TL;DR
This paper introduces the Bayesian Residual Transform (BRT), a novel Bayesian-based method for multi-scale signal decomposition. BRT decomposes signals into a sum of residual signals, each representing information at a different scale, offering a new approach to physiological signal processing, particularly for noise suppression.
What Problem Does This Paper Address?
Physiological signals, like ECGs, are often corrupted by noise, making them difficult to process and interpret. Traditional methods like linear filters can degrade signal quality. While multi-scale decomposition techniques like scale-space theory and wavelet transforms have been successful, they have limitations. Scale-space theory can struggle with fine-grained noise suppression, and wavelet transforms can introduce oscillation artifacts. This paper aims to address these limitations by proposing a new Bayesian-based multi-scale decomposition method that offers a different perspective on signal analysis and noise suppression.
Methods
The core of the paper is the Bayesian Residual Transform (BRT). BRT models a signal $f(t)$ as the sum of $n$ residual signals, $r_i(t)$, each capturing information at a specific scale: $f(t) = \sum_{i=1}^{n} r_i(t)$. The key idea is that residual signals at lower scales capture fine-grained details, while those at higher scales capture coarser features.
The BRT is realized through a deep cascading framework. This means that the residual signal at scale $j$, $r_j(t)$, is computed based on the signal representation at the previous scale, $f_{\Sigma,j}(t)$, and the expected signal representation at the next scale, $f_{\Sigma,j+1}(t)$, where $f_{\Sigma,j+1}(t) = E(f_{\Sigma,j+1}(t)|f_{\Sigma,j}(t))$. Specifically, $r_j(t) = f_{\Sigma,j}(t) - f_{\Sigma,j+1}(t)$. This process is iterated, with the final residual $r_n(t)$ being the coarsest representation.
For practical implementation, the paper uses kernel regression (specifically, Nadaraya-Watson kernel regression with a Gaussian kernel) to estimate the conditional expectations.
For noise suppression, a simple thresholding strategy is employed. After decomposing the noisy signal using BRT into residual signals $r_i(t)$, a noise threshold $\theta_j$ is estimated for each scale $j$ using the median absolute deviation (MAD) method. Residual signals are then thresholded: if the absolute value of a residual signal at a specific location is below the threshold, it's set to zero; otherwise, it's kept. Finally, the noise-suppressed residual signals are summed to reconstruct the denoised signal.
Datasets
The primary dataset used for evaluation is the MIT-BIH Normal Sinus Rhythm Database. This database contains ECG recordings. The authors extracted 18 low-noise segments and then artificially contaminated them with white Gaussian noise to create noisy signals with various Signal-to-Noise Ratios (SNRs) ranging from 12 dB to 2.5 dB. This allowed for a quantitative evaluation of noise suppression performance.
Key Insights and Ideas
- Bayesian Perspective on Multi-scale Decomposition: Moving beyond traditional scale-space and wavelet approaches, BRT frames signal decomposition through a Bayesian lens, modeling signals as a sum of statistically characterized residual components at different scales.
- Cascading Framework for Tractability: The deep cascading framework makes the BRT computationally tractable by computing residuals sequentially, leveraging previous scale computations.
- Decoupling Noise and Signal: The BRT's multi-scale nature allows for the separation of noise characteristics (often found at finer scales) from the underlying signal's structural information (more prominent at coarser scales). This is crucial for effective noise suppression without sacrificing signal integrity.
- Adaptive Noise Thresholding: The use of Donoho's noise level estimation method (MAD-based) provides an adaptive threshold for each scale, allowing for more precise noise removal.
Key Results
- Feasibility Demonstrated: The study successfully demonstrates the feasibility of using BRT for physiological signal processing, specifically for noise suppression.
- Competitive Noise Suppression: In quantitative SNR analysis on ECG signals, the BRT-based noise suppression method achieved strong SNR improvements across various input SNRs, performing comparably to the SURE wavelet denoising method and outperforming other tested wavelet methods (HSURE, UNI, MINIMAX).
- Effect of Parameters:
- The number of scales ($n$) significantly impacts SNR improvement, with gains diminishing after $n=3$.
- The standard deviation parameter ($\lambda$) also plays a crucial role, with 1SD generally providing the strongest results for the tested ECG signals.
- Signal Preservation: Visual inspection of the denoised ECG signals shows that the BRT effectively reduces noise artifacts while preserving the essential characteristics of the original signal.
Limitations and Future Work
- Parameter Selection: The performance of BRT is sensitive to parameter choices, particularly the number of scales ($n$) and the standard deviation ($\lambda$). Careful selection is needed for optimal results, and this may vary across different signal types.
- Adaptive Thresholding for Nonstationary Noise: The current noise suppression method uses a simple thresholding strategy. Future work aims to investigate more adaptive thresholding schemes to better handle nonstationary noise, which is common in physiological signals.
- Beyond Noise Suppression: The authors plan to explore BRT for more advanced multi-scale analysis tasks, such as feature extraction for signal classification.
- Extension to High-Dimensional Signals: Future research will focus on extending BRT to handle high-dimensional physiological signals (e.g., VCG) and medical imaging data (e.g., MRI, OCT).
Summary updated