Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2020 Jul 21:14:599.
doi: 10.3389/fnins.2020.00599. eCollection 2020.

Automatic Reconstruction of Mitochondria and Endoplasmic Reticulum in Electron Microscopy Volumes by Deep Learning

Jing Liu  1   2 , Linlin Li  1 , Yang Yang  3 , Bei Hong  1   2 , Xi Chen  1 , Qiwei Xie  1   4 , Hua Han  1   2   5
Affiliations

Automatic Reconstruction of Mitochondria and Endoplasmic Reticulum in Electron Microscopy Volumes by Deep Learning

Jing Liu et al. Front Neurosci. .

Abstract

Together, mitochondria and the endoplasmic reticulum (ER) occupy more than 20% of a cell's volume, and morphological abnormality may lead to cellular function disorders. With the rapid development of large-scale electron microscopy (EM), manual contouring and three-dimensional (3D) reconstruction of these organelles has previously been accomplished in biological studies. However, manual segmentation of mitochondria and ER from EM images is time consuming and thus unable to meet the demands of large data analysis. Here, we propose an automated pipeline for mitochondrial and ER reconstruction, including the mitochondrial and ER contact sites (MAMs). We propose a novel recurrent neural network to detect and segment mitochondria and a fully residual convolutional network to reconstruct the ER. Based on the sparse distribution of synapses, we use mitochondrial context information to rectify the local misleading results and obtain 3D mitochondrial reconstructions. The experimental results demonstrate that the proposed method achieves state-of-the-art performance.

Keywords: 3D reconstruction; electron microscopes; endoplasmic reticulum; mitochondria; segmentation.

PubMed Disclaimer

Figures

Figure 1
Figure 1
Workflow of our proposed method. Briefly, there are four main steps in the complete method. (1) Data preparation: First, serial images are obtained by SEM and then aligned to create 3D image stacks using a non-linear registration method. (2) Model training: From the aligned data, we annotate some images to form a training dataset, and then separately train two networks to process mitochondria and ER. (3) Ultrastructure reconstruction: 2D mitochondria were predicted using the trained network and connected to form 3D mitochondria. The ER is segmented by the trained network. Then, we detect mitochondria-ER contact sites from the above results. (4) Data analysis: We measure the performance of our method and calculate biological measurements to conduct various analyses.
Figure 2
Figure 2
Mouse cortex neural tissue acquired by ATUM-SEM. (A) An example of an aligned image stack that covers approximately 20 × 20 × 10 μm through the ATUM-SEM method. (B,C) Examples of mitochondria and other ultrastructures. The green arrows indicate mitochondria; the red arrows indicate vesicles; the yellow arrow indicates the Golgi body, and the purple arrow indicates the endoplasmic reticulum. (D–F) Examples of mitochondria segmentation in our training dataset.
Figure 3
Figure 3
The architecture and training scheme of the proposed network for segmenting mitochondria. Left: The three dotted boxes represent the backbone network (purple), the detection subnetwork (blue) and the recursive segmentation subnetwork (green). The black block, blue blocks, orange blocks, green blocks, purple blocks, and red blocks indicate the input image, convolution layers, classification layers, regression layers, and the fixed-size feature maps obtained from the RoIAlign and mask channels, respectively. Right: A simplified execution scheme for segmenting mitochondria. The green box indicates the outputs from the detection subnetwork, which may be smaller than the ground truth. After the first iteration of the segmentation subnetwork, the eight prediction directions are checked, and the direction is the input for the next iteration. In this example, the position in the orange circle is selected, and then the field of view (FoV) moves to the orange box.
Figure 4
Figure 4
Performance comparisons with the baseline approaches. From left to right: comparisons to the manual ground truth from U-Net, FFN-2d, Mask R-CNN and our proposed network. In each comparison image, green pixels represent true positives (TP), blue pixels represent false positives (FP), red pixels represent false negatives (FN), and black pixels represent true negatives (TN). The insert at the top right shows enlarged details pointed to by the white arrow. Qualitatively, the segmentation by the proposed network contains fewer false positives than that of FFN-2d and more true positives than those of U-Net and Mask R-CNN.
Figure 5
Figure 5
The architecture of the neural network for the endoplasmic reticulum. The yellow blocks and gray blocks denote two different residual blocks, and the numbers above the blocks indicate the number of feature map channels. The blue plus signs indicate a sum operation rather than concatenation.
Figure 6
Figure 6
Examples of training samples and predictions for ER segmentation. (A) An example of manual annotations of the nucleus (yellow), ER, and the nuclear membrane (red). (B,C) An inference from the trained neural network and the corresponding relabeled result. The red and green pixels indicate the ER and the nuclear membrane, respectively.
Figure 7
Figure 7
The 3D morphology of mitochondria for different neuron locations. (A–C) Raw EM images and 3D reconstructed mitochondria in axons, neuron somas and dendrites. (D) Top: The volumes of mitochondria in the cell body (n = 86) are greater than those in axons/dendrites (n = 81) (one-sided Mann–Whitney test, p < 0.0001). Bottom: The surface area of mitochondria in the cell body (n = 86) is greater than that in axons/dendrites (n = 81) (one-sided Mann–Whitney test, p = 0.0002). (E,F) Example of cross section analysis in non-nanotunneling and nanotunneling mitochondria, respectively. The red arrows indicate mitochondria. (G) The cross-sectional areas of non-nanotunneling and nanotunneling mitochondria.
Figure 8
Figure 8
The 3D morphology of ER in neurons. (A–D) Left: the original EM image; Right: ER 3D morphology (pink) under different directions. (E) Mitochondria and ER distance distribution in neurons. (F) Percentage of different mitochondria and ER distances in neurons.
See this image and copyright information in PMC

Similar articles

See all similar articles

Cited by

See all "Cited by" articles

References

    1. Abrisch R. G., Gumbin S. C., Wisniewski B. T., Lackner L. L., Voeltz G. K. (2020). Fission and fusion machineries converge at ER contact sites to regulate mitochondrial morphology. J. Cell Biol. 219:e201911122. 10.1083/jcb.201911122 - DOI - PMC - PubMed
    1. Agus M., Calì C., Al-Awami A., Gobbetti E., Magistretti P., Hadwiger M. (2019). Interactive volumetric visual analysis of glycogen-derived energy absorption in nanometric brain structures. Comput. Graph. Forum 38, 427–439. 10.1111/cgf.13700 - DOI
    1. Bhatti J. S., Bhatti G. K., Reddy P. H. (2017). Mitochondrial dysfunction and oxidative stress in metabolic disorders–a step towards mitochondria based therapeutic strategies. Biochim. Biophys. Acta Mol. Basis Dis. 1863, 1066–1077. 10.1016/j.bbadis.2016.11.010 - DOI - PMC - PubMed
    1. Calì C., Agus M., Kare K., Boges D. J., Lehväslaiho H., Hadwiger M., et al. . (2019). 3D cellular reconstruction of cortical glia and parenchymal morphometric analysis from serial block-face electron microscopy of juvenile rat. Prog. Neurobiol. 183:101696. 10.1016/j.pneurobio.2019.101696 - DOI - PubMed
    1. Cardona A., Saalfeld S., Schindelin J., Arganda-Carreras I., Preibisch S., Longair M., et al. . (2012). TrakEM2 software for neural circuit reconstruction. PLoS ONE 7:e38011. 10.1371/journal.pone.0038011 - DOI - PMC - PubMed

LinkOut - more resources