A Novel Singular Value Decomposition-Based Denoising Method in 4-Dimensional Computed Tomography of the Brain in Stroke Patients with Statistical Evaluation

WonSeok Yang, Jun-Yong Hong, Jeong-Youn Kim, Seung-ho Paik, Seung Hyun Lee, Ji-Su Park, Gihyoun Lee, Beop Min Kim, Young-Jin Jung2020 · DOI

TL;DR

This paper introduces a novel Singular Value Decomposition (SVD)-based denoising technique for 4D Computed Tomography Perfusion (4D-CTP) brain scans. By effectively removing noise without increasing radiation dose, this method promises to improve image quality and diagnostic accuracy for stroke patients, potentially enabling the use of lower radiation settings.

What Problem Does This Paper Address?

Strokes are a major health concern, and timely diagnosis is critical for effective treatment. 4D-CTP is the gold standard for diagnosing hyperacute strokes, but it exposes patients to high radiation doses, posing health risks. While low-dose scanning techniques exist, they often introduce Gaussian noise, degrading image quality and hindering accurate interpretation of brain structures and blood vessels. This degradation can reduce diagnostic success rates. Therefore, there's a pressing need for methods that can improve image quality in 4D-CTP without increasing radiation exposure.

Methods

The core of this research is a denoising technique leveraging Singular Value Decomposition (SVD). The authors adapted SVD, a mathematical tool for matrix decomposition, to process 4D CTP data. The process involves:

  1. Reshaping the 4D CTP data: The 4D data (spatial dimensions x time) is transformed into a 2D spatio-temporal matrix.
  2. SVD Decomposition: The 2D matrix is decomposed into three matrices: U, S (containing singular values), and V.
  3. Noise Component Identification: The singular values in the S matrix are analyzed. The authors empirically identified the second singular value as a thresholding component. Singular values below a certain percentage (5-40%) of this threshold are considered noise.
  4. Reconstruction: A new diagonal matrix (Sth) is created where singular values above the threshold are retained, and those below are set to zero. The denoised image is then reconstructed using U, Sth, and V.
  5. Quantitative Analysis: Signal-to-Noise Ratio (SNR) and Contrast-to-Noise Ratio (CNR) were calculated using MATLAB to objectively assess image quality.
  6. Statistical Analysis: Paired t-tests were performed using SPSS to compare the quantitative scores between original and denoised images, with p-values indicating statistical significance.

Datasets

The study utilized retrospective 4D-CTP dynamic images from 20 stroke-suspected patients collected between January and December 2016 at Kosin University Medical Center. Patients already diagnosed with ischemic or hemorrhagic strokes were excluded. No new benchmark datasets were introduced.

Key Insights and Ideas

  • SVD as a powerful denoising tool for medical imaging: The paper demonstrates that SVD can effectively separate signal from noise in complex 4D medical imaging data, a task that is challenging with traditional methods.
  • The "sweet spot" for denoising: While higher threshold percentages (more singular values removed) generally led to higher SNR and CNR, the 25% threshold setting was identified as potentially optimal. This is because it achieved good statistical significance while still preserving diagnostic information, suggesting a trade-off between noise reduction and image detail.
  • Achieving high quality without high dose: The most significant insight is that this SVD-based denoising can significantly improve image quality to a level comparable to high-dose scans, but using lower radiation doses. This has direct implications for patient safety.
  • The second singular value's importance: The empirical finding that the second singular value is a good indicator for thresholding suggests a specific structural property of noise in this type of data that SVD can exploit.

Key Results

  • Improved SNR and CNR: The SVD-based denoising technique led to significant improvements in both SNR and CNR at the basal ganglia level. For example, a 43% increase in CNR was observed at the 40% threshold setting.
  • Optimal Thresholding: The 25% threshold setting showed the best statistical significance (p < 0.001) for both SNR and CNR, with a 14% increase in SNR and a 37% increase in CNR.
  • Visual Quality: Qualitative comparisons showed that denoised images (using the 25% threshold) were visually comparable to high-radiation dose images, with sharper margins and better differentiation of anatomical structures.
  • No Radiation Increase: The technique successfully improved image quality without requiring an increase in radiation dose.

Limitations and Future Work

The authors acknowledge several limitations:

  • Perfusion Maps Not Analyzed: The SVD technique was not applied to perfusion maps, which are crucial for stroke diagnosis. Future work should investigate its effectiveness on these maps.
  • Small Sample Size: The study used a relatively small sample size (20 patients). A larger cohort is needed to confirm the findings.
  • Quantitative Focus: The study focused on quantitative image quality metrics. Future research should assess the diagnostic accuracy of the denoised images compared to non-SVD images.

Promising future directions include:

  • Investigating the application of SVD to CT perfusion maps.
  • Evaluating the effectiveness of SVD on images acquired with even lower tube current settings.
  • Assessing the diagnostic accuracy of the denoised images.
  • Developing a standardized method for identifying the optimal threshold level.
  • Further studies with artificial biological objects using 3D bioprinting for more accurate evaluation.

Summary updated

BibTeX
@misc{yang2020novel,
  author = {WonSeok Yang and Jun-Yong Hong and Jeong-Youn Kim and Seung-ho Paik and Seung Hyun Lee and Ji-Su Park and Gihyoun Lee and Beop Min Kim and Young-Jin Jung},
  title = {{A Novel Singular Value Decomposition-Based Denoising Method in 4-Dimensional Computed Tomography of the Brain in Stroke Patients with Statistical Evaluation}},
  year = {2020},
  doi = {10.3390/s20113063},
}
Figure