Solving the inverse problem of microscopy deconvolution with a residual Beylkin-Coifman-Rokhlin neural network

1. Problem Statement

Context: Light microscopy (LM) captures biological samples through an optical convolution process, mathematically described as $I(x) = A(PSF) * o(x) + \varepsilon$, where $I$ is the observed image, $o$ is the true object, $A$ is the Point Spread Function (PSF) operator, and $\varepsilon$ represents noise and artifacts.
The Challenge: The goal is optic deconvolution: recovering the ground truth $o(x)$ from the blurred, noisy image $I(x)$.

• Traditional Methods: Explicit methods (e.g., Richardson-Lucy) rely on accurate PSF models. However, they often fail due to inaccurate PSF estimation, noise, and the ill-posed nature of the inverse problem.
• Deep Learning (DL) Limitations: While data-driven DL models (U-Net, DDPM, ESRGAN, etc.) have shown promise in super-resolution and denoising, they often neglect the underlying physical texture of the optical problem. This leads to:
  • Over-parameterization: Models require massive numbers of trainable parameters to learn the physics implicitly.
  • Instability: Without physical constraints, models may learn irrelevant features, leading to artifacts in real-world microscopy data where ground truth is unavailable.
  • Efficiency: Large models result in high computational costs and long runtimes.

2. Methodology

The authors propose a Physics-Informed Neural Network (PINN) called Multi-Stage Residual-BCR Net (m-rBCR). The methodology bridges integral operator theory with deep learning.

A. Theoretical Foundation: BCR Representation

• Inverse Problem Formulation: Deconvolution is treated as solving an inverse problem for an integral operator. The inverse operator $A^{-1}$ is approximated as a pseudo-differential operator.
• Beylkin-Coifman-Rokhlin (BCR) Theory: The authors utilize the BCR wavelet representation to approximate integral and pseudo-differential operators.
  • Standard wavelet representations suffer from sparsity issues and high computational complexity ($O(N \log N)$).
  • BCR introduces a nonstandard form of wavelets that reduces complexity to $O(N)$ by representing operators as sparse band matrices.
  • The inverse operator is decomposed into a recursive chain of local convolution operators and orthogonal matrices.

B. Network Architecture: m-rBCR

The architecture translates the mathematical decomposition of the BCR theory into a neural network structure:

1. Backbone (Pseudo-differential Operator): The core of the network approximates the operator $B = (K^T K + \varepsilon I)^{-1} K^T$, where $K^T$ represents the forward convolution cluster and $B$ represents the pseudo-differential inverse.
2. Residual Dense Structure: To address the strong non-linearity and noise in microscopy data (which the original linear BCR theory cannot handle), the authors integrate Residual Dense Blocks (RDN).
   • This adds non-linearity and robustness, allowing the network to learn complex noise patterns while adhering to the physical decomposition.
3. Multi-Stage Learning:
   • To mitigate accumulated errors from truncating the wavelet decomposition levels (a requirement for computational efficiency), the model employs a multi-stage approach.
   • Inputs from different resolution levels are reintegrated as "posteriors" at each stage. This allows the network to recover truncated signals and refine the restoration iteratively.
4. Key Components:
   • Sequence Convolution Block (SCB): Implements the forward operator $K^T$.
   • Residual Dense Blocks: Implement the pseudo-differential operator $B$.
   • Feature Fusion: Concatenates features from previous stages and current inputs to guide the learning process.

3. Key Contributions

1. Physics-Informed Architecture: The first application of BCR wavelet theory to design a neural network specifically for optic deconvolution, embedding the mathematical structure of the inverse problem directly into the network topology.
2. Efficiency vs. Performance: The model achieves state-of-the-art performance with a drastically reduced parameter count.
   • Parameter Reduction: m-rBCR uses ~30 times fewer parameters than the benchmark MIMO-U-Net and ~210 times fewer than ESRGAN.
   • Speed: It runs ~3 times faster than MIMO-U-Net and up to ~300 times faster than DDPM.
3. Robustness to Noise: By leveraging physical constraints (the inverse operator structure), the model is less prone to generating artifacts compared to "black-box" DL models when applied to real, noisy microscopy data.

4. Experimental Results

The model was validated on four datasets: two simulated (BioSR, ImageNet) and two real-world (dSTORM, Widefield-Confocal).

• Performance Metrics (PSNR & SSIM):
  • Simulated BioSR: m-rBCR achieved the highest PSNR (24.89) and SSIM (0.78), outperforming MIMO-U-Net (PSNR 23.95) and all other DL models.
  • Simulated ImageNet: m-rBCR ranked second-best (PSNR 21.41), closely following MIMO-U-Net (PSNR 22.35) but with significantly higher efficiency.
  • Real dSTORM: m-rBCR achieved the highest PSNR (20.13), outperforming MIMO-U-Net (18.91). It successfully recovered details from extremely weak inputs where other models failed or introduced artifacts.
  • Real Widefield-Confocal: m-rBCR led with a PSNR of 23.10, surpassing DDPM (22.27) and others.
• Qualitative Analysis: Visual inspection confirmed that m-rBCR restored high-frequency details without introducing the noisy background textures or artifacts seen in DDPM or the blurring seen in Richardson-Lucy.
• Hyperparameter Robustness: Experiments showed optimal performance at a decomposition level $K=12$ and residual block depth $RDN=7$.

5. Significance and Conclusion

This paper demonstrates that incorporating physical priors (specifically the BCR wavelet decomposition of integral operators) into neural network design is a superior strategy for microscopy deconvolution compared to purely data-driven approaches.

• Scientific Impact: It provides a mathematically grounded framework for solving ill-posed inverse problems in imaging, moving beyond "black-box" deep learning.
• Practical Impact: The m-rBCR model offers a highly efficient solution for biological research. Its low parameter count and fast runtime make it suitable for real-time processing or deployment on hardware with limited computational resources, while maintaining or exceeding the restoration quality of much larger models.
• Future Work: The authors suggest exploring more advanced wavelet-based fusion strategies to further improve learning performance.

Code Availability: The source code is available at https://github.com/leeroyhannover/m-rBCR.