The data sets of influenza infected human eHAP cells, BPAE cells (Invitrogen FluoCells Slide #1) and mouse kidney tissue (Invitrogen FluoCells Slide #3), shown and analyzed in (Figures 2, 4, 5, Supplementary Figure S1; Supplementary Figure S3) have been acquired with a Zeiss LSM880 confocal microscope, using a Plan-Apochromat 20X/0.8 NA objective. A sequential acquisition for DAPI (excitation 405 nm, detection in the range 420–462 nm), AlexaFluor 488 and AlexaFluor 568 (excitation 561 nm, detection in the range 570–615 nm) was used to acquire z-stacks with the total size up to 13 um, every 0.5 um. Pixel size was 0.20 um.

The beads used to show the segmentation workflow (Figure 1) were TetraSpeck™ Microspheres, 0.1 µm; images were acquired on a Zeiss Observer.Z1 using Micro-Manager (https://micro-manager.org/) software with a Hamamatsu ORCA-spark Digital CMOS camera, using a 63X/1.4 NA objective. Pixel size is 0.08 um.

The segmentation obtained by running the ZELDA protocols is achieved using scikit-image (van der Walt et al., 2014) (version 0.18.1) and SciPy (Virtanen et al., 2020) (version 1.6.3) modules for image processing in python.

The resulting measurements are handled as Pandas data frames (McKinney, 2011) (version 1.2.4) and plotted with Matplotlib (Hunter, 2007) (version 3.4.2). JupyterLab Version 3.0.14 was used to handle the result tables (pandas), calculate the jaccard index (scikit-learn (Pedregosa et al., 2011)), and plot the data (matplotlib). Additionally, the latest version of napari-zelda uses datatable package to handle results (https://github.com/h2oai/datatable).

ZELDA plugin for napari (“napari-zelda”) can be installed through the “Install/Uninstall Package(s)” menu in napari (napari contributors, 2019), and its interface can be added with “Plugins/Add dock widget”.

Alternatively, the installation can be done downloading the repository, navigating to it with the Anaconda prompt and using the command “pip install -e.” within the downloaded folder.

The plugin widgets have been created using magicgui (https://github.com/napari/magicgui), while the GUI plots included in the “Data Plotter” protocol are obtained with matplotlib.backends.backend_qt5agg (https://matplotlib.org/2.2.2/_modules/matplotlib/backends/backend_qt5agg.html).

The template for the plugin has been obtained from cookiecutter-napari-plugin (https://github.com/napari/cookiecutter-napari-plugin).

Once the user has selected a specific base protocol, a JSON file is used by the plugin to load the right widgets in the GUI.

The “Design a new Protocol” option saves the custom workflow as a list of widgets that will be sequentially loaded the next time that the newly created protocol is launched. It will be visible just after re-launching napari and ZELDA plugin.

ZELDA plugin for napari (“napari-zelda”) makes available to the end-user the segmentation, measurement, and “parent-child” relation of two object populations. It ultimately allows to plot the results and explore the data in the same Graphical User Interface (GUI).

The current version of the plugin includes three different “protocols” to ease the image analysis of 3D data sets. Each protocol is a set of individual steps (functions) that return images (as napari layers), or results (printed plots in .tiff or tables in .csv format).

The first protocol, called “Segment a single population” (Figure 1A), can be used to segment both the 2D or 3D data sets. The basic workflow of this protocol (Figure 1B) includes simple steps, such as Gaussian Blur, Threshold, and Distance Map, to identify the seed points for the subsequent segmentation of the objects of interest. The user can then set the “min dist” parameter in the “Show seeds” function to improve the accuracy of cell counting, before calling the “Watershed” segmentation (Figures 1C–H). The detected objects can eventually be measured and the results table automatically saved (Supplementary Figure S2A).

Similar workflows have been previously implemented in useful tools such as MorphoLibJ (Legland et al., 2016) and in the latest versions of CellProfiler (McQuin et al., 2018), although with the limitation of being exclusively applied to the 2D image analysis, or lacking an embedded and flexible 3D viewer. In contrast, ZELDA provides an integration of a basic 3D object segmentation workflow with napari 3D rendering GUI. Notably, in ZELDA the individual workflow steps are also accessible as single functions that can be optionally used, or fine-tuned individually, without having to restart the entire workflow from scratch.

The second protocol, “Segment two populations and relate” (Figure 2A), implements the segmentation of two populations of objects in parallel, using the same workflow described above, with an additional step that allows establishing the “parent-child” relation between the two object populations (Figures 2B–D).

To run reproducible image analysis with ZELDA, both described protocols include a “log” functionality that stores the parameters used at each step. The log is shown in the GUI and can be optionally saved as a .txt file, together with the other results (Supplementary Figure S2B).

Once segmented, measured, and optionally related two object populations, the “Data Plotter” protocol (Figure 2E) allows to load a result table, and plot histograms or scatterplots of the measured properties. The plots are shown directly in the napari GUI and can be automatically saved as images to a specific folder. This has the advantage of avoiding the employment of additional software for data visualization.

Given that ZELDA does not require any coding skill, life science researchers are hugely facilitated by the integration of multiple bioinformatics tools in a single GUI.

Computer scientists and developers continuously propose new algorithms to tackle biological problems that frequently require extensive coding skills. However, users might have the necessity to reproduce a specific published workflow (such as the one in Figure 1B), without knowing a scripting language or necessarily having any background in image analysis. We made this possible by implementing a method that allows the customization of the image analysis protocols available in ZELDA. Indeed, by simply running the fourth option called “Design a New Protocol”, a user can create a new custom protocol (Figure 3A). Every step of the base protocols is listed in a JSON database and the relative GUI widgets (used for the software layout) are available as ready-to-use modules to build personalized protocols. The different functions, such as threshold, gaussian blur or distance map etc., can be chosen in a drop-down menu at specific steps of the new protocol (Figure 3B). By using the saving option (Figure 3C), the JSON database will be automatically updated (Figure 3D), and the ordered series of GUI widgets will be available the next time that ZELDA plugin will be launched (Figure 3E). Once developed, custom protocols can be shared with the community using the “Import and Export Protocols” option (Figure 3F).

In order to assess the accuracy in the segmentation of 2D and 3D data sets, we compared the results obtained by the ZELDA plugin for napari with those generated by two of the most widely used software in the bioimage field: ImageJ and CellProfiler.

As 2D data sets, we used images of cells (Figure 4A) at a low confluence (∼30% of the field of view area) with a cytoplasmic staining to identify parent objects, and a second one for cellular organelles (children objects), with the final goal of correctly assign the organelles to the containing cell (parent-child relation).

Intriguingly, ZELDA performed almost equivalently to ImageJ (Figure 4B) in identifying parent objects (Jaccard index J = 0.987 ± 0.003), child objects (J = 0.920 ± 0.054), and in the parent-child relation (J = 0.993 ± 0.001). This means that, assuming ImageJ segmentation as ground truth (Figures 4G–I), ZELDA will correctly label the pixels of an organelle as belonging to the corresponding cell cytoplasm in 99% of the cases (Figures 4D–F).

However, the adherence with CellProfiler labelling was slightly less striking (Figure 4C) although this difference might be due to the many more parameters available in the CellProfiler GUI, such as the “declump method” in the “watershed” module etc., that haven’t been implemented in ZELDA GUI to keep the software interface and its utilization as simple as possible. Nonetheless, the agreement on the identification of the parent cytoplasms found with CellProfiler (Figure 4J) was around 88% of the pixels, while for both the child objects segmentation and the parent-child relation (Figure 4K–L) it was ∼82%.

Benchmarking the segmentation of 3D data sets has proven to be slightly more complicated, since not all the available modules in CellProfiler support the 3D data processing. For example, in version 4.2.1 the “smooth” module that operates a Gaussian blur filter, is available just for the 2D data pipeline, while another one has to be used for the 3D case. The same holds for morphological operations such as those executed by the “ExpandOrShrinkObjects”. Trying to circumvent this lack of interchangeable 2D/3D functions could result in a more elaborated and time-consuming construction of the CellProfiler pipeline. Conversely, the versatile protocols supplied with ZELDA (Figure 1; Figures 2A–D) allowed the 3D segmentation and parent-child relation in fewer steps and about twice quicker than the CellProfiler “Test mode” (Figure 5J).

We then analyzed a collection of z-stacks of mouse kidney glomeruli, as 3D data sets (Figures 5A–C). In this tissue, phalloidin staining (Figure 5B) was used for the identification of the glomerular structures, and DAPI staining (Figure 5A) to pinpoint the cell nuclei contained in each glomerulus. The resulting segmentation of the two populations and parent-child relation obtained by ZELDA (Figures 5D–F) were compared with the output of a CellProfiler pipeline which included solely the 3D data compatible modules (Figures 5G–I).

Unfortunately, the labelling agreement between the two software was reduced with respect to the 2D analysis. A performance comparison of the 3D segmentation revealed a variation of the Jaccard index across the z-stack, with maximum values typically around the mid-slice, where the staining intensity of the confocal microscopy data set was stronger (Supplementary Figure S3A). We then considered the maxima of the Jaccard index across the z-stacks (Figure 5K), assessing an accordance around 63% for the parent objects (Jaccard index J = 0.632 ± 0.033), 73% for the children (J = 0.735 ± 0.032), and of 64% for the parent-child relation (J = 0.643 ± 0.029) (Supplementary Figure S3B).

We further investigated the reason for the lack of agreement on 3D data sets labelling between ZELDA and CellProfiler, and found that the difference was due to the absence of 3D equivalents for some modules (e.g., the “ExpandOrShrink” morphological operations), or lack of a unique naming for the 2D and 3D version of the same method in CellProfiler (e.g., “Gaussian Blur”). Indeed, pre-processing the z-stacks with the ZELDA and proposing the resulting smoothed 3D data sets to CellProfiler, successfully increased the accordance in identifying parents, children, and parent-child relation above the 99% of the pixels (Figure 5L).

Therefore, ZELDA can represent a faster interactive alternative to CellProfiler for the exploratory analysis of 3D data sets.