Introduction
Researchers from the University of Canterbury and Massey University have introduced a deep-learning approach to a classic problem in signal processing: cleaning up "messy" signals without losing valuable information. Their framework selectively removes second- and third-order nonlinearities from digital signals, such as those used in Optical Coherence Tomography (OCT). By utilizing a novel "amplitude stack" as input, the system can sharpen blurred images and correct distortions while allowing scientists to isolate specific nonlinear effects for better material analysis.
The Problem: Nonlinearities in Signal Processing
Many high-end sensing and imaging technologies—including MRI, Fourier-transform spectrometry, and OCT—acquire raw signals containing sine-like modulations. These are converted into images via a Fourier Transform. In an ideal scenario, these modulations are linear. However, hardware limitations and material properties introduce nonlinearities.
As shown in Figure 1, these manifest in two primary ways:
- Second-order nonlinearities: Broaden signal peaks, causing a loss of resolution (blurring).
- Third-order nonlinearities: Distort peaks, creating asymmetrical "tails" that degrade image quality.
Traditional linearization methods are often rigid and remove all nonlinearities simultaneously. The authors argue that selective removal is a powerful tool for material analysis, as specific nonlinear signatures can serve as biomarkers or chemical signatures.
Methodology: The Amplitude Stack and U-Net
The core innovation lies in the "Amplitude Stack" input and a modified U-Net architecture.
The Amplitude Stack
Instead of feeding a 1D signal directly into the network, the researchers multiply the raw signal by 32 different complex exponent "correction factors." Performing a Fourier Transform on each creates a 32x1024 2D "stack." Within this stack, different regions are "perfectly corrected" at different rows.
Architecture and Training
The team utilized a modified U-Net, known for preserving fine-grained details in medical imaging. They trained two distinct networks:
- Network #1: Removes second-order nonlinearity to sharpen peaks.
- Network #2: Removes third-order nonlinearity to eliminate distortions.
The models were trained on 200,000 synthetic signals to establish "ground truth" and then validated on real-world OCT data from objects like mirrors, sapphire glass, and fruit.
Key Results and Insights
The neural network learns to identify "sweet spots" in the amplitude stack where the signal is naturally compensated, "stitching" these regions together for a clean output.
- High Precision: Synthetic tests yielded a 99.98% mean Goodness-of-Fit (GoF).
- Real-world Performance: In tests on sapphire glass and fruit, Network #1 significantly reduced peak width, effectively de-blurring the images.
- Efficiency: Processing takes approximately 39 ms on a modern GPU (Quadro RTX 6000).
Limitations and Future Work
While promising, the 39 ms processing time remains a bottleneck for real-time 3D imaging, which requires processing thousands of signals per image. Additionally, the networks occasionally produce tiny non-zero values in empty background areas. Future work aims to create a sequential pipeline to automatically remove all nonlinearity orders.
Reference: Maliszewski, K. A., et al. (2024). Selective nonlinearities removal from digital signals. arXiv:2403.09731v1 [eess.IV].
Summary updated