CTP by leveraging anatomical structure and temporal blood flow, its "local" nature limits its ability to capture long-range correlations. This locality can lead to the absence or artifacts in delicate structures like capillaries and the insula, which are critical for diagnosing cerebrovascular diseases. This paper aims to bridge this gap by introducing a "non-local" approach to TTV deconvolution.

Methods

The core innovation is the development of a Non-Local Tensor Total Variation (NL-TTV) deconvolution method. Unlike traditional TTV which only considers adjacent voxels (local neighborhood), NL-TTV extends the regularization to search for similar patterns across the entire brain volume, including distant regions. This is achieved by:

  • Non-local Regularization: Instead of just looking at 4-connected neighbors, NL-TTV searches for similar patches within a defined window around each voxel. The similarity is quantified using an exponential function of the difference between patches. This allows the method to leverage global anatomical symmetry and structural similarity.
  • Spatio-temporal Integration: The NL-TTV regularization is applied to the 4D spatio-temporal flow-scaled residue impulse functions, meaning it considers similarities across both space and time.
  • Efficient Optimization: To address the increased computational cost of non-local methods, the authors propose an efficient optimization algorithm. This involves a combination of steepest gradient descent and an accelerated update for the NL-TTV term, reducing the number of NL-TTV updates per iteration while maintaining accuracy.
  • Parallel Computing: The algorithm's design also lends itself to parallelization, allowing for faster processing on multi-core processors.

Datasets

The paper evaluates its method using two types of data:

  • Simulated Low-Dose CTP Data: This data was generated by adding correlated Gaussian noise to high-dose CTP data. This is a common practice for evaluating CT algorithms when ethical concerns or practical limitations prevent repeated scans on the same subjects.
  • Clinical Dataset: The method was also tested on a clinical dataset of 10 subjects admitted to Weill Cornell Medical College. These subjects included individuals with brain deficits (due to aneurysmal subarachnoid hemorrhage or ischemic stroke) and normal subjects.

Key Insights and Ideas

  • Long-Range Correlations are Crucial: The most significant insight is that capturing long-range anatomical similarities in the brain is essential for accurate low-dose CTP deconvolution. Neglecting these global connections leads to the loss of fine details critical for diagnosis.
  • Preserving Delicate Structures: NL-TTV's ability to preserve small vessels and capillaries without dilation or rupture is a major advantage over local methods, directly impacting diagnostic confidence.
  • Balancing Accuracy and Efficiency: The paper demonstrates that it's possible to achieve improved accuracy with non-local methods without an insurmountable computational cost, thanks to efficient optimization and parallelization strategies.

Key Results

  • Improved Quantitative Accuracy: The proposed NL-TTV method significantly outperforms existing methods (sSVD, bSVD, Tikhonov, and local TTV) in terms of quantitative accuracy, as measured by PSNR and SSIM. The paper states these improvements are statistically significant (p < 0.05).
  • Enhanced Preservation of Fine Structures: Qualitative results show that NL-TTV better preserves delicate anatomical structures, such as capillaries, compared to local TTV, which can lead to blurring or dilation.
  • Efficient Convergence: The algorithm converges within 10 iterations.
  • Reasonable Running Time: After accelerated optimization, the entire CTP data for one subject can be processed in approximately 30 minutes on a single PC.

Limitations and Future Work

  • Computational Cost: While the proposed optimization significantly reduces the time complexity compared to brute-force non-local methods, it is still higher than local TTV. The paper notes that further optimization and parallel computing on clusters could further speed up the experiments.
  • Generalizability: The paper focuses on brain CTP. Future work could explore the applicability of NL-TTV to other imaging modalities or anatomical regions.
  • Clinical Validation: While clinical data was used, further extensive clinical validation and integration into clinical workflows would be beneficial.

Summary updated

BibTeX
@inproceedings{fang2015efficient,
  author = {Ruogu Fang and Ming Ni and Junzhou Huang and Qianmu Li and Tao Li},
  title = {{Efficient 4D Non-local Tensor Total-Variation for Low-Dose CT Perfusion Deconvolution}},
  booktitle = {MCV @ MICCAI},
  year = {2015},
  doi = {10.1007/978-3-319-42016-5\_16},
}
Figure