LUCYD: A Feature-Driven Richardson-Lucy Deconvolution Network

Tomás Chobola, Gesine Müller, Veit Dausmann, Anton Theileis, Jan Taucher, Jan Huisken, Tingying Peng

Introduction

In the life sciences, microscopy often produces images marred by blur and noise, forcing a choice between classical, interpretable algorithms and massive, data-hungry "black-box" AI models. Traditional methods like the Richardson-Lucy (RL) algorithm are mathematically grounded but struggle with noise and high computational costs in 3D. Conversely, deep learning models like CARE (Content-Aware Image Restoration) provide high-quality results but lack physical grounding and require extensive training data.

LUCYD (Feature-Driven Richardson-Lucy Deconvolution Network) bridges this gap. By embedding the physics of image formation directly into a lightweight neural network, it achieves state-of-the-art restoration with significantly fewer parameters and greater robustness.

Methodology

The LUCYD architecture integrates the logic of the RL algorithm into a hybrid deep learning framework. It consists of three primary components:

  1. Correction Module: A U-shaped structure that extracts multi-scale features to generate an initial sharp estimate mask.
  2. Update Module: This module mimics the iterative steps of the RL formula. Crucially, it replaces fixed Point Spread Functions (PSFs) with learnable convolutional layers (forward and backward projectors), allowing the network to adapt to the specific blurring characteristics of the imaging system.
  3. Feature Fusion Blocks (FFBlocks): These facilitate cross-scale information exchange, ensuring the preservation of both fine details and global structures.
The LUCYD architecture, featuring Correction and Update modules integrated with Feature Fusion Blocks.

By performing a single forward pass, LUCYD effectively replaces the multiple iterations required by traditional RL, drastically reducing computation time while maintaining physical constraints.

Performance and Efficiency

A standout feature of LUCYD is its extreme efficiency. While standard models like CARE utilize approximately 1,000,000 parameters, LUCYD achieves superior results with only 24,964 parameters—a 40x reduction. This compactness makes it highly suitable for resource-constrained environments and real-time applications.

The model was evaluated on simulated phantoms, biological simulations (nuclei and membranes), and real-world 3D light-sheet and widefield microscopy. Across these datasets, LUCYD consistently outperformed both CARE and previous RL-based networks (RLN) in terms of Structural Similarity (SSIM) and Peak Signal-to-Noise Ratio (PSNR).

Visual comparison of image restoration: (a) input blurry image y, (b-d) intermediate and final outputs, and (e) ground truth x.

Robustness and Generalization

Because LUCYD is "physics-informed," it generalizes better than purely data-driven models. It maintains high restoration quality even when noise levels exceed those seen during training—a common failure point for standard deep learning. In real-world applications, such as imaging starfish embryos or human bone cancer cells, LUCYD effectively suppressed background haze and recovered fine structural details that competing methods missed.

Conclusion

LUCYD represents a significant step forward in microscopy image restoration. By combining the interpretability of classical physics with the power of deep learning, it offers a fast, robust, and incredibly efficient solution. Future work may focus on handling non-uniform PSFs and developing unsupervised training variants to further broaden its applicability in biological research.

Check out the code: The authors have made LUCYD open-source on GitHub.

Summary updated

BibTeX
@inproceedings{chobola2023lucyd,
  author = {Tomás Chobola and Gesine Müller and Veit Dausmann and Anton Theileis and Jan Taucher and Jan Huisken and Tingying Peng},
  title = {{LUCYD: A Feature-Driven Richardson-Lucy Deconvolution Network}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2023},
  doi = {10.1007/978-3-031-43993-3\_63},
}

Loading related content...

Figure