NOISE REMOVAL FOR MEDICAL X-RAY IMAGES IN MULTIWAVELET DOMAIN

Jianming Lu, Ling Wang, Yeqiu Li, Takashi YahagiInternational Journal of Image and Graphics 2008 · DOI

TL;DR

This paper introduces a novel "covariance shrink" (CS) method for denoising medical X-ray images using multiwavelets. By adaptively calculating thresholds for different subbands of the multiwavelet transform, the CS method significantly outperforms traditional scalar wavelet and existing multiwavelet denoising techniques, leading to clearer medical images.

What Problem Does This Paper Address?

Medical X-ray images are often plagued by noise, primarily "quantum mottle," which arises from the inherent statistical nature of X-ray detection (modeled as a Poisson process, not Gaussian noise). This noise degrades image quality, making it harder for radiologists to accurately diagnose conditions. While wavelet-based denoising has been a popular approach, traditional scalar wavelets have limitations in time-frequency localization. Multiwavelets offer potential improvements due to their ability to simultaneously achieve orthogonality, symmetry, and short support. However, existing multiwavelet denoising methods, particularly those relying on univariate thresholding, struggle to fully exploit the correlated nature of multiwavelet coefficients, leading to suboptimal noise reduction. This paper aims to bridge this gap by developing a more effective multiwavelet denoising strategy tailored for the specific noise characteristics of medical X-ray images.

Methods

The proposed system follows a standard "transform-denoise-inverse transform" pipeline. The core innovation lies in the Covariance Shrink (CS) method for thresholding wavelet coefficients.

  1. Multiwavelet Decomposition: The noisy X-ray image is first prefiltered and then decomposed into multiple subbands using a 3-level orthogonal periodic multiwavelet transform. This process generates approximation coefficients and detail coefficients (horizontal, vertical, and diagonal) at different resolution levels.
  2. Covariance Shrink (CS) Thresholding: This is the key novel component. Instead of applying a single threshold across all subbands or using simple univariate methods, the CS method:
  3. Subband Adaptive Thresholding: It computes a separate, adaptive threshold for each detail subband.
  4. Bivariate Soft Thresholding: It utilizes a bivariate soft thresholding rule, which accounts for the correlation between coefficients within a subband.
  5. Covariance Estimation: The method explicitly estimates the covariance matrix of the noise within each subband. This estimation is crucial for determining the appropriate threshold.
  6. Threshold Formulation: The thresholds are carefully formulated based on the average coefficient magnitudes within each subband and the estimated noise variance. The paper provides specific equations for calculating these thresholds, which depend on the decomposition level, subband direction (horizontal, vertical, diagonal), and the noise level.
  7. Multiwavelet Reconstruction: After thresholding the detail coefficients, the image is reconstructed using the inverse multiwavelet transform.

The paper also briefly discusses traditional multiwavelet methods (MWV) and BayesShrink (BS) as comparison points, noting that MWV often uses scalar hard thresholding, and BS uses univariate soft thresholding.

Datasets

The paper evaluates its proposed method on "medical X-ray images with different doses." While specific public datasets are not explicitly named, the experiments use a "Spinel" X-ray image, which is presented in various figures and tables to demonstrate the denoising performance. The quality of the denoised images is assessed using metrics like Peak Signal-to-Noise Ratio (PSNR) and Mean-to-Standard-Deviation Ratio (MSR) calculated within a Region of Interest (ROI) specified by a doctor.

Key Insights and Ideas

  • Poisson Noise Matters: The paper correctly identifies that medical X-ray noise is often Poisson-distributed, not Gaussian, which necessitates specialized denoising approaches.
  • Multiwavelets Offer Advantages: Multiwavelets, with their desirable properties like orthogonality and symmetry, provide a more powerful framework for signal decomposition compared to scalar wavelets.
  • Subband Adaptivity is Crucial: Applying different thresholds to different subbands of the multiwavelet decomposition is key to achieving better denoising results, as noise and signal characteristics vary across these subbands.
  • Exploiting Coefficient Correlation: The novel CS method's strength lies in its ability to account for the inherent correlations between multiwavelet coefficients within a subband, leading to more effective noise suppression than univariate methods.
  • Covariance Estimation is Key: Accurately estimating the noise covariance matrix within each subband is fundamental to the success of the CS thresholding strategy.

Key Results

The experimental results demonstrate that the proposed Covariance Shrink (CS) method significantly outperforms traditional multiwavelet denoising (MWV) and BayesShrink (BS) in terms of both objective metrics and visual quality.

  • Quantitative Improvements: Tables I and II show that the proposed method consistently achieves higher PSNR and MSR values across different noise levels (doses) compared to BS and MWV. For instance, in Table I, the proposed method's PSNR is 22.89 dB for a specific dose, while BS is 22.52 dB and MWV is 22.10 dB.
  • Qualitative Improvements: Figures 3 and 4 visually confirm the superiority of the CS method. The denoised images produced by the proposed method are noticeably cleaner, with better preservation of edges and image details, while effectively suppressing noise. The paper states that the proposed method "gives better results than conventional methods" and "achieves the highest performance."

Limitations and Future Work

  • Limitations: The paper acknowledges that while the CS method shows improved results, finding "more accommodating convergent values for wavelet coefficients" remains an area for future exploration. The current threshold formulation might not be universally optimal for all types of noise or signals.
  • Future Work: The authors suggest future research directions including:
  • Finding more optimal thresholding parameters.
  • Generalizing the CS method to other types of wavelets beyond multiwavelets.
  • Extending the approach to handle different kinds of noise beyond the Poisson model.

Summary updated

Abstract

When a signal is embedded in an additive Gaussian noise, its estimation is often done by finding a wavelet basis that concentrates the signal energy over few coefficients and by thresholding the noisy coefficients. However, in many practical problems such as medical X-ray image, astronomical and low-light images, the recorded data is not modeled by Gaussian noise but as the realization of a Poisson process. Multiwavelet is a new development to the body of wavelet theory. Multiwavelet simultaneously offers orthogonality, symmetry and short support which are not possible in scalar 2-channel wavelet systems. After reviewing this recently developed theory, a new theory and algorithm for denoising medical X-ray images using multiwavelet multiple resolution analysis (MRA) are presented and investigated in this paper. The proposed covariance shrink (CS) method is used to threshold wavelet coefficients. The form of thresholds is carefully formulated which is the key to more excellent results obtained in the extensive numerical simulations of medical image denoising compared to conventional methods.
BibTeX
@article{lu2008noise,
  author = {Jianming Lu and Ling Wang and Yeqiu Li and Takashi Yahagi},
  title = {{NOISE REMOVAL FOR MEDICAL X-RAY IMAGES IN MULTIWAVELET DOMAIN}},
  journal = {International Journal of Image and Graphics},
  year = {2008},
  doi = {10.1142/s0219467808002952},
}
Figure