Label-free phenotyping of human microvessel networks

Luca Rappez, Akinola Akinbote, Marta Cherubini, Virginie Uhlmann, Kristina HaasebioRxiv 2024 · DOI

TL;DR

This paper introduces VascuMap, a deep-learning tool that enables label-free segmentation and comprehensive spatial analysis of entire microvessel networks from brightfield microscopy images. This approach overcomes limitations of traditional fluorescent labeling methods, paving the way for more accurate, high-throughput, and non-invasive studies of vascular development and disease.

What Problem Does This Paper Address?

Studying the intricate spatial heterogeneity of microvessel networks is crucial for understanding various biomedical applications, from development to disease. However, existing methods for analyzing these networks primarily rely on fluorescent labeling. This approach has several drawbacks: it can interfere with the biological sample's homeostasis, limit studies to specific precursor cells, require sample fixation (preventing live measurements), and often focuses on small, cropped images, neglecting the global network structure and leading to biased interpretations. The paper aims to address these limitations by developing a label-free method that can analyze entire microvessel networks with high accuracy.

Methods

The core of VascuMap is a deep learning approach, specifically a U-Net convolutional neural network (CNN), for vessel segmentation. To handle the challenges of brightfield microscopy, such as complex ECM optical properties and varying vessel heights, the authors developed a custom "virtual refocusing" technique. This involves analyzing a 3D z-stack of images, identifying the optimal focal plane for different regions using the tenengrad variance method, and then reconstructing a single, in-focus 2D image of the entire microvessel network.

After segmentation, the vessel mask is converted into a graph representation. Nodes in this graph represent vessel junctions, and edges represent vessel pathways. This graph structure allows for the calculation of a comprehensive set of morphological and topological metrics, providing detailed insights into the network's structure and complexity.

Datasets

The paper describes the generation of a training dataset using microscopy images of microvessels. Ground truth labels were manually drawn on microscopy images (combining brightfield and fluorescent channels) using an iPad and Apple Pencil. These images were then processed into 256x256 pixel tiles. The paper mentions acquiring 385 images using a Leica Thunder Imager and 812 using a Leica Stellaris 8 microscope. The dataset was split into 70% for training and 30% for validation.

Key Insights and Ideas

  • Label-free is feasible and effective: The study demonstrates that label-free segmentation using brightfield microscopy can achieve segmentation accuracy comparable to fluorescence-based methods, which is a significant step towards non-invasive imaging.
  • Virtual refocusing is key for 3D brightfield: The custom virtual refocusing technique is crucial for overcoming the challenges of imaging planar microvessels in 3D z-stacks, enabling the analysis of the entire network in a single, in-focus image.
  • Morphological and topological metrics together reveal biological differences: Combining detailed morphological and topological metrics derived from a graph representation of the vascular network allows for the clear distinction between different culture conditions, highlighting the biological relevance of these features.
  • High-throughput potential: The compatibility with widefield label-free microscopy significantly accelerates sample acquisition, making VascuMap ideal for high-throughput applications like drug toxicity and safety screens.

Key Results

  • VascuMap achieves high segmentation accuracy, comparable to state-of-the-art fluorescence-based models, even when segmenting from brightfield microscopy.
  • The method successfully distinguishes between microvessels cultured in different media conditions (basal vs. complete), with the complete medium promoting larger, more interconnected vessels, as evidenced by metrics like fractal dimension and lacunarity.
  • Principal Component Analysis (PCA) of the collected metrics clearly separates samples based on their culture conditions, demonstrating the power of VascuMap's feature extraction.

Limitations and Future Work

  • Limitations: While the paper highlights the comparable accuracy of brightfield segmentation to fluorescence, it notes a slight improvement in Dice accuracy when using fluorescence. The current method is primarily focused on 2D analysis after virtual refocusing, and the authors suggest extending this to full 3D segmentation.
  • Future Work: The authors envision VascuMap being extended to analyze microvasculature from other species and modalities, including human micro-CT or imaging studies in chick and zebrafish models. They also suggest exploring more advanced deep learning architectures, such as 3D CNNs for label-free 3D segmentation and graph neural networks for more nuanced topological analysis. Furthermore, they propose integrating methodologies from fields like ecology and sociology for more in-depth quantitative studies of metric co-occurrences.

Summary updated

Abstract

Abstract Understanding the spatial heterogeneity in blood vessel formation and development is crucial for various biomedical applications. Traditional methods for in-vitro microvessel segmentation rely on fluorescent labeling, which either interferes with the sample homeostasis, limits the study to a restricted set of precursor cells, or requires sample fixation, thus preventing live measurements. Moreover, these methods often focus on small, cropped images, neglecting global spatial heterogeneity of microvasculature, leading to biased data interpretation. To overcome these limitations, we present VascuMap, a deep-learning-based tool for label-free vessel segmentation and spatial analysis. VascuMap enables a comprehensive examination of entire vessel networks, capturing both morphological and topological features across the full vascular bed. Our method achieves high segmentation accuracy, comparable to the state-of-the-art fluorescence-based models. VascuMap’s capabilities extend to characterizing vasculature generated from label-free patient-derived samples, a vital step towards personalized medicine. Its compatibility with widefield label-free microscopy also accelerates sample acquisition, making it ideal for high-throughput systems crucial for drug toxicity and safety screens.
BibTeX
@misc{rappez2024label,
  author = {Luca Rappez and Akinola Akinbote and Marta Cherubini and Virginie Uhlmann and Kristina Haase},
  title = {{Label-free phenotyping of human microvessel networks}},
  year = {2024},
  doi = {10.1101/2024.02.20.581133},
  howpublished = {bioRxiv},
}
Figure