Limits...
Extracellular space preservation aids the connectomic analysis of neural circuits.

Pallotto M, Watkins PV, Fubara B, Singer JH, Briggman KL - Elife (2015)

Bottom Line: Any improvement in segmentation error rates would therefore directly reduce the time required to analyze 3D EM data.ECS preserved tissue is easier to segment using machine learning algorithms, leading to significantly reduced error rates.We conclude that preservation of ECS benefits multiple aspects of the connectomic analysis of neural circuits.

View Article: PubMed Central - PubMed

Affiliation: Circuit Dynamics and Connectivity Unit, National Institute of Neurological Disorders and Stroke, National Institutes of Health, Bethesda, United States.

ABSTRACT
Dense connectomic mapping of neuronal circuits is limited by the time and effort required to analyze 3D electron microscopy (EM) datasets. Algorithms designed to automate image segmentation suffer from substantial error rates and require significant manual error correction. Any improvement in segmentation error rates would therefore directly reduce the time required to analyze 3D EM data. We explored preserving extracellular space (ECS) during chemical tissue fixation to improve the ability to segment neurites and to identify synaptic contacts. ECS preserved tissue is easier to segment using machine learning algorithms, leading to significantly reduced error rates. In addition, we observed that electrical synapses are readily identified in ECS preserved tissue. Finally, we determined that antibodies penetrate deep into ECS preserved tissue with only minimal permeabilization, thereby enabling correlated light microscopy (LM) and EM studies. We conclude that preservation of ECS benefits multiple aspects of the connectomic analysis of neural circuits.

No MeSH data available.


Related in: MedlinePlus

Warping error of automated segmentations.(A) Split/merger curves parameterized by threshold of the network output. Shaded area is the standard error of the mean, lines are the mean number of splits/mergers per object, and the circles are the mean of the minimum total splits + mergers per object. (B) Network segmentations at the minimum warping error threshold, same regions as in Figure 2A. Warping errors were 0.68 (0.6% ECS), 0.32 (5.8% ECS), 0.52 (11.3% ECS), and 0.56 (23.9% ECS) for the four examples shown. (C) Cumulative distributions of the number of ground truth intracellular pixels per object for the four datasets. The difference in these distributions motivated the use of a Bernoulli sampling procedure for calculating the adapted Rand Error metrics in Figure 2C. Scale bars: 2 μm. ECS: Extracellular space.DOI:http://dx.doi.org/10.7554/eLife.08206.011
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4764589&req=5

fig2s1: Warping error of automated segmentations.(A) Split/merger curves parameterized by threshold of the network output. Shaded area is the standard error of the mean, lines are the mean number of splits/mergers per object, and the circles are the mean of the minimum total splits + mergers per object. (B) Network segmentations at the minimum warping error threshold, same regions as in Figure 2A. Warping errors were 0.68 (0.6% ECS), 0.32 (5.8% ECS), 0.52 (11.3% ECS), and 0.56 (23.9% ECS) for the four examples shown. (C) Cumulative distributions of the number of ground truth intracellular pixels per object for the four datasets. The difference in these distributions motivated the use of a Bernoulli sampling procedure for calculating the adapted Rand Error metrics in Figure 2C. Scale bars: 2 μm. ECS: Extracellular space.DOI:http://dx.doi.org/10.7554/eLife.08206.011

Mentions: We then quantified the number of misclassified pixels, the classification error (Figure 2A, third column), on test images. Similar pixel classification errors were observed regardless of ECS fraction (Figure 2B), indicating that networks were equally able to learn the statistics of the images. We, however, were interested primarily in the degree to which ECS preservation affected the ability of neural networks to segment actual cells (Jain et al., 2010). We therefore joined (segmented) regions of neighboring intracellular pixels and measured segmentation performance using two common segmentation metrics (see 'Materials and methods'). We color-coded segmentations using a four-color map method (Figure 2A, fourth column) such that neighboring objects do not share a color. This approach helps to highlight the locations of two predominant error-types: mergers between neurons that should have been segmented as independent objects as well as the splitting of neurons into multiple objects. We observed a significant reduction in segmentation errors that correlated with increasing ECS fraction, regardless of the error metric we used (Figure 2C, D, Figure 2—figure supplement 1). Therefore, despite sharing similar pixel classification errors, images that contained some degree of ECS were easier to automatically segment. This is because ECS provides more separation between neighboring neurons leading to, for example, fewer mergers between cells.


Extracellular space preservation aids the connectomic analysis of neural circuits.

Pallotto M, Watkins PV, Fubara B, Singer JH, Briggman KL - Elife (2015)

Warping error of automated segmentations.(A) Split/merger curves parameterized by threshold of the network output. Shaded area is the standard error of the mean, lines are the mean number of splits/mergers per object, and the circles are the mean of the minimum total splits + mergers per object. (B) Network segmentations at the minimum warping error threshold, same regions as in Figure 2A. Warping errors were 0.68 (0.6% ECS), 0.32 (5.8% ECS), 0.52 (11.3% ECS), and 0.56 (23.9% ECS) for the four examples shown. (C) Cumulative distributions of the number of ground truth intracellular pixels per object for the four datasets. The difference in these distributions motivated the use of a Bernoulli sampling procedure for calculating the adapted Rand Error metrics in Figure 2C. Scale bars: 2 μm. ECS: Extracellular space.DOI:http://dx.doi.org/10.7554/eLife.08206.011
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4764589&req=5

fig2s1: Warping error of automated segmentations.(A) Split/merger curves parameterized by threshold of the network output. Shaded area is the standard error of the mean, lines are the mean number of splits/mergers per object, and the circles are the mean of the minimum total splits + mergers per object. (B) Network segmentations at the minimum warping error threshold, same regions as in Figure 2A. Warping errors were 0.68 (0.6% ECS), 0.32 (5.8% ECS), 0.52 (11.3% ECS), and 0.56 (23.9% ECS) for the four examples shown. (C) Cumulative distributions of the number of ground truth intracellular pixels per object for the four datasets. The difference in these distributions motivated the use of a Bernoulli sampling procedure for calculating the adapted Rand Error metrics in Figure 2C. Scale bars: 2 μm. ECS: Extracellular space.DOI:http://dx.doi.org/10.7554/eLife.08206.011
Mentions: We then quantified the number of misclassified pixels, the classification error (Figure 2A, third column), on test images. Similar pixel classification errors were observed regardless of ECS fraction (Figure 2B), indicating that networks were equally able to learn the statistics of the images. We, however, were interested primarily in the degree to which ECS preservation affected the ability of neural networks to segment actual cells (Jain et al., 2010). We therefore joined (segmented) regions of neighboring intracellular pixels and measured segmentation performance using two common segmentation metrics (see 'Materials and methods'). We color-coded segmentations using a four-color map method (Figure 2A, fourth column) such that neighboring objects do not share a color. This approach helps to highlight the locations of two predominant error-types: mergers between neurons that should have been segmented as independent objects as well as the splitting of neurons into multiple objects. We observed a significant reduction in segmentation errors that correlated with increasing ECS fraction, regardless of the error metric we used (Figure 2C, D, Figure 2—figure supplement 1). Therefore, despite sharing similar pixel classification errors, images that contained some degree of ECS were easier to automatically segment. This is because ECS provides more separation between neighboring neurons leading to, for example, fewer mergers between cells.

Bottom Line: Any improvement in segmentation error rates would therefore directly reduce the time required to analyze 3D EM data.ECS preserved tissue is easier to segment using machine learning algorithms, leading to significantly reduced error rates.We conclude that preservation of ECS benefits multiple aspects of the connectomic analysis of neural circuits.

View Article: PubMed Central - PubMed

Affiliation: Circuit Dynamics and Connectivity Unit, National Institute of Neurological Disorders and Stroke, National Institutes of Health, Bethesda, United States.

ABSTRACT
Dense connectomic mapping of neuronal circuits is limited by the time and effort required to analyze 3D electron microscopy (EM) datasets. Algorithms designed to automate image segmentation suffer from substantial error rates and require significant manual error correction. Any improvement in segmentation error rates would therefore directly reduce the time required to analyze 3D EM data. We explored preserving extracellular space (ECS) during chemical tissue fixation to improve the ability to segment neurites and to identify synaptic contacts. ECS preserved tissue is easier to segment using machine learning algorithms, leading to significantly reduced error rates. In addition, we observed that electrical synapses are readily identified in ECS preserved tissue. Finally, we determined that antibodies penetrate deep into ECS preserved tissue with only minimal permeabilization, thereby enabling correlated light microscopy (LM) and EM studies. We conclude that preservation of ECS benefits multiple aspects of the connectomic analysis of neural circuits.

No MeSH data available.


Related in: MedlinePlus