Misic, a general deep learning-based method for the high-throughput cell segmentation of complex bacterial communities
Curation statements for this article:-
Curated by eLife
Evaluation Summary:
In this work, Panigrahi et. al. develop a powerful deep-learning-based cell segmentation platform (MiSiC) capable of accurately segmenting brightfield, fluorescence, and phase contrast images of bacteria cells densely packed within both homogenous and heterogeneous cell populations. This algorithm, if further optimized and disseminated to the community, will have a large impact in microbial studies in that it will allow for automated analyses of essential every aspect of bacterial cell biology, including cell morphology, cell cycle, cell-cell communications, protein localization dynamics and a variety of cellular processes using time-lapse imaging.
(This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #2 agreed to share their name with the authors.)
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
- Microbiology and Infectious Disease (eLife)
Abstract
Studies of bacterial communities, biofilms and microbiomes, are multiplying due to their impact on health and ecology. Live imaging of microbial communities requires new tools for the robust identification of bacterial cells in dense and often inter-species populations, sometimes over very large scales. Here, we developed MiSiC, a general deep-learning-based 2D segmentation method that automatically segments single bacteria in complex images of interacting bacterial communities with very little parameter adjustment, independent of the microscopy settings and imaging modality. Using a bacterial predator-prey interaction model, we demonstrate that MiSiC enables the analysis of interspecies interactions, resolving processes at subcellular scales and discriminating between species in millimeter size datasets. The simple implementation of MiSiC and the relatively low need in computing power make its use broadly accessible to fields interested in bacterial interactions and cell biology.
Article activity feed
-
-
Author Response:
Reviewer #1 (Public Review):
In this work, Panigrahi et. al. develop a powerful deep-learning-based cell segmentation platform (MiSiC) capable of accurately segmenting bacteria cells densely packed within both homogenous and heterogeneous cell populations. Notably, MiSiC can be easily implemented by a researcher without the need for high-computational power. The authors first demonstrate MiSiC's ability to accurately segment cells with a variety of shapes including rods, crescents and long filaments. They then demonstrate that MiSiC is able to segment and classify dividing and non-dividing Myxococcus cells present in a heterogenous population of E. coli and Myxococcus. Lastly, the authors outline a training workflow with which MiSiC can be trained to identify two different cell types present in a mixed population …
Author Response:
Reviewer #1 (Public Review):
In this work, Panigrahi et. al. develop a powerful deep-learning-based cell segmentation platform (MiSiC) capable of accurately segmenting bacteria cells densely packed within both homogenous and heterogeneous cell populations. Notably, MiSiC can be easily implemented by a researcher without the need for high-computational power. The authors first demonstrate MiSiC's ability to accurately segment cells with a variety of shapes including rods, crescents and long filaments. They then demonstrate that MiSiC is able to segment and classify dividing and non-dividing Myxococcus cells present in a heterogenous population of E. coli and Myxococcus. Lastly, the authors outline a training workflow with which MiSiC can be trained to identify two different cell types present in a mixed population using Myxococcus and E. coli as examples.
While we believe that MiSiC is a very powerful and exciting tool that will have a large impact on the bacterial cell biological community, we feel explanations of how to use the algorithm should be more greatly emphasized. To help other scientists use MiSiC to its fullest potential, the range of applications should be clarified. Furthermore, any inherent biases in MiSiC should be discussed so that users can avoid them.
We thank the reviewer for the positive feedback and comments to help disseminate MiSiC to the broad bacterial cell biology community as it is meant to. As described above we have largely addressed this comment via the redaction of a comprehensive handbook. As detailed below, we now also provide precise measurements of the MiSiC segmentation accuracy compared to ground truth for the various imaging modalities and bacterial species segmentation.
Major Concerns:
- It is unclear to us how a MiSiC user should choose/tune the value for the noise variance parameter. What exactly should be considered when choosing the noise variance parameter? Some possibilities include input image size, cell size (in pixels), cell density, and variance in cell size. Is there a recommended range for the parameter? These questions along with our second minor correction can be addressed with a paragraph in the Discussion section.
Setting the noise parameters is now detailed in the handbook (section 1.d). A set of thumb rules and recommendations are provided. In addition a paragraph explaining the importance of noise addition for images with sparse bacterial cell density has been added in the results section.
“Associated Figure S1. Background noise can lead to spurious cell detection by MiSiC. SI images retain the shape/curvature information of the intensities in a raw image through eigenvalues of the hessian of the image and an arctan function, creating the smooth areas corresponding to cell bodies and propagating noisy regions where there is no shape information. Thus, MiSiC segments the cells by discriminating between “smooth” and “rough” regions. In effect, when adjusting the size parameter, scaling smooths out the image noise, leading to background regions that have a smoother SI than in the raw image. Some of these areas could be falsely detected as bacterial cells. This effect is shown here: When an image with uniform and random intensity values is segmented with MiSiC with increasing smoothening (here using a gaussian blur filter), spurious cell detection becomes apparent. In addition, since the SI keeps the shape information and not the intensity values, background objects that are of relatively low contrast (ie dead cells or debris) may be detected as cells. All these artifacts can be mitigated by adding synthetic noise to the scaled images.”
- Could the authors expand on using algorithms like watershed, conditional random fields, or snake segmentation to segment bacteria when there is not enough edge information to properly separate them? How accurate are these methods at segmenting the cells? Should other MiSiC parameters be tuned to increase the accuracy when implementing these methods?
We thank the reviewer for raising this point as it is important to make clear that post-processing algorithms can certainly improve the accuracy of MiSiC masks downstream. To show this specifically, we further processed MiSiC masks of Bacillus subtilis filamentous cells to resolve division septa using the watershed algorithm. This example is now provided as Figure S3. Importantly, there is no particular MiSiC adjustment that needs to be performed prior to running these processing steps, which can be done directly in Image-J or its bacterial cell analysis plug-in, MicrobeJ. It is worth noting that the post- processing strategy may depend on the scientific question under consideration. In the handbook, we also give an example of post-processing methods that may be used.
“Associated Figure S3. Refining cell separations with watershed. Watershed methods may be used to obtain a more accurate segmentation of septate filaments such as Bacillus subtilis. In this example applying this method to the MiSiC mask effectively resolves cell boundaries that are not captured in the prediction but are visible by eye (arrows).”
- Can the MiSiC's ability to accurately segment phase and brightfield images be quantitatively compared against each other and against fluorescent images for overall accuracy? A figure similar to Fig. 2C, with the three image modalities instead of species would nicely complement Fig. 2A. If the segmentation accuracy varies significantly between image modalities, a researcher might want to consider the segmentation accuracy when planning their experiments. If the accuracy does not vary significantly, that would be equally useful to know.
This is a very important issue that was also raised by reviewer 3 and which we decided to address in full. For each imaging modality and distinct species, we measured the Jaccard Index as a function of the threshold set for the Intersection over Union (ioU). The resulting curves are now provided in two separate Figures 2 and 3 and a supplemental Figure S2; they provide a robust measure of the segmentation for each modality/tested species.
“Figure 2. MiSiC predictions under various imaging modalities. a) MiSiC masks and corresponding annotated masks of fluorescence, phase contrast and bright field images of a dense E. coli microcolony. b) Jaccard index as a function of IoU threshold for each modality determined by comparing the MiSiC masks to the ground truth (see Methods). The obtained Jaccard score curves are the average of analyses conducted over three biological replicates and n=763, 811, 799 total cells for Fluorescence, Phase Contrast and Bright Field, respectively (bands are the maximum range, the solid line is the median). The fluorescence images were pre-processed using a Gaussian of Laplacian filter to improve MiSiC prediction (see methods).”
“Associated Figure S2. MiSiC predictions under various imaging modalities. a) MiSiC masks and corresponding annotated masks of fluorescence, phase contrast and bright field images of a dense M. xanthus microcolony. b) Jaccard index as a function of IoU threshold for each modality determined by comparing the MiSiC masks to the ground truth (see Methods). The obtained curves are the average of analyses conducted over three biological replicates and n=193,206,211 total cells for Fluorescence, Phase Contrast and Bright Field, respectively. The fluorescence (bands are the maximum range, the solid line is the median) images were pre-processed using a Gaussian of Laplacian filter to improve MiSiC prediction (see methods). c) A human observer is slightly less performant than MiSiC. The same ground truth as used in Figure 2 (dashed lines) was compared to an independent observer’s annotation (solid lines) and Jaccard score curves were constructed as shown in Figure 2. BF: Bright Field, PC: Phase Contrast, Fluo: Fluorescence.”
“Figure 3. MiSiC predictions in various bacterial species and shapes. a) MiSiC masks and corresponding annotated masks of phase contrast images of another Pseudomonas aeruginosa (rod-shape), Caulobacter crescentus (crescent shape) and Bacillus subtilis (filamentous shape). b) Jaccard index as a function of IoU threshold for each species determined by comparing the MiSiC masks to the ground truth (see Methods). The obtained Jaccard score curves are the average of analyses conducted over three biological replicates and n=1149,101,216 total cells for P. aeruginosa, B. subtilis and C. crescentus, respectively (bands are the maximum range, solid line the median). Note that the B. subtilis filaments are well predicted but edge information is missing for optimal detection of the cell separations.”
- The ability of MiSiC to segment dense clusters of cells is an exciting advancement for cell segmentation algorithms. However, is there a minimum cell density required for robust segmentation with MiSiC? The algorithm should be applied to a set of sparsely populated images in a supplemental figure. Is the algorithm less accurate for sparse images (perhaps reflected by an increase in false-positive cell identifications)? Any possible biases related to cell density should be noted.
In fact, MiSiC performs well both with densely or sparsely populated images. In the case of sparsely populated images it is however possible that non-cell objects can occasionally appear in the MiSiC mask. As mentioned above, inclusion of noise can help remove these objects in the sparsely populated images. This issue is now fully explained in a supplemental Figure S1. Of note, non-cell objects -if they were to remain after noise addition- can be eliminated using additional general morphometric filters or specific models fitting bacterial cells, as for example those included in Microbe-J and Oufti. These points are now clarified in the text.
“Associated Figure S1. Background noise can lead to spurious cell detection by MiSiC. SI images retain the shape/curvature information of the intensities in a raw image through eigenvalues of the hessian of the image and an arctan function, creating the smooth areas corresponding to cell bodies and propagating noisy regions where there is no shape information. Thus, MiSiC segments the cells by discriminating between “smooth” and “rough” regions. In effect, when adjusting the size parameter, scaling smooths out the image noise, leading to background regions that have a smoother SI than in the raw image. Some of these areas could be falsely detected as bacterial cells. This effect is shown here: When an image with uniform and random intensity values is segmented with MiSiC with increasing smoothening (here using a gaussian blur filter), spurious cell detection becomes apparent. In addition, since the SI keeps the shape information and not the intensity values, background objects that are of relatively low contrast (ie dead cells or debris) may be detected as cells. All these artifacts can be mitigated by adding synthetic noise to the scaled images.”
and:
“Along similar lines, non-cell objects can appear in the MiSiC masks and while some can be removed by the introduction of noise, an easy way to do it is to apply a post-processing filter, for example using morphometric parameters to remove objects that are not bacteria. This can be easily done using Fiji, MicrobeJ or Oufti."
- It is exciting to see the ability of MiSiC to segment single cells of M. xanthus and E. coli species in densely packed colonies (Fig. 4b). Although three morphological parameters after segmentation were compared with ground truth, the comparison was conducted at the ensemble level (Fig. 4c). Could the authors use the Mx-GFP and Ec-mCherry fluorescence as a ground truth at the single cell level to verify the results of segmentation? For example, for any Ec cells identified by MiSiC in Fig. 4b, provide an index of whether its fluorescence is red or green. This single-cell level comparison is most important for the community.
We have now performed this comparison and determined Jaccard indexes for E. coli and Myxococcus detection using the individual fluorescence images as a reference (figure 5b). Since we were only able to make this comparison in relatively small fields we also kept the comparison of expected morphometric parameters in large images. Taken together, these data now demonstrate that semantic classification as performed does well separate Myxococcus cells from E. coli cells (see more details in our response to reviewer 3).
Reviewer #2 (Public Review):
Panigrahi and co-authors introduce a program that can segment a variety of images of rod-shaped bacteria (with somewhat different sizes and imaging modalities) without fine-tuning. Such a program will have a large impact on any project requiring segmentation of a large number of rod-shaped cells, including the large images demonstrated in this manuscript. To my knowledge, training a U-Net to classify an image from the image's shape index maps (SIM) is a new scheme, and the authors show that it performs fairly well despite a small training set including synthetic data that, based on Figure 1, does not closely resemble experimental data other than in shape. The authors discuss extending the method to objects with other shapes and provide an example of labelling two different species - these extensions are particularly promising.
The authors show that their network can reproduce results of manual segmentation with bright field, phase and fluorescence input. Performance on fluorescence data in Fig. 1 where intensities vary so much is particularly good and shows benefits of the SIM transformation. Automated mapping of FtsZ show that this method can be immediately useful, though the authors note this required post-processing to remove objects with abnormal shapes. The application in mixed samples in Fig. 4 shows good performance. However, no Python workflow or application is provided to reproduce it or train a network to classify mixtures in different experiments.
We thank the reviewer for the positive comment. As discussed in our answer to reviewer 1, the classification presented in Figure 4 (now Figure 5) is meant to provide an example of how MiSiC can be further used to train networks to classify species in interspecies communities by generating two datasets, one per species of interest, to further train a U-Net. Here, the secondary U-Net was developed to specifically discriminate Myxococcus from E. coli, which is a very specialized application. Hence it was not included in the MiSiC package. Nevertheless the code is accessible at https://github.com/pswapnesh/MyxoColi (which is mentioned in the Methods).
Performance was compared between SuperSegger with default parameters and MiSiC with tuned parameters for a single data set. Perhaps other SuperSegger parameters would perform better with the addition of noise, and it's unclear that adding Gaussian noise to a phase contrast image is the best way to benchmark performance. An interesting comparison would be between MiSiC and other methods applying neural networks to unprocessed data such as DeepCell and DeLTA, with identical training/test sets and an attempt to optimize free parameters.
In fact, we believe that it does make sense to test how MiSiC performs in the presence of noise and show that it is robust, making it suitable for use on complex multi-tile images. For this analysis we kept the comparison with Superseger, which provides a reference as it is done on a data set optimized for Superseger segmentation. Importantly, we keep the parameters constant throughout the analysis because it would not be feasible to tweek parameters tile-by-tile in a multi-tile image. This analysis shows that MiSiC is more adapted for this application.
INSTALLATION: I installed both the command line and GUI versions of MiSiC on a Windows PC in a conda environment following provided instructions. Installation was straightforward for both. MiSiCgui gave one error and required reinstallation of NumPy as described on GitHub. Both give an error regarding AVX2 instructions. MiSiCgui gives a runtime error and does not close properly. These are all fairly small issues. Performance on a stack of images was sufficiently fast for many applications and could be sped up with a GPU implementation.
We have updated the pip install script available in GitHub for MiSiCgui that remediates some of these issues : There is no more numpy error, it closes properly and there are only warning messages concerning future deprecations in the napari packages. We have tested in Windows 10, Linux Ubuntu 18, and Mac OS Catalina. For the moment it seems impossible to install in Mac OS BigSur maybe due to the python 3.7 requirement. We will work on this problem in the near future. We have removed the command line interface as we are developing future version with an easiest way to provide MiSiC as Napari or FIJI/ImageJ plugin
TESTING: I tested the programs using brightfield data focused at a different plane than data presumably used to train the MiSiC network, so cells are dark on a light background and I used the phase option which inverts the image. With default settings and a reasonable cell width parameter (10 pixels for E. coli cells with 100-nm pixel width; no added noise since this image requires no rescaling) MiSiCgui returned an 8-bit mask that can be thresholded to give segmentation acceptable for some applications. There are some straight-line artifacts that presumably arise from image tiling, and the quality of segmentation is lower than I can achieve with methods tuned to or trained on my data. Tweaking magnification and added noise settings improved the results slightly. The MiSiC command line program output an unusable image with many small, non-cell objects. Looking briefly at the code, it appears that preprocessing differs and it uses a fixed threshold.
We thank the reviewer for testing the programs. Tiling related artifacts may now be avoided by excluding a few pixels at the border in the new version of MiSiC code. This is now implemented in the MiSiC.segment function as segment(im,invert = False,exclude = 16). Without seeing the reviewers data it is difficult for us to see how the segmentation (which is said to be acceptable) could be further improved. The command line program has now been removed in favor of continuous development on the graphical interface.
Reviewer #3 (Public Review):
The authors aimed to develop a 2D image analysis workflow that performs bacterial cell segmentation in densely crowded colonies, for brightfield, fluorescence, and phase contrast images. The resulting workflow achieves this aim and is termed "MiSiC" by the authors.
I think this tool achieves high-quality single-cell segmentations in dense bacterial colonies for rod-shaped bacteria, based on inspection of the examples that are shown. However, without a quantification of the segmentation accuracy (e.g. Jaccard coefficient vs. intersection over union, false positive detection, false negative detection, etc), it is difficult to pass a final judgement on the quality of the segmentation that is achieved by MiSiC.
We thank the reviewer for this comment. To address it we divided the previous Figure 2 into two figures (and associated supplemental figures) separately showing how MiSiC performs (i), to segment two very distinct bacterial species E. coli and Myxococcus under various imaging modalities. (ii) to segment other bacterial species: rods (P. aeruginosa), filaments (B. subtilis) and crescent shapes (C. crescentus). The results now clearly show both the strength and limitations of the system.
A particular strength of the MiSiC workflow arises from the image preprocessing into the "Shape Index Map" images (before the neural network analysis). These shape index maps are similar for images that are obtained by phase contrast, brightfield, and fluorescence microscopy. Therefore, the neural network trained with shape index maps can apparently be used to analyze images acquired with at least the above three imaging modalities. It would be important for the authors to unambiguously state whether really only a single network is used for all three types of image input, and whether MiSiC would perform better if three separate networks would be trained.
A single network is using a shape-index-map rather than the original images as an input. As mentioned by the reviewer this is a major strength of the workflow given that it permits segmentation, independent of the imaging modality, which we now measure for each modality.
As the reviewer hints, three different models specific to each modality (CP, Fluorescence and BF) could also be used to train three networks, allowing the direct end-to-end segmentation of raw images. In theory, this could improve the segmentation (although this might lead to negligible benefits given the actual segmentation quality).
-
Reviewer #3 (Public Review):
The authors aimed to develop a 2D image analysis workflow that performs bacterial cell segmentation in densely crowded colonies, for brightfield, fluorescence, and phase contrast images. The resulting workflow achieves this aim and is termed "MiSiC" by the authors.
I think this tool achieves high-quality single-cell segmentations in dense bacterial colonies for rod-shaped bacteria, based on inspection of the examples that are shown. However, without a quantification of the segmentation accuracy (e.g. Jaccard coefficient vs. intersection over union, false positive detection, false negative detection, etc), it is difficult to pass a final judgement on the quality of the segmentation that is achieved by MiSiC.
A particular strength of the MiSiC workflow arises from the image preprocessing into the "Shape Index …
Reviewer #3 (Public Review):
The authors aimed to develop a 2D image analysis workflow that performs bacterial cell segmentation in densely crowded colonies, for brightfield, fluorescence, and phase contrast images. The resulting workflow achieves this aim and is termed "MiSiC" by the authors.
I think this tool achieves high-quality single-cell segmentations in dense bacterial colonies for rod-shaped bacteria, based on inspection of the examples that are shown. However, without a quantification of the segmentation accuracy (e.g. Jaccard coefficient vs. intersection over union, false positive detection, false negative detection, etc), it is difficult to pass a final judgement on the quality of the segmentation that is achieved by MiSiC.
A particular strength of the MiSiC workflow arises from the image preprocessing into the "Shape Index Map" images (before the neural network analysis). These shape index maps are similar for images that are obtained by phase contrast, brightfield, and fluorescence microscopy. Therefore, the neural network trained with shape index maps can apparently be used to analyze images acquired with at least the above three imaging modalities. It would be important for the authors to unambiguously state whether really only a single network is used for all three types of image input, and whether MiSiC would perform better if three separate networks would be trained.
-
Reviewer #2 (Public Review):
Panigrahi and co-authors introduce a program that can segment a variety of images of rod-shaped bacteria (with somewhat different sizes and imaging modalities) without fine-tuning. Such a program will have a large impact on any project requiring segmentation of a large number of rod-shaped cells, including the large images demonstrated in this manuscript. To my knowledge, training a U-Net to classify an image from the image's shape index maps (SIM) is a new scheme, and the authors show that it performs fairly well despite a small training set including synthetic data that, based on Figure 1, does not closely resemble experimental data other than in shape. The authors discuss extending the method to objects with other shapes and provide an example of labelling two different species - these extensions are …
Reviewer #2 (Public Review):
Panigrahi and co-authors introduce a program that can segment a variety of images of rod-shaped bacteria (with somewhat different sizes and imaging modalities) without fine-tuning. Such a program will have a large impact on any project requiring segmentation of a large number of rod-shaped cells, including the large images demonstrated in this manuscript. To my knowledge, training a U-Net to classify an image from the image's shape index maps (SIM) is a new scheme, and the authors show that it performs fairly well despite a small training set including synthetic data that, based on Figure 1, does not closely resemble experimental data other than in shape. The authors discuss extending the method to objects with other shapes and provide an example of labelling two different species - these extensions are particularly promising.
The authors show that their network can reproduce results of manual segmentation with bright field, phase and fluorescence input. Performance on fluorescence data in Fig. 1 where intensities vary so much is particularly good and shows benefits of the SIM transformation. Automated mapping of FtsZ show that this method can be immediately useful, though the authors note this required post-processing to remove objects with abnormal shapes. The application in mixed samples in Fig. 4 shows good performance. However, no Python workflow or application is provided to reproduce it or train a network to classify mixtures in different experiments.
Performance was compared between SuperSegger with default parameters and MiSiC with tuned parameters for a single data set. Perhaps other SuperSegger parameters would perform better with the addition of noise, and it's unclear that adding Gaussian noise to a phase contrast image is the best way to benchmark performance. An interesting comparison would be between MiSiC and other methods applying neural networks to unprocessed data such as DeepCell and DeLTA, with identical training/test sets and an attempt to optimize free parameters.
INSTALLATION: I installed both the command line and GUI versions of MiSiC on a Windows PC in a conda environment following provided instructions. Installation was straightforward for both. MiSiCgui gave one error and required reinstallation of NumPy as described on GitHub. Both give an error regarding AVX2 instructions. MiSiCgui gives a runtime error and does not close properly. These are all fairly small issues. Performance on a stack of images was sufficiently fast for many applications and could be sped up with a GPU implementation.
TESTING: I tested the programs using brightfield data focused at a different plane than data presumably used to train the MiSiC network, so cells are dark on a light background and I used the phase option which inverts the image. With default settings and a reasonable cell width parameter (10 pixels for E. coli cells with 100-nm pixel width; no added noise since this image requires no rescaling) MiSiCgui returned an 8-bit mask that can be thresholded to give segmentation acceptable for some applications. There are some straight-line artifacts that presumably arise from image tiling, and the quality of segmentation is lower than I can achieve with methods tuned to or trained on my data. Tweaking magnification and added noise settings improved the results slightly. The MiSiC command line program output an unusable image with many small, non-cell objects. Looking briefly at the code, it appears that preprocessing differs and it uses a fixed threshold.
-
Reviewer #1 (Public Review):
In this work, Panigrahi et. al. develop a powerful deep-learning-based cell segmentation platform (MiSiC) capable of accurately segmenting bacteria cells densely packed within both homogenous and heterogeneous cell populations. Notably, MiSiC can be easily implemented by a researcher without the need for high-computational power. The authors first demonstrate MiSiC's ability to accurately segment cells with a variety of shapes including rods, crescents and long filaments. They then demonstrate that MiSiC is able to segment and classify dividing and non-dividing Myxococcus cells present in a heterogenous population of E. coli and Myxococcus. Lastly, the authors outline a training workflow with which MiSiC can be trained to identify two different cell types present in a mixed population using Myxococcus and E. …
Reviewer #1 (Public Review):
In this work, Panigrahi et. al. develop a powerful deep-learning-based cell segmentation platform (MiSiC) capable of accurately segmenting bacteria cells densely packed within both homogenous and heterogeneous cell populations. Notably, MiSiC can be easily implemented by a researcher without the need for high-computational power. The authors first demonstrate MiSiC's ability to accurately segment cells with a variety of shapes including rods, crescents and long filaments. They then demonstrate that MiSiC is able to segment and classify dividing and non-dividing Myxococcus cells present in a heterogenous population of E. coli and Myxococcus. Lastly, the authors outline a training workflow with which MiSiC can be trained to identify two different cell types present in a mixed population using Myxococcus and E. coli as examples.
While we believe that MiSiC is a very powerful and exciting tool that will have a large impact on the bacterial cell biological community, we feel explanations of how to use the algorithm should be more greatly emphasized. To help other scientists use MiSiC to its fullest potential, the range of applications should be clarified. Furthermore, any inherent biases in MiSiC should be discussed so that users can avoid them.
Major Concerns:
It is unclear to us how a MiSiC user should choose/tune the value for the noise variance parameter. What exactly should be considered when choosing the noise variance parameter? Some possibilities include input image size, cell size (in pixels), cell density, and variance in cell size. Is there a recommended range for the parameter? These questions along with our second minor correction can be addressed with a paragraph in the Discussion section.
Could the authors expand on using algorithms like watershed, conditional random fields, or snake segmentation to segment bacteria when there is not enough edge information to properly separate them? How accurate are these methods at segmenting the cells? Should other MiSiC parameters be tuned to increase the accuracy when implementing these methods?
Can the MiSiC's ability to accurately segment phase and brightfield images be quantitatively compared against each other and against fluorescent images for overall accuracy? A figure similar to Fig. 2C, with the three image modalities instead of species would nicely complement Fig. 2A. If the segmentation accuracy varies significantly between image modalities, a researcher might want to consider the segmentation accuracy when planning their experiments. If the accuracy does not vary significantly, that would be equally useful to know.
The ability of MiSiC to segment dense clusters of cells is an exciting advancement for cell segmentation algorithms. However, is there a minimum cell density required for robust segmentation with MiSiC? The algorithm should be applied to a set of sparsely populated images in a supplemental figure. Is the algorithm less accurate for sparse images (perhaps reflected by an increase in false-positive cell identifications)? Any possible biases related to cell density should be noted.
It is exciting to see the ability of MiSiC to segment single cells of M. xanthus and E. coli species in densely packed colonies (Fig. 4b). Although three morphological parameters after segmentation were compared with ground truth, the comparison was conducted at the ensemble level (Fig. 4c). Could the authors use the Mx-GFP and Ec-mCherry fluorescence as a ground truth at the single cell level to verify the results of segmentation? For example, for any Ec cells identified by MiSiC in Fig. 4b, provide an index of whether its fluorescence is red or green. This single-cell level comparison is most important for the community.
-
Evaluation Summary:
In this work, Panigrahi et. al. develop a powerful deep-learning-based cell segmentation platform (MiSiC) capable of accurately segmenting brightfield, fluorescence, and phase contrast images of bacteria cells densely packed within both homogenous and heterogeneous cell populations. This algorithm, if further optimized and disseminated to the community, will have a large impact in microbial studies in that it will allow for automated analyses of essential every aspect of bacterial cell biology, including cell morphology, cell cycle, cell-cell communications, protein localization dynamics and a variety of cellular processes using time-lapse imaging.
(This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the …
Evaluation Summary:
In this work, Panigrahi et. al. develop a powerful deep-learning-based cell segmentation platform (MiSiC) capable of accurately segmenting brightfield, fluorescence, and phase contrast images of bacteria cells densely packed within both homogenous and heterogeneous cell populations. This algorithm, if further optimized and disseminated to the community, will have a large impact in microbial studies in that it will allow for automated analyses of essential every aspect of bacterial cell biology, including cell morphology, cell cycle, cell-cell communications, protein localization dynamics and a variety of cellular processes using time-lapse imaging.
(This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #2 agreed to share their name with the authors.)
-
