A napari-plugin for clustering objects according to their properties.
Note: support for other layers besides the Labels layer as described below is currently only available in the development version. This note will be removed as soon as there is a new release.
Jump to:
For clustering objects according to their properties, the starting point is a grey-value image and another layer containing derived measurements in the .features property. Here are the supported layer types along with examples of what they may represent:
Labelslayer containing a label image representing a segmentation of objects.Pointslayer containing points representing centroid coordinates of objects.Surfacelayer containing a surface representing a segmentation of an object.Labelslayer containing a time-lapse label image representing tracking results, where each label number/color correspond to a unique track ID.
Check the examples data folder to learn how to load these data with code.
The label image should not contain objects with the label 0 as these objects cannot be separated from the background, which is 0 as well in many images and would lead into erroneous behaviour when performing the clustering.
For segmenting objects, you can for example use the Voronoi-Otsu-labelling approach
in the napari plugin napari-segment-blobs-and-things-with-membranes.
In case you have 2D time-lapse data you need to convert it into a suitable shape using the menu Tools > Utilities > Convert 3D stack to 2D time-lapse (time-slicer) (documentation).
The first step is deriving measurements from the labeled image and the corresponding pixels in the grey-value image.
Use the menu Tools > Measurement tables > Regionprops (scikit-image, nsr) to get to the measurement widget (documentation).
Select the image, the corresponding label image and the measurements to analyse and click on Run.
A table with the measurements will open and afterwards, you can save and/or close the measurement table.
At this point it is recommended to close the table and the Measure widget to free space for following steps.
You can also load your own measurements. You can do this using the menu Tools > Measurement > Load from CSV (nsr).
If you load custom measurements, please make sure that there is a label column that specifies which measurement belongs to which labeled object.
Make sure to avoid the label 0 as this is reserved for the background. Tables for time-lapse data need to include an additional column named frame.
Once measurements were saved in the labels layer which was analysed, you can then plot these measurements using the menu Tools > Visualization > Plot measurements (ncp).
In this widget, you can select the labels layer which was analysed and the measurements which should be plotted
on the X- and Y-axis. Click on Plot to draw the data points in the plot area.
Under advanced options, you can also select the plot type histogram which will visualize a 2D histogram. 2D histogram visualization is recommended if you have a very high number of data points.
If you choose the same measurement for the X and the Y axis, a histogram will be shown.
Under advanced options you will also find the checkbox determining whether not-selected data points should be hidden (shown in grey) or automatically clustered as another cluster.
You can manually select a region in the plot. To use lasso (freehand) tool use left mouse click, and to use a rectangle - right click. The resulting manual clustering will also be visualized in the original image. To optimize visualization in the image, turn off the visibility of the analysed labels layer.
Hold down the SHIFT key while annotating regions in the plot to manually select multiple clusters.
Manual clustering results can be saved by going to Tools > Measurement > Show table (nsr), and clicking on Save as csv.
The saved table will contain a "MANUAL_CLUSTER_ID" column. This column is overwritten every time different clusters are manually selected.
When you plot your time-lapse datasets you will notice that the plots look slightly different. Datapoints of the current time frame are highlighted in bright color and you can see the datapoints move through the plot while you navigate through time:
You can also manually select groups using the lasso tool and plot a measurement per frame and see how the group behaves in time. Furthermore, you could also select a group in time and see where the datapoints lie in a different feature space:
If you have custom measurements from tracking data where each column specifies measurements for a track instead of a label at a specific time point, the frame column must not be added. Check below how tracking data and features should look like.
The Points layer typically contains coordinates of objects of interest (for example, object centroids).
To get these coordinates, you can apply spot detection algorithms (check references here with scikit-image, here with pyclesperanto and here for the spotflow plugin) or, if you have segmentation results, use objects centroids as coordinates. This last approach can be done via Tools > Points > Create points from labels centroids (nppas) from the napari-process-points-and-surfaces plugin:
You can load object features to these points by assigning them to the .feature attribute of the Points layer, like this in Python:
points_layer.features = features_table
The number of rows in the table should match the number of points.
You can cluster these features using the same algorithms explained furhter down, or manually, and get points colored accordingly, like shown below:
Check also this notebook to learn how to load these data from code.
The Surface layer could contain a surface representing a segmentation result.
To generate this surface from a labeled image containing the segmentation results, a classical algorithm is the Marching Cubes. It is available in scikit-image, and you can also apply it via Tools > Surfaces > Create surface from any label (marching cubes, scikit-image, nppas) from the napari-process-points-and-surfaces plugin. Choose which label id number you want to turn into a surface and click on Run:
You will notice that the surface layer will be created.
You can derive quantitative measurements from the vertices of a surface via Tools > Measurement tables > Surface quality table (vedo, nppas) from napari-process-points-and-surfaces plugin:
Surface vertex measurements can be plotted and classified the same way with the plotter (Tools > Visualization > Plot measurements (ncp)):
Check this notebook to learn how to load these data from code.
The Labels layer can be also used to display tracking results.
These notebooks show you examples of how to load and format tracking features from Mastodon in a way compatible with napari-clusters-plotter.
For example, if you have a time-lapse labeled image where each label number represents a unique track ID, you can load tracking features to this Labels layer and use the plotter to cluster them. In the 'gif' below, the Tracks layer is NOT used for clustering, it is just shown along as a convenience. There is currently no support for the Tracks layer.
Check this notebook to learn how to load these data from code.
For getting more insights into your data, you can reduce the dimensionality of the measurements, using these algorithms:
- Uniform Manifold Approximation Projection (UMAP)
- t-distributed stochastic neighbor embedding (t-SNE)
- Principal Component Analysis (PCA)
- Non-linear dimensionality reduction through Isometric Mapping (Isomap)
- Multi-dimensional Scaling (MDS)
To apply them to your data use the menu Tools > Measurement post-processing > Dimensionality reduction (ncp).
Select the label image that was analysed and in the list below, select all measurements that should be dimensionality reduced.
By default, all measurements are selected in the box.
You can read more about parameters of both algorithms by hovering over question marks or by clicking on them.
When you are done with the selection, click on Run and after a moment, the table of measurements will re-appear with two additional columns representing the reduced dimensions of the dataset.
These columns are automatically saved in the labels layer and can be further processed by other plugins.
Afterwards, you can again save and/or close the table.
If manual clustering, as shown above, is not an option, you can automatically cluster your data, using these implemented algorithms:
- k-means clustering (KMEANS)
- Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN)
- Gaussian Mixture Model (GMM)
- Mean Shift (MS)
- Agglomerative clustering (AC)
Therefore, click the menu Tools > Measurement post-processing > Clustering (ncp),
select the analysed labels layer.
Select the measurements for clustering, e.g. select only the UMAP measurements.
Select the clustering method KMeans and click on Run.
The table of measurements will reappear with an additional column ALGORITHM_NAME_CLUSTERING_ID containing the cluster
ID of each datapoint.
Afterwards, you can save and/or close the table.
Return to the Plotter widget using the menu Tools > Visualization > Plot measurement (ncp).
Select UMAP_0 and UMAP_1 as X- and Y-axis and the ALGORITHM_NAME_CLUSTERING_ID as Clustering, and click on Plot.
The easiest way to install this plugin is to install the devbio-napari plugin collection. The napari-clusters-plotter is part of it.
-
Get a python environment, e.g. via mini-conda. If you never used mamba/conda environments before, please follow the instructions in this blog post first.
-
Create a new environment, for example, like this:
mamba create --name ncp-env python=3.9
- Activate the new environment via conda:
mamba activate ncp-env
mamba install -c conda-forge napari
Afterwards, you can install napari-clusters-plotter, e.g. via conda:
mamba install -c conda-forge napari-clusters-plotter
Error: Could not build wheels for hdbscan which use PEP 517 and cannot be installed directly
This can happen if you used pip for the installation. To solve this error, install hdbscan via conda before installing the plugin:
mamba install -c conda-forge hdbscan
ValueError: numpy.ndarray size changed, may indicate binary incompatibility. Expected 96 from C header, got 88 from PyObject
Similar to the above-described error, this error can occur when importing hdbscan through pip or in the wrong order. This can be fixed by installing packages separately through conda and in the following order:
mamba install -c conda-forge napari hdbscan
pip install napari-clusters-plotterContributions are very welcome. Tests can be run with pytest, please ensure the coverage at least stays the same before you submit a pull request.
Distributed under the terms of the BSD-3 license, "napari-clusters-plotter" is free and open source software
This project was supported by the Deutsche Forschungsgemeinschaft under Germany’s Excellence Strategy – EXC2068 - Cluster of Excellence "Physics of Life" of TU Dresden. This project has been made possible in part by grant number 2021-240341 (Napari plugin accelerator grant) from the Chan Zuckerberg Initiative DAF, an advised fund of the Silicon Valley Community Foundation.
If you encounter any problems, please file an issue along with a detailed description.