Neuronal synapses are important cellular communication hubs that transmit signals between neurons to enable brain function (Figure 1A). The transmission of the signal is mediated by chemical neurotransmitters that are synthesized in the presynaptic neuron and stored in the lumen of synaptic vesicles (SVs). When an action potential reaches the synapse, calcium enters the presynaptic terminal via voltage-gated calcium channels and triggers the exocytic fusion of release-ready SVs with the plasma membrane (Sudhof, 2013). The released neurotransmitter molecules rapidly diffuse through the synaptic cleft to activate postsynaptic receptors of the receiving neuron. Depending on its type, the neurotransmitter can either help (i.e. excitatory transmitters) or hinder (i.e. inhibitory transmitters) the postsynaptic neuron from firing its own action potential. At the presynaptic terminal, shortly after their fusion with the plasma membrane, SVs are recycled via endocytosis and refilled with neurotransmitter to sustain synaptic transmission (Soykan et al., 2016).

To study these highly dynamic processes that occur on timescales of milliseconds to seconds and to monitor the molecular components that orchestrate them, functional imaging tools are crucial. For instance, the exocytosis and endocytosis of SVs can be tracked using synapto-pHluorins, chimeric proteins comprising a pH sensitive variant of GFP genetically fused to the luminal domain of a SV membrane protein. As the pH of a SV under physiological conditions is acidic (∼pH 5.5) the fluorescence of pHluorin is quenched at steady-state (Miesenböck et al., 1998). Upon arrival of an action potential (e.g. electrical stimulation of a neuron) SVs release their neurotransmitter content by fusing with the plasma membrane, thereby exposing their lumen to the neutral pH of the extracellular medium (∼pH 7.4), upon which pHluorin becomes strongly fluorescent. Subsequent endocytosis of SV membrane proteins from the plasma membrane and rapid acidification of the endocytic vesicle lumen via the vesicular ATPase results in re-quenching of the fluorescence signal elicited from pHluorin molecules (Figures 1B,C) (Miesenböck et al., 1998; Wienisch and Klingauf, 2006; Kavalali and Jorgensen, 2014). Other assays for synaptic function include molecular sensors that detect the pre- or postsynaptic influx of calcium (Figures 1D,E) (Zhao et al., 2011), as well as reporters that monitor changes in membrane voltage or detect the released neurotransmitters (Lin and Schnitzer, 2016). These tools all have in common that they need to detect fast changes in fluorescence signals in small, often diffraction-limited structures to enable the visualization of specific synaptic functions.

The optical detection of synaptic fluorescence is in general performed on live cell or tissue cultures using widefield microscopy to capture as much signal as possible on large, fast and highly sensitive detectors with a diffraction limited resolution of ∼200 nm. In order to reduce bleaching and phototoxicity, the light exposure is kept to a minimum. This produces relatively noisy images with abundant out-of-focus signal and many cellular structures that are irrelevant for the downstream analysis (Figures 1F–H). The resulting time-lapse movies are thus hard to analyze via automated image analysis workflows that only rely on intensity thresholding and detection, often resulting in the analysis being performed manually (Balaji, 2007; Soykan et al., 2017), using custom scripts not easily usable for a wider research community (Chanaday and Kavalali, 2018), or (i.e. for calcium imaging) via advanced software toolboxes that require programming expertise (Kaifosh et al., 2014; Pachitariu et al., 2016; Giovannucci et al., 2019; Robbins et al., 2021) (Supplementary Table S1). To eliminate the need for tedious, time consuming and biased manual image analysis, the key challenge is to develop robust and easy-to-use software tools that do not require in-depth expertise in programming and image analysis. These tools need to be implemented rigorously, using scientific software engineering to enable easy, transparent and fully reproducible analysis of synaptic imaging data. In recent years the research community has recognized this need and more of such tools have become available for different applications. For example, EZcalcium has been developed for the analysis of calcium imaging data (Cantu et al., 2020). However, due to its implementation in MATLAB, EZcalcium and similar tools require expensive software licenses, thereby limiting their accessibility.

Here we present SynActJ an easy-to-use, entirely free and open source Fiji plugin (Schindelin et al., 2012a) combined with an R shiny app (R-Core-Team, 2020; RStudio-Team, 2020; Chang et al., 2021) that allows the fully automatic detection and analysis of synaptic activity in time-lapse movies. The segmentation parameters can be easily and conveniently adjusted and verified over an entire dataset after which the image analysis can be performed in batch. It is fully reproducible by allowing the saving and loading of the segmentation parameters. SynActJ is implemented as a Fiji plugin in Java and can thus be easily installed via a Fiji update site. For the necessary data processing and quality control of the extracted traces a convenient R Shiny app is provided as well as an example script for more detailed analysis and plotting.

The Fiji plugin can be accessed via the Cellular Imaging update site: https://sites.imagej.net/Cellular-Imaging/. The code for the Shiny app can be downloaded from github: https://github.com/schmiedc/SynActJ_Shiny. Documentation for using the plugin and the app can be found here: https://schmiedc.github.io/SynActJ/. A test dataset is made available on Zenodo: https://doi.org/10.5281/zenodo.5644945. The code is distributed free and open-source under the MIT license: https://github.com/schmiedc/SynActJ/blob/master/LICENSE. The source code of the Fiji plugin can be downloaded from github: https://github.com/schmiedc/SynActJ. Contact and support: https://forum.image.sc/u/schmiedc/.

Typically, live cell neuronal cultures are imaged using a widefield system with a large field of view. The tissue is illuminated with as little light as possible, to reduce bleaching and phototoxicity. The data sets, while rich in quantitative data, contain out-of-focus signal and many different biological structures that are irrelevant to the analysis (Figures 1F,G). For synapto-pHluorin imaging, upon applying field stimulation to the live neuron culture with an electric current, the synaptic boutons respond with a sharp elevated fluorescence signal corresponding to the exocytic fusion and subsequent release of neurotransmitter. Following stimulation, the signal rapidly drops back to baseline on a timescale of seconds as the pHluorin tag gets quenched when SVs are endocytosed and acidified (Figures 1H–H‴). These events of briefly elevated signal with relatively small responses over the baseline signal are hard to detect by eye and require that brightness contrast settings are set correctly for perceiving the specific signal responses (Figures 1I,J).

SynActJ achieves accurate automatic segmentation of active synaptic boutons (Figures 2A,B). It solves the segmentation problem by carving out the portion of the image signal that increases around the time of stimulation (Figure 2C). This is achieved by dividing the maximum intensity projection of 10 frames post-stimulation by the maximum intensity projection of the frames before the stimulation. This allows the program to selectively detect responding boutons within the image while removing any non-relevant signal from the remainder of the field of view (Figures 2E,F). In order to improve detection of weakly responding synapses an ImageJ implementation of the Laplacian-of-Gaussian (LoG) filter (Sage et al., 2005) is applied to enhance blob-like structures of a defined size range followed by a maximum detection to spot individual events within the diffraction limited resolution of ∼200 nm (Figure 2G). To create an accurate segmentation of the bouton area, an intensity threshold is applied. The detection and segmentation masks are then used in a marker-controlled watershed using the MorpholibJ Fiji plugin (Legland et al., 2016). This limits the segmentation to objects that have passed the detection threshold and enables the separation of touching detected structures (Figure 2H). The segmentation can be further refined by performing a size and object circularity filter. To quantitatively measure the background the maximum intensity projection of the movie is filtered with a large Gaussian blur and an intensity threshold is used to segment the broad signal of the tissue. The resulting mask of the field of view is finally inverted to create a segmentation of the background.

SynActJ provides a graphical user interface which enables the user to interactively adjust the parameters of the workflow via a preview (Figure 2D). Importantly the user has access to the entire dataset and, thus, is able to quickly verify and adjust the parameters on any of the movies the analysis is applied to. SynActJ ensures full reproducibility of the workflow by allowing the user to conveniently save, load and access all the processing settings. Once suitable segmentation parameters have been defined by the user, the entire image analysis can be executed on the dataset in batch, e.g. different movies of different stimulations or treatments (Figure 2D). The segmentation exactly outlines the area of the signal response after stimulation (Figures 2A,B,H) and the raw average signal is measured over time in this area. Since the segmentation is outlining the bouton, it is vital that any drift is corrected prior to segmentation using plug-ins such as Correct 3D drift (Parslow et al., 2014).

The resulting measurements are saved in the specified output folder and can then be processed easily using the R Shiny app provided (Figure 3A). The app allows rapid processing of the individual traces extracted by the image analysis and provides overview plots for quick and efficient quality control of the data. For each movie the number of ROIs, area of segmented structures and background are plotted (Figure 3B). Furthermore, for each movie the raw traces, raw background signals, as well as the average raw traces and background values are plotted and provided as a data table (Figure 3C). The raw values can be further filtered and processed in custom data analysis workflows.

The Shiny App also performs basic as well as advanced data processing on each individual movie. The average background value is subtracted from the average signal (Figure 3D). To compare the differences in peak responses to the stimulation, the background-corrected signal is surface-normalized by dividing the signal by the average value of the trace in the frames prior to stimulation (Figure 3E). To compare how fast the signal returns to baseline a further normalization to the peak of the trace is performed (Figure 3F). The result of this data processing is again provided as a data table accessible for further custom statistical analysis and plotting. For instance in order to characterize and quantitatively compare the rate of endocytosis one can fit an exponential decay function to the peak-normalized data and compute the time constant τ (tau: time it took the signal to return to 1/e of the peak) (Figure 3F).

To benchmark our workflow we applied SynActJ to an existing dataset that has previously been analyzed manually (Soykan et al., 2017). In this dataset, cultured hippocampal neurons expressing Synaptophysin-pHluorin were treated with DMSO or para-nitro-Blebbistatin (Kepiro et al., 2014), a specific inhibitor for Myosin-II, and stimulated with 50 APs (10 Hz, 5 s) to determine whether Myosin-II activity is involved in SV exo-/endocytosis. In line with the previously proposed role for Myosin-II at synapses (Chandrasekar et al., 2013; Flores et al., 2014; Yue and Xu 2014), para-nitro-Blebbistatin treatment led to reduced Synaptophysin-pHluorin exocytosis after stimulation, as well as delayed endocytosis. Our automated workflow was able to segment more boutons compared to the previously performed manual analysis, which was limited to about 20 ROIs for responding boutons and five background ROIs (Figure 4A). For further validation we asked how many of the manually selected ROIs also contain an automatically detected ROI. We found that 64% of the manually selected ROIs overlapped with the automatically detected ones (Figure 4B). We further determined which objects the automatic analysis missed in the manual dataset and found that many of those structures were large objects such as clusters of boutons or large uniform, potentially out of focus objects (Figure 4C). Overall the automatic segmentations of SynActJ seemed to be more restrictive in terms of object size and shape compared to the manual segmentation. To remove false positive automatic detections that by chance showed an increase around the stimulation frame, we removed traces in which the response was below two standard deviations of the variation of the background. Further, we removed traces with peaks much later than the stimulation frame. This filter operations removed 312 traces from the original set, leaving 6,623 traces from the automatic segmentation that conform to the necessary response upon stimulation.

The automatic analysis was able to reproduce the results of the manual analysis: Surface-normalized plots revealed a decreased exocytic response to stimulation (Figures 4D,D′) while peak-normalized signals displayed a delay in the return to baseline (Figures 4F,F′). The automatic analysis also showed a reduced peak signal (Figures 4E,E′) and a significant increase in the time constant for returning to baseline intensity upon para-nitro-Blebbistatin application (Figures 4G,G′). Interestingly, automatic analysis detected a lower peak signal than analysis by manual assessment. This likely is attributed to the fact our automated analysis workflow is superior with respect to the detection of weakly responding synaptic events, thereby increasing the accuracy of the analysis and reducing bias towards larger and strong responding structures (Figure 4C). This is particularly overt for the DMSO control, for which manual detection overall used much fewer active boutons (Figures 4E,E′). Finally, the variability in the extracted parameters is greatly reduced with the standard deviation dropping by a factor of 1.5–2.4 (Figures 4E,G; Tables 1, 2), further illustrating the increased precision of the automated approach.

To test if our plugin can be applied to other modes of functional imaging of synaptic activity, we further validated our approach on calcium imaging data. Calcium imaging suffers from similar challenges as those seen for Synapto-pHluorins such as relatively noisy and background-rich live images with fast, yet comparably small fluorescence signal changes over baseline (Figure 5A–C). We recorded time-lapse movies of cultured hippocampal neurons expressing the presynaptic calcium sensor Synaptophysin-RGECO and stimulated successively with 10, 20 and 40 APs (all at 40 Hz and 0.25, 0.5 and 1 s, respectively). Similar to the data from Synapto-pHluorin imaging experiments, we were able to rapidly produce accurate segmentations of active boutons in which calcium influx was revealed by RGECO (Figures 5A,B). Automated image analysis by SynActJ successfully extracted traces thereby enabling the detection of calcium spikes with overall results near-identical to those produced by manual analysis (Figures 5D,E), yet in only a fraction of the time.

Biological image data is notoriously difficult to robustly analyze automatically. Not only do the biological and experimental variability hamper computational analysis but also noise and background introduced by hard limitations in the acquisition of biological data. Thus, it is often easier for highly trained scientists to perform these image analysis tasks by hand. For analyzing synaptic activity another key challenge are the brief changes in fluorescent intensity, limited to many small structures throughout a large field of view with heterogeneous signal. This renders manual detection also biased towards strong and large structures that are easier to detect. The difficulty to manually detecting active boutons effectively also limits the number and completeness of detected events that can be analyzed (Soykan et al., 2017). Furthermore, manual selection of responding synaptic terminals using circular ROIs does not allow for precisely outlining the boutons, which is advantageous if there is a slight sample drift or tissue movement. However, this leads to the inclusion of some background area in the final quantification making the measurement less accurate.

The automated analysis of SynActJ efficiently overcomes these challenges. As an easy-to-use workflow it allows the robust and transparent application of well-tested classical image analysis on complex time-lapse movies. It is delivered as an easily installed Fiji plugin for the image analysis as well as a Shiny app that are centered on easy-to-use GUIs. The segmentation is interactive and can be fine-tuned using particle size and shape filters. Basic as well as advanced data analysis is provided via an R Shiny app. SynActJ thereby replaces the tedious manual analysis, which may be prone to user-dependent bias, increases as demonstrated the overall accuracy and precision of the analysis, and enables a fully reproducible data analysis by allowing a standardized ROI selection. Together with its compatibility with a wide variety of optical sensors, we believe SynActJ will become an important toolset for the analysis of synaptic activity and function.