Leonardo: A Toolset to Remove Sample-Induced Aberrations in Light Sheet Microscopy Images

Tingying Peng, Yu Liu, G.a. Müller, Lennart Kowitz, Tomáš Chobola, Kurt Weiss, Paul Maier, Jie Luo, Malte Roeßing, M Stenzel, Anika Grüneboom, Johannes C. Paetzold, Ali Ertürk, Nassir Navab, Carsten Marr, Jianxu Chen, Jan HuiskenResearch Square 2025 · DOI

TL;DR

This paper introduces "Leonardo," a novel AI-powered toolbox designed to significantly improve the quality of light sheet microscopy images by removing common sample-induced aberrations. Leonardo tackles issues like stripe artifacts, blur, and ghosting, enabling clearer and more accurate imaging of complex biological samples.

What Problem Does This Paper Address?

Light sheet fluorescence microscopy (LSFM), also known as SPIM, is a powerful technique for imaging biological specimens, especially living ones, due to its gentle illumination and rapid volumetric acquisition. However, when imaging inhomogeneous samples (like tissues with varying densities or refractive indices), several aberrations degrade image quality. These include:

  • Light Absorption: Causes stripe artifacts, appearing as dark shadows, especially when light travels through dense regions.
  • Light Scattering: Widens the light sheet, leading to reduced sectioning quality and blur as the focal plane goes deeper into the sample.
  • Light Refraction: Mismatches in refractive indices can misposition the light sheet, causing optical distortions and "ghost artifacts."
  • Limited Depth of Field: Detection can also suffer from blur and loss of contrast in deeper sample regions.

These issues collectively limit the ability to obtain high-resolution, artifact-free 3D images, especially for larger or more complex specimens. Existing methods often struggle to address all these aberrations simultaneously or can introduce their own artifacts.

Methods

Leonardo is a two-module toolbox:

  • Leonardo-DeStripe: This module addresses stripe artifacts caused by light absorption. It employs a deep learning approach using a graph neural network (GNN) to analyze Fourier space. It identifies and isolates stripe-corrupted Fourier coefficients using a wedge-shaped mask and then reconstructs them by learning relationships with neighboring coefficients. This method is designed to be adaptive to sample-specific stripe characteristics and preserves structural details using anisotropic total variation. It can also handle multi-directional stripes by extending the GNN to process multiple orientations.
  • Leonardo-Fuse: This module tackles blur and ghost artifacts arising from light scattering and refraction. It works by fusing data from dual-sided illumination and/or dual-sided detection. The core idea is to intelligently combine information from different views, prioritizing regions with higher image quality. It estimates "fusion boundaries" that delineate transitions between "good" and "bad" image quality regions, using local saliency levels and prior knowledge of light propagation. This allows it to selectively stitch together the best parts of different datasets, effectively eliminating blur and ghost artifacts. It can handle both illumination-axis fusion and detection-axis fusion.

The toolbox is designed with a moderate parameter size for efficient training and execution, allowing it to run on standard workstations. It also includes a user-friendly Napari-based graphical interface.

Datasets

The paper validates Leonardo on a variety of biological systems, including:

  • Embryonic zebrafish: Both relatively translucent and optically opaque larval zebrafish.
  • Immunolabeled mice: Up to 2 centimeters in size.
  • Murine hind leg and heart tissue: Optically cleared samples.
  • Zebrafish brain tissue: Optically cleared samples.

Specific datasets mentioned for evaluation and comparison include:

  • Synthetic stripes: Simulated on adipose tissue for initial testing of Leonardo-DeStripe.
  • Live zebrafish embryos (2 days post-fertilization): Used for evaluating Leonardo-Fuse.
  • PEGSOS-cleared mouse brain: Labeled with THY1-eGFP, used for simulating light falloff.
  • Transgenic zebrafish (Tg (H2B:GFP)): Used for evaluating Leonardo-Fuse with dual-sided detection.

The paper also provides links to publicly available datasets for researchers to use.

Key Insights and Ideas

  • AI for Aberration Correction: The paper demonstrates the power of AI, specifically GNNs and sophisticated fusion strategies, to overcome fundamental optical limitations in microscopy.
  • Sample-Specific Adaptability: Leonardo-DeStripe's GNN approach allows it to adapt to the unique characteristics of stripes in different samples, a significant improvement over fixed-filter methods.
  • Intelligent Fusion: Leonardo-Fuse's ability to estimate and adapt fusion boundaries based on local image quality and physical properties of the specimen is crucial for handling complex geometries and scattering.
  • Synergistic Workflow: The paper highlights that combining DeStripe and Fuse in a specific order (Fuse first, then DeStripe for large samples) can optimize results and computational efficiency.
  • Broad Applicability: The techniques developed are not limited to SPIM but have potential for other microscopy modalities facing similar stripe or fusion challenges.

Key Results

  • Leonardo-DeStripe:
    • Outperformed the state-of-the-art MDSR method in quantitative evaluations, achieving higher PSNR (35.5 vs. 29.2) and SSIM (0.972 vs. 0.928) on synthetic stripes.
  • Successfully removed stripes in diverse specimens, including zebrafish, murine heart, and brain tissue, even in challenging multi-directional SPIM setups.
  • Significantly improved downstream analysis, such as cell segmentation using Cellpose, compared to raw or MDSR-enhanced images.
  • Leonardo-Fuse:
    • Effectively eliminated blur and ghost artifacts in scattering tissues, producing more contrasted images than existing methods like BigStitcher-Fuse and MST-SR.
  • Demonstrated superior performance in complex geometries, such as the zebrafish larva with an eye shadow, by accurately adapting fusion boundaries.
  • The fused image stack contained up to 1.5 times more information at any depth compared to the best individual input stack.
  • Integrated Workflow: The optimized workflow for large samples (Fuse then DeStripe) reduced processing time by approximately 35% compared to a normal workflow.

Limitations and Future Work

  • Inter-slice Information: The authors suggest that incorporating inter-slice information into Leonardo-DeStripe could further improve stripe removal by leveraging the smooth continuity of biological specimens across slices.
  • Processing Speed: While efficient, further acceleration of Leonardo-Fuse's feature extraction step using single-precision floating-point numbers is proposed. Real-time operation and integration with adaptive microscopy systems are envisioned.
  • Multi-view Reconstruction: While Leonardo enhances inputs for multi-view reconstruction, it doesn't directly perform multi-view fusion or deconvolution itself. The authors note that existing toolboxes for this exist and can benefit from Leonardo's preprocessing.
  • Generalization: While Leonardo is designed to be generalizable, further testing across a wider range of microscopy techniques and sample types is suggested.

Summary updated

Abstract

<title>Abstract</title> Selective plane illumination microscopy (SPIM, also known as light sheet fluorescence microscopy) is the method of choice for studying organ morphogenesis and function, as it permits gentle and rapid volumetric imaging of biological specimens over days. In inhomogeneous samples, however, sample-induced aberrations, including light absorption, scattering, and refraction, can degrade the image, particularly as the focal plane gets deeper into the sample. Here, we present Leonardo, the first toolbox capable of resolving all sample-induced aberrations by using two modules: (1) DeStripe removes stripe artifacts in SPIM caused by light absorption; (2) Fuse reconstructs a single high-quality image from dual-sided illumination and/or dual-sided detection, while eliminating blur and optical distortions caused by light scattering and refraction. The efficacy of Leonardo is validated on a wide range of biological systems, from minimally invasive experiments on sensitive specimens (relatively translucent embryonic and optically opaque larval zebrafish) to immunolabeled mice up to 2 centimeters in size. We provide model code and a Napari-based graphical user interface, enabling the SPIM community to easily apply Leonardo to advance light sheet imaging of inhomogeneous and complex specimens.
BibTeX
@inproceedings{peng2025leonardo,
  author = {Tingying Peng and Yu Liu and G.A. Müller and Lennart Kowitz and Tomáš Chobola and Kurt Weiss and Paul Maier and Jie Luo and Malte Roeßing and M Stenzel and Anika Grüneboom and Johannes C. Paetzold and Ali Ertürk and Nassir Navab and Carsten Marr and Jianxu Chen and Jan Huisken},
  title = {{Leonardo: A Toolset to Remove Sample-Induced Aberrations in Light Sheet Microscopy Images}},
  booktitle = {Research Square},
  year = {2025},
  doi = {10.21203/rs.3.rs-5853941/v1},
}
Figure