Automated Analysis of Ultrastructure: A Study Based on Large-Scale Hyperspectral Electron Microscopy

Academic Background

Electron Microscopy (EM) is a key technique for studying biological ultrastructure, revealing the fine details of cells at biomolecular resolution. In recent years, advancements in automation and digitization have enabled EM to capture large areas of cells and tissues at nanoscale resolution. However, EM images are typically grayscale and involve massive data volumes, often requiring laborious manual annotation, which limits their application in large-scale studies. To address this issue, researchers have begun exploring automated methods for extracting biomolecular assembly information, thereby accelerating the understanding of biological ultrastructure.

The background of this study lies in the challenges faced by EM in biomedical research, particularly in automatically extracting biomolecular information from large-scale EM data. This paper proposes an automated analysis method based on hyperspectral energy-dispersive X-ray (EDX) imaging, aiming to extract biomolecular assemblies from conventionally processed tissues in an unsupervised manner, thereby reducing manual intervention and improving analysis efficiency.

Source of the Paper

This paper was co-authored by B. H. Peter Duinkerken, Ahmad M. J. Alsahaf, Jacob P. Hoogenboom, and Ben N. G. Giepmans, from the Department of Biomedical Sciences at the University Medical Center Groningen and the Department of Imaging Physics at Delft University of Technology, Netherlands. The paper was published in npj Imaging in 2024.

Research Process

1. Large-Scale EDX Imaging and Spatial-Elemental Context

The study first developed a workflow for large-scale EDX imaging and applied it to an entire section of pancreatic islets. By extending pixel dwell time and frame accumulation, spectral richness was ensured. Simultaneously, the use of relatively high beam currents (4-5 nA) and two symmetrically positioned EDX detectors overcame the issue of long acquisition times. The study visualized the distribution of three elements (phosphorus, osmium, and iron) using false-color images, clearly distinguishing biological features such as granules, nuclei, lysosomes, and rough endoplasmic reticulum within cells.

2. Spectral Mixture Analysis and Biological Feature Extraction

To extract biological features from hyperspectral images (HSI), the study employed Spectral Mixture Analysis (SMA). First, representative spectra were obtained through manual annotation, and then linear unmixing was used to decompose the HSI into relative abundance maps of each spectrum. The study also introduced dimensionality reduction techniques, such as manifold learning, using PACMAP (Pairwise Controlled Manifold Approximation Projection) to reduce the HSI to two dimensions, thereby better identifying biological features.

3. Large-Scale Spectral Mixture Analysis

In large-scale EDX imaging, the data volume is immense. The study reduced the data volume through subsampling and excluding overlapping regions. By using PACMAP for dimensionality reduction, the purest spectra (i.e., endmembers) were identified, and a linear mixture model was used for unmixing. The study demonstrated the abundance maps of different biological structures in pancreatic and skin tissues, comparing them with expected biochemical compositions.

4. Automated Detection and Segmentation of EM Images

The study also explored automated segmentation methods based on Convolutional Neural Networks (CNNs), combining EDX endmember abundance maps with the Segment Anything Model (SAM) for automatic segmentation. Using spatial prompts, the SAM model was able to automatically detect and segment five distinct organelles and biostructures, showcasing the potential of EDX HSI in unsupervised EM analysis.

Main Results

1. Large-Scale EDX Imaging: The study successfully developed a workflow for large-scale EDX imaging, enabling the differentiation of various biological features within cells, such as granules, nuclei, lysosomes, and rough endoplasmic reticulum, at ultrastructural resolution.

2. Spectral Mixture Analysis: Through SMA, the study extracted spectral features of biological structures and generated corresponding abundance maps, illustrating the spatial distribution of different biological features.

3. Dimensionality Reduction and Endmember Extraction: Using PACMAP for dimensionality reduction, the study identified the purest spectra and performed unmixing with a linear mixture model, demonstrating the abundance maps of different biological structures in pancreatic and skin tissues.

4. Automated Segmentation: By combining EDX endmember abundance maps with the SAM model, the study achieved automated detection and segmentation of EM images, highlighting the potential of EDX HSI in unsupervised EM analysis.

Conclusions and Significance

This paper presents an automated analysis method based on large-scale hyperspectral EDX imaging, enabling the unsupervised extraction of biomolecular assemblies and reducing manual intervention, thereby improving the efficiency of EM data analysis. This method not only accelerates the understanding of biological ultrastructure but also provides new tools and insights for future large-scale EM research.

Research Highlights

1. Innovative Method: This study is the first to apply large-scale hyperspectral EDX imaging to biomedical research, proposing an unsupervised automated analysis method for extracting biomolecular assemblies from conventionally processed tissues.

2. Efficient Data Analysis: Through SMA and dimensionality reduction techniques, the study was able to extract biological features from large-scale EDX data, reducing the workload of manual annotation.

3. Automated Segmentation Technology: By combining EDX endmember abundance maps with the SAM model, the study achieved automated detection and segmentation of EM images, showcasing the potential of EDX HSI in unsupervised EM analysis.

Additional Valuable Information

All data and analysis codes from this study have been made publicly available. Researchers can access related resources via www.nanotomy.org and GitHub to further explore and apply the methods proposed in this paper.

Through this research, the application of EM in the biomedical field is poised for new breakthroughs. This method is expected to be extended to more tissues and samples in the future, further advancing the understanding of biological ultrastructure.