MorphoFeatures: unsupervised exploration of cell types, tissues and organs in volume electron microscopy

Valentyna Zinchenko, Johannes Hugger, Virginie Uhlmann, Detlev Arendt, Anna KreshukbioRxiv 2022 · DOI

TL;DR

This paper introduces "MorphoFeatures," a novel unsupervised deep learning method to extract comprehensive morphological descriptors from 3D electron microscopy data. This approach enables the unbiased exploration and identification of cell types, tissues, and organs in entire organisms, bridging the gap between detailed ultrastructural imaging and large-scale biological analysis.

What Problem Does This Paper Address?

Electron microscopy (EM) offers unparalleled detail of cellular ultrastructure, but analyzing large volumes of this data has been a major bottleneck. While segmenting cells in EM volumes is becoming more routine, extracting meaningful, comprehensive morphological descriptions automatically across an entire organism has been challenging. Existing methods often rely on manual feature engineering or are task-specific, limiting their applicability for broad, exploratory analyses. This paper aims to overcome these limitations by developing a generalizable, unsupervised method to learn rich morphological features directly from 3D EM data, enabling large-scale, unbiased exploration of cellular and tissue organization.

Methods

The core of the method is a deep learning pipeline that learns "MorphoFeatures" from segmented 3D EM data. This pipeline uses a convolutional neural network trained in a self-supervised manner. It extracts three key components of cell morphology:

  • Shape: Learned from 3D point clouds and surface normals of cell membranes.
  • Coarse Texture: Extracted from downsampled raw EM data.
  • Fine Texture: Captured from high-resolution crops of raw EM data.

The self-supervised training employs two complementary objectives:

  1. Autoencoder Reconstruction Loss: Encourages the network to learn a comprehensive representation by reconstructing the original cell volume.
  2. Contrastive Loss: Pulls representations of similar cell views closer together in feature space while pushing dissimilar ones apart, ensuring that learned features reflect morphological similarity.

To identify tissues and organs, they introduce MorphoContextFeatures, which augment a cell's MorphoFeatures with the average MorphoFeatures of its immediate neighbors. This allows for the identification of tissues as groups of morphologically similar cells, even if the individual cells within the tissue are not morphologically identical.

Datasets

The primary dataset used is a 3D serial block-face electron microscopy (SBEM) volume of the entire marine annelid Platynereis dumerilii at the 6 days post-fertilization stage. This dataset includes:

  • A whole-animal SBEM volume with nanometer resolution.
  • Automated segmentation of 11,402 cells and nuclei.
  • Whole-animal gene expression maps covering differentiation genes and transcription factors.

This dataset is publicly available and was used for both training and evaluation.

Key Insights and Ideas

  • Unsupervised learning unlocks large-scale EM analysis: By avoiding manual feature engineering and task-specific supervision, the method allows for unbiased exploration of complex biological structures.
  • Morphological features are interpretable: Despite being learned by a deep network, individual MorphoFeatures correspond to visually understandable cellular properties like cytoplasm thickness, nuclear shape, and organelle appearance.
  • Neighborhood context is crucial for tissue identification: Incorporating information from neighboring cells (MorphoContextFeatures) is essential for grouping cells into tissues and organs, revealing organizational principles beyond individual cell morphology.
  • MorphoFeatures offer superior resolution: The learned features can distinguish cell types with finer granularity than previously used manually defined features, even identifying cell types not readily apparent in explicit feature sets.
  • Bridging morphology and genetics: The framework effectively links morphological clusters with gene expression data, providing a powerful tool for understanding the genetic basis of cellular identity and function.

Key Results

  • High accuracy in cell classification: A logistic regression model trained on MorphoFeatures achieved 96% accuracy in classifying 7 broad cell types, outperforming a model trained on 140 manually defined features (94% accuracy).
  • Improved bilateral symmetry matching: MorphoFeatures were 38% more precise in identifying symmetric cell partners compared to manually defined features, indicating a more nuanced representation of morphology.
  • Identification of distinct cell types and tissues: The method successfully clustered cells into morphologically coherent groups, which correlated well with gene expression profiles, revealing distinct cell types (e.g., enteric neurons, photoreceptors) and tissues (e.g., foregut structures, nervous system ganglia).
  • Discovery of novel structures: The MorphoContextFeatures approach enabled the identification of previously uncharacterized tissues, such as an anterior extension of the axochord surrounding the foregut and a potential infracerebral gland.

Limitations and Future Work

  • Data quality dependence: The performance of the pipeline is dependent on the quality of the initial cell segmentation and the EM data itself. Intensity variations across volumes can be a challenge.
  • Neurite tracing limitations: For neurons, only somata were segmented due to resolution limits, meaning neurite morphology was not directly incorporated into the features.
  • Future directions:
    • Improving robustness to non-biological intensity variations.
  • Adapting the pipeline for other imaging modalities (e.g., light microscopy, X-ray tomography).
  • Integrating other data modalities (e.g., transcriptomics, proteomics) with MorphoFeatures.
  • Leveraging advancements in self-supervised learning, such as models that only rely on positive samples.
  • Exploring graph-based contrastive learning for MorphoContextFeatures.

Summary updated

Abstract

Abstract Electron microscopy (EM) provides a uniquely detailed view of cellular morphology, including organelles and fine subcellular ultrastructure. While the acquisition and (semi-)automatic segmentation of multicellular EM volumes is now becoming routine, large-scale analysis remains severely limited by the lack of generally applicable pipelines for automatic extraction of comprehensive morphological descriptors. Here, we present a novel unsupervised method for learning cellular morphology features directly from 3D EM data: a convolutional neural network delivers a representation of cells by shape and ultrastructure. Applied to the full volume of an entire three-segmented worm of the annelid Platynereis dumerilii , it yields a visually consistent grouping of cells supported by specific gene expression profiles. Integration of features across spatial neighbours can retrieve tissues and organs, revealing, for example, a detailed organization of the animal foregut. We envision that the unbiased nature of the proposed morphological descriptors will enable rapid exploration of very different biological questions in large EM volumes, greatly increasing the impact of these invaluable, but costly resources.
BibTeX
@misc{zinchenko2022morphofeatures,
  author = {Valentyna Zinchenko and Johannes Hugger and Virginie Uhlmann and Detlev Arendt and Anna Kreshuk},
  title = {{MorphoFeatures: unsupervised exploration of cell types, tissues and organs in volume electron microscopy}},
  year = {2022},
  doi = {10.1101/2022.05.07.490949},
  howpublished = {bioRxiv},
}
Figure