ExoJ – a Fiji/ImageJ2 plugin for automated spatiotemporal detection and analysis of exocytosis

Exocytosis is a fundamental biological process that conveys sets of chemical information to the extracellular environment and delivers membrane proteins and lipids to the plasma membrane (PM) (Jahn and Südhof, 1999). Briefly, the process consists of the transport, docking, priming and fusion of intracellular compartments to the PM (Verhage and Sørensen, 2008). This eventually leads to the deposition of proteins and lipids into the PM, and the release of their luminal content to sustain important physiological functions and to respond to external stimuli. The variety of released cargo is vast, including neurotransmitters (Antonucci et al., 2012; Budnik et al., 2016; Filannino et al., 2024; Gundelfinger et al., 2003), antibodies and antigens (Buzas, 2023; Marcoux et al., 2021; Zitvogel et al., 1998), enzymes (Steffen et al., 2008), genetic materials (Crescitelli et al., 2013; Ehnfors et al., 2009; Fujita et al., 2016; Hazrati et al., 2022; Mathivanan et al., 2010; Zernecke et al., 2009), extracellular vesicles (EVs) (Bebelman et al., 2018; Minciacchi et al., 2015; van Niel et al., 2022), cytokines (Aiello et al., 2020; Lamichhane et al., 2015; MacKenzie et al., 2001; Mesri and Altieri, 1998; Szabó et al., 2014), signaling or adhesion molecules (Dequidt et al., 2007; Park et al., 2021), and receptors and transporters (Bakr et al., 2021; Guček et al., 2019; Heijnen et al., 1999; Jullie et al., 2014; Kassassir et al., 2023; Martinez-Arca et al., 2001; Park et al., 2006; Passafaro et al., 2001). Precise spatiotemporal control of exocytosis is thus essential for regulating diverse physiological and pathological processes. This also underlies the need to characterize the structural dynamics of the fusion machinery, the biochemical profile of the biological contents and the spatiotemporal dynamics of their release. In addition to a collection of biochemical approaches, methods using live-cell imaging of fluorescently tagged intracellular vesicles allow for spatiotemporal monitoring and quantitative analysis of content release (Ge et al., 2010). In particular, total internal reflection fluorescence microscopy (TIRFM) has become the imaging modality of choice owing to its inherent high signal-to-noise ratio (SNR) with a reduced phototoxicity and increased temporal resolution compared to widefield microscopy (Bebelman et al., 2020; Miesenböck et al., 1998). The TIRFM evanescent field of illumination enables the recording of events close to the PM, minimizing the disturbance from fluorescently labeled vesicles in the cytoplasm. Specific labeling of vesicle exocytosis is achieved by tagging the content and/or the vesicle membrane with a pH-sensitive fluorescent protein (FP) variant of GFP, known as pHluorin (Miesenböck et al., 1998; Sankaranarayanan et al., 2000). The FP pHluorin is quenched in the acidic vesicular lumen and brightens in the neutral extracellular environment upon vesicle fusion to the PM (hereafter termed a fusion event). When performing TIRFM, a fusion event appears as an abrupt brightening followed by spreading of the fluorescence signal (Miesenböck et al., 1998). Besides the vesicle-specific protein marker, the choice of the FP greatly influences the monitoring of the dynamics of exocytosis steps as evidenced by others (Liu et al., 2021; Martineau et al., 2017; Shen et al., 2014). Taken together, the subsequent amount of data generated by fluorescent time series poses the need of a robust analysis pipeline to identify fusion events in an unbiased manner. Manual annotation of each candidate event represents a time-consuming task and is prone to selection biases. This is particularly notable for features like the xy location of fusion events (potentially indicating intracellular hotspots), and the fluorescence intensity (estimating relative protein amount), the apparent size and the duration of the events. Because these characteristics exhibit significant variability between cells, achieving consistent measurements from one cell to another through manual analysis becomes challenging. Numerous algorithms have been developed to address these challenges, with varying degrees of user involvement throughout the detection workflow (Bebelman et al., 2020; Huang et al., 2007; Jullie et al., 2014; Mahmood et al., 2023; Moro et al., 2021a; Sebastian et al., 2006; Urbina and Gupton, 2021; Wang et al., 2018; Yuan et al., 2015). These solutions are, however, optimized for a particular population of vesicular cargoes and hence specific applications. Here, we present Project-ExoJ (hereafter named ExoJ; https://www.project-exoj.com), a solution developed as a Fiji/ImageJ2 (Schindelin et al., 2012) software plugin that automates the identification of fluorescently reported fusion events. We designed ExoJ to be fully user-configurable, with a graphical user interface (GUI) to set up a series of parameters that define a genuine population of fusion events according to the experimental conditions. To improve user experience, we provided tools for visualizing and further reporting features such as spatial location, intensity over time, apparent size and duration. To illustrate ExoJ capability, we focused on fusions events reported by tetraspanin (TSPAN) and vesicular soluble N-ethylmaleimide-sensitive fusion protein attachment protein receptor (v-SNARE) proteins coupled to pHluorin. Recordings of quantitative features revealed significant differences between and among labeled vesicle populations. Assessment of ExoJ performance using synthetic data revealed a highly robust and reliable identification tool that was insensitive to noise encountered in experimental settings.

At least three modes of exocytosis have been identified (full-collapse, kiss-and-run and compound exocytosis) and characterized according to their fusion dynamic patterns (Wu et al., 2014). Although each mode has its own fluorescence fluctuation pattern, the intensity decays to a certain extent right after vesicle fusion to the PM (Fig. S1A,B). This makes automatic recognition possible. Hence, we describe a fusion event as a transient diffraction-limited or large object that displays a sudden increase followed by an exponentially decreasing fluorescence intensity into the background. This fluorescence decrease is characterized by the mean lifetime decay τ (Fig. 1A). With this definition, we designed a processing pipeline broken into three main steps. Each step is managed in its own GUI dialog in which every parameter can be manually adjusted (Fig. 1). To facilitate parameter search optimization, the pipeline includes a ‘preview’ button that enables users to visualize the results for a specific set of parameters before committing to the analysis. Additionally, a ‘back’ button provides the flexibility to revert to previous steps and refine earlier parameter selections, ensuring a streamlined workflow (Fig. 1B–D, Table 1). These parameters can be saved and called in the pipeline to improve reproducibility.

The first step consists of detecting vesicles seen as bright spots from an image series (Fig. 1A,B, Table 2). Before performing spot detection, a custom photobleaching correction algorithm is applied to compensate for the variations in image intensity within fluorescent time sequences, as described in the Materials and Methods section. Although ExoJ does not include a masking algorithm, it provides flexibility for users to define specific regions of interest (ROIs) for analysis using either the drawing selection toolbox or by selecting previously saved ROIs from the Fiji built-in ROI manager tool (Fig. S1C). Users can choose to perform the spot detection on either the original movie or on its first-order differential movie (Transform to dF movie; Fig. 1B, Table 2). The dF movie highlights changes between frames by showing the difference between two consecutive frames from the original movie. ExoJ offers two detection algorithms: a wavelet-based method (default option) or a local maximum method (by unchecking the wavelet filter box; Fig. 1B). For the wavelet-based option, we employ the multiscale à trous wavelet transform algorithm (Olivo-Marin, 2002). To selectively keep user-specified sized spots from image series, images are convolved with a set of wavelet functions that are scaled according to the minimum and maximum spot size (minimum and maximum fusion event apparent size; Fig. 1B, see parameter 1). The resulting images are then decomposed into high- and low-frequency components. On the low-frequency component images, pixel values are now equal to wavelet coefficients C_wavelet as a result of wavelet transformation (Show wavelet lowpass image; Fig. 1B). The coefficient C_wavelet coarsely translates the similarity between sets of pixels and the user-defined wavelet functions. The C_wavelet value increases when there is a close resemblance between the intensity signal and the user-defined wavelet functions.

The detection result on single images can be previewed to ensure that proper fluorescence spots were accurately detected (Fig. 1B, Table 2). Once determined, the parameter is applied for the whole image series to extract the XY location of all fluorescent spots with a local maxima method within an adaptive window sized to ⁠.

The second step consists of building individual time-lapse trajectories of previously localized fluorescent spots (Fig. 1C). Although the starts of single trajectories are due to the appearance of a bright spot, the ends are not solely due to the fusion to the PM but could result from limitations in imaging conditions (e.g. low SNR and loss of focus). To account for this caveat, we combine a simple but yet sufficient multi-frame nearest-neighbor approach with a gap-closing algorithm (Chenouard et al., 2014; Crocker and Grier, 1996). Here, our plugin introduces three cut-off parameters that need to be tailored according to spot behavior, consisting of a spatial searching range, a temporal searching window and a minimal event size (Fig. 1C, Table 3). Spot size, direction and intensity are not considered during the frame-to-frame tracking process. Spots within the user-defined spatial and temporal searching range are assigned to the same trajectory, minimizing their global lateral displacement.

Various features have been considered to streamline the accurate identification of different types of fusion events in different cellular contexts (Bebelman et al., 2020; Diaz et al., 2010; Moro et al., 2021b; Sebastian et al., 2006; Urbina et al., 2018; Wang et al., 2018; Yuan et al., 2015). Here, we combine advantages of previous methods to define a versatile processing protocol. We reason that all fusion events display a statistically significant and transient fluorescence peak fluctuation F above the local background F₀, followed by an exponential fluorescence decay (Fig. 1D). Considering single spot trajectories, we first perform consecutive adjacent image subtraction to normalize the background fluorescence intensity (Jullie et al., 2014; Sebastian et al., 2006). This step results in a high SNR first-order differential fluorescent image series dF (Fig. 1D). For each candidate fusion event, we calculate σ_dF as the MAD of dF instead of solely considering the normalized peak change in fluorescence intensity ΔF/F₀ (Bebelman et al., 2020; Jullie et al., 2014; Moro et al., 2021a; Urbina et al., 2018). To refine the estimation of σ_dF, the trajectories of single candidate events are extended before (resp. after) the appearance (resp. disappearance) of the spot according to user entries (Expanding Frames; Fig. 1D). We eventually proceed with a moving linear regression on the fluorescence peak intensity profile to refine the onset time t₀ of candidate fusion events. This step helps considering fluorescence saturation and successive events at the same xy location. The algorithm also evaluates the maximal displacement of candidate events relative to their initial xy location at t₀.

Additional measurements are made by our plugin to describe the population of candidate fusing vesicles, including the mean lifetime τ which relates to the fusion dynamics and serves as a proxy for the fusion duration; the apparent size of the fusion event and the normalized peak change in fluorescence intensity (ΔF/F₀), which estimates the relative amount of fluorescently labeled proteins. In particular, the estimated apparent size reflects the entire fusion event, which could involve one or more vesicles or even larger compartments. Hence, it may not represent the size of individual vesicles within the fusion complex.

To account for various types of fusion and/or image series acquired under different experimental conditions, we integrate user-defined entries to modulate the definition of a genuine fusion event and hence the identification requirements of the algorithm (Table 4). In particular, candidate events with dF higher than kσ_dF (Detection threshold; Fig. 1D), limited displacements (Max. displacement), duration (Upper/Lower decay limit) and estimated size (Upper/Lower apparent size limit) at t₀ that comply with user inputs are seen as genuine fusion events (Fig. 1D). For the last two parameters, the goodness-of-fit, reported by the coefficient of determination R², is set as a threshold value (Min. R²) above which events are selectively kept (Fig. 1D, see parameters 8 and 9). Once the identification parameters are set, the plugin summarizes features of user-defined genuine events in a result table (Fig. S2A). We implemented a comprehensive suite of tools for visualizing, curing and exporting events, consisting of responsive pop-up windows (resp. histograms) for individual (resp. population-level) data of identified events, ensuring an expedited review process (Fig. S2) (see Materials and Methods for more details). In particular, the implemented tools automatically generate intensity (F), the first-order differential peak (dF), apparent size (full width at half maximum; FWHM) and mean radial intensity over time (spatial dynamics) profiles along with the detected event sequence (movie or montage). These visualizations dynamically adjust to the event selected in the result table.

To evaluate whether our tool can accurately detect fusion events, we used HeLa cells expressing pHluorin-tagged CD9, CD63 and CD81 TSPANs, which are commonly used to track extracellular vesicle (EV) exocytosis (Crescitelli et al., 2013; Kowal et al., 2016; Mathieu et al., 2021; Théry et al., 2018), and v-SNARE VAMP2 and VAMP7 proteins, which are components in the fusion machineries in many cellular contexts (Burgo et al., 2013; Gupton and Gertler, 2010; Han et al., 2017; Jahn and Scheller, 2006; Vats and Galli, 2022; Verderio et al., 2012; Wang et al., 2018) (Fig. 2A). To account for different image qualities and/or fusion reporter signal intensity (Fig. 2A), we iteratively refined the detection threshold values kσ_wavelet (Fig. 1B, Table 2), dF/σ_dF (Fig. 1D, Table 4) and the time window centered around candidate events at t₀ (Expanding frames; Fig. 1D, Table 4). During the identification process, we used the Preview button to optimize parameter search until we reached a plateau in the number of detected TSPAN- and v-SNARE-reported fusion events (Fig. 1B–D, Table 1). Although detection thresholds allow for distinct fusion intensity with respect to the local background, our algorithm explores a user-set time window to optimize the capture of fusion events of different dynamics. No further adjustments were made before running the identification process of each vesicle population in a batch-processed manner. We eventually reported a total number of 481 and 315 analyzed events for the TSPAN and v-SNARE population, respectively. These numbers represent the sum of events identified across all analyzed cells (Fig. 2B). Next, we used the ExoJ interactive visualization tools to review the features of each individual event and identify potentially missed events (see ‘Built-in tools’ section in Materials and Methods for more details; Fig. S2A,B). This in-depth review typically took ∼5 to 10 min per analyzed cell, depending on the number of detected events. Ultimately, we removed 10 TSPAN–pHluorin and 2 v-SNARE–pHluorin detection hits that did not correspond to fusion events but rather to filopodia tips coming in and out of focus, extracellular fluorescent objects and stationary vesicles.

The capacity to modulate the identification algorithm ensures that our plugin is adaptive in detecting distinct types of fluorescently reported fusion event with high accuracy.

Next, we investigated the potential of ExoJ for quantitative analysis of exocytosis. TSPAN CD9, CD63 and CD81 are known to form protein microdomains at the PM (Hemler, 2005; Le Naour et al., 2006). Recent studies have revealed that these TSPAN proteins are enriched to varying degrees in the membrane of EV subpopulations, including exosomes and ectosomes (Crescitelli et al., 2013; Escola et al., 1998; Kowal et al., 2016; Mathieu et al., 2021; Verweij et al., 2018). Exosomes are formed within late endosomes as intraluminal vesicles, and further released upon fusion with the PM. Their secretion can be monitored using pHluorin-tagged proteins (Fig. 2A). Previous studies suggest that TSPAN CD9 is primarily enriched in other EV subtypes that directly bud from the PM, making them undetected with our current approach (Kowal et al., 2016; Mathieu et al., 2021; Verweij et al., 2018). We examined post-fusion features of TSPAN subpopulations and compared them to transport vesicles reported with v-SNARE–pHluorin (Fig. 2A). In HeLa cells, the detection algorithm recorded a statistically similar fusion activity between individual reporter populations (Fig. 2B) with no correlation with the intensity and fusion apparent size (Fig. 2C,D). Despite TSPAN–pHluorin-reported events displaying a two- to three-fold higher fusion intensity than the v-SNAREs, our unbiased automatic detection method was able to capture fusion events reported by both, highlighting the effectiveness of our approach even in the presence of significantly varying signal intensity (Fig. 2A,C). The recorded difference among v-SNARE–pHluorin subpopulations is consistent with previous findings in various cellular contexts, where v-SNAREs VAMP2 and VAMP7 were observed to be differently localized in endosomes at different stages (Gupton and Gertler, 2010; Wang et al., 2018) (Fig. 2C,E). Furthermore, this difference in fusion intensity strongly suggests a difference in the number of fluorescently labeled protein reporters associated with each fusion event. The estimated apparent size of TSPAN–pHluorin events aligns with previous qualitative analysis at supra-optical electron microscopy (EM) resolution using a dynamic correlative light electron microscopy approach (Verweij et al., 2018) and falls within the range reported for v-SNARE-pHluorin events (Fig. 2D) (Altick et al., 2009). In addition, this closely matches early quantitative EM studies in the central nervous system (Roizin et al., 1967). It is important to consider that our system resolution limit is 246 nm when interpreting the apparent size results (see the ‘Practical considerations for live-cell imaging analysis’ below). Therefore, even if there are actual size differences in fusion events smaller than 246 nm, we would not be able to detect them. Focusing on the fusion duration, we first noted the short-lived fluorescence signal of v-SNARE–pHluorin, corresponding to either a rapid post-fusion lateral diffusion at the PM or an endocytic process (Alberts et al., 2006; Miesenböck et al., 1998; Urbina et al., 2018; Verweij et al., 2018) (Fig. 2E). We also found that subpopulations of TSPAN–pHluorin displayed a significantly distinct fusion event duration (Fig. 2E). The signal duration at the PM of CD81-pHluorin was almost two- and three-fold longer than that for CD63– and CD9–pHluorin, respectively. This could not previously be measured with this level of accuracy (Verweij et al., 2018). The divergence in fusion duration could not be explained by a noticeable difference in apparent size, nor a significantly higher amount of TSPANs (Fig. 2C,D). For CD63 and CD81, we hypothesized that the observed difference could either reflect the various regulatory fusion machineries and/or the heterogeneity in the properties of TSPAN-containing vesicles released by the cell, potentially reflecting differences in multi-vesicular body (MVB) fusion machinery or cargo composition (Edgar et al., 2016; Larios et al., 2020). Indeed, TSPANs are known to form discrete microdomains characterized by TSPAN–TSPAN interactions (typically less than 120 nm) and interactions with specific partner proteins (Charrin et al., 2009; Rubinstein et al., 1996; van Deventer et al., 2021). It is important to note that these observations do not definitively distinguish between MVB fusion and other fusion events involving TSPANs. Further investigation employing complementary techniques like immunoelectron microscopy or cargo-specific assays would provide a more conclusive classification of the released vesicles. Furthermore, the signal duration of CD9–pHluorin was similar to v-SNARE-labeled population of vesicles, as previously observed (Verweij et al., 2018). This suggests that CD9–pHluorin bursts of fluorescence might reflect post-fusion lateral diffusion over the PM or endocytosis, similar to what is seen for v-SNARE–pHluorin, rather than exosome secretion in most events.

Altogether, our computer vision-assisted tool enabled us to record and evaluate features of different types of cargo and/or vesicles, providing quantitative insights into post-fusion dynamics.

To fully assess the performance of ExoJ, we simulated movies with synthetic data corresponding to fluorescence signal of fusion events randomly distributed at the cell surface (Fig. 3A). In our simulation, we modeled a random number of events across a wide range of normalized intensity ΔF/F₀, apparent size and duration τ features including data collected in this study and in different cellular contexts and/or protein reporters from others (Persoon et al., 2019; Urbina et al., 2018; Verweij et al., 2018; Wang et al., 2018; Yuan et al., 2015). We also simulated events featured as non-relevant, exhibiting distinct signal decay behaviors (see Materials and Methods for details) (Fig. S3A). We additionally tested the influence of increasing Gaussian noise signal which coarsely recapitulates local background variation due to noise sources (Fig. 3A). ExoJ was able to accurately capture simulated events within the range of typical signal-to-noise ratio (SNR) observed experimentally, with an identification error rate (also equates to 100 – accuracy) as low as 1.5% up to 3%, which remains comparable to that presented by Urbina and colleagues (Urbina et al., 2018) (Fig. 3B). At each noise increment, we thoroughly adjusted the algorithm requirements by refining kσ_wavelet (step 1), dF/σ_dF (step 3), both goodness-of-fit R² and the time window to optimize event identification (Fig. 1D, Table 4). These parameters were optimized for three randomly selected simulated movies, saved and further applied to all remaining time series. Altogether, we spent on average 15 min for the analysis of each simulated movie. In response to incremental Gaussian noise, ExoJ detection capability was significantly degraded with an error rate up to 7±1.4% (mean±s.e.m.), which still performs at a higher level of accuracy under equal noise condition (Urbina et al., 2018). We further compared ExoJ to Automated Multivesicular Body Exocytosis (AMvBE), a manual detection in ImageJ and Fiji (Bebelman et al., 2020). We proceeded with the AMvBE-assisted manual detection of all simulated movies, which took from 30 to 40 min per movie depending on the incremental Gaussian noise condition. We optimized the identification parameters to maximize the capture of all relevant events. In simulated movies with typical experimental SNR, manual detection exhibited a significantly lower accuracy, with error rates ranging from 19% to 25%. Even when relaxing the detection criteria to improve capture rate, manual detection with AMvBE still worsened, reaching up to 34.2%±8.0% (see Materials and Methods for more details). This result demonstrated that ExoJ offered a time-efficient and accurate alternative to manual analysis method.

Error rate alone is an incomplete measure on simulated movies with disparate fusion activity. Thus, we introduced standard metrics such as sensitivity, precision, F₁ score and specificity which altogether score and provide a more balanced assessment on ExoJ ability to correctly detect events while accurately discarding those featured as non-relevant (see Materials and Methods for details). Results on simulated data without added noise demonstrated that ExoJ is a highly robust tool with scoring metrics above 99% (Fig. 3C), which bettered previously published algorithms (Sebastian et al., 2006; Urbina et al., 2018). Increasing the contribution of Gaussian noise significantly impaired the ability to correctly capture fusion events. On average, the sensitivity was down to 89.6% in the lowest condition. However, these captured events were accurately identified (mean precision of 97.8%) while avoiding those non-relevant (mean specificity of 97.3%) (Fig. 3C). Notably, the F₁ score decreased from 0.99 to 0.93, reflecting the trade-off between sensitivity and precision under increasing noise. Although the highest sensitivity, precision and specificity scores achieves 100%, the impact of increasing Gaussian noise signal was significant with a sensitivity, precision and specificity score as low as 66.7%, 80% and 75%, respectively. The reported lowest scoring results occurred with the strongest added Gaussian noise signal. In addition, there was no correlation between ExoJ scoring results and the ratio of simulated events (non-relevant versus relevant) featured per movie (Fig. S3B). The flexibility of ExoJ enabled us to relax the algorithm requirements to account for changes in imaging SNR without compromising its ability to effectively capture fusion events.

Finally, we asked whether the event detection procedure is reliable enough to handle biological variability between individual cells, particularly under the most challenging noise conditions. To mimic this, we took advantage of the random spatial distribution of Gaussian noise added to simulated timeseries. We added a high noise level with a sigma value of 1, followed by event identification using constant set of parameters previously optimized for single simulated movies. We repeated this procedure ten times, scored ExoJ performance at each round and reported the overall relative variation using the MAD. The scoring results were grouped into four categories based on the simulated fusion activity (see Materials and Methods). The average MAD values indicated a well-maintained performance with a range of variations of 3.1%–4.1%, 3.8%–6.9% and 4.5%–7.8%, for accuracy, F₁ score and specificity, respectively. Although the F₁ score and specificity exhibited variations up to 20% in some individual measurements under the challenging noise condition (σ_G=1), the reported average MAD values remained consistently low. This result suggested that ExoJ can handle biological variability as shown by the reported average MAD values of ExoJ performance.

Taken together, assessment of ExoJ performance revealed a robust tool in identifying fusion events insensitive to experimental noise.

Here we presented ExoJ, a computer-vision assisted tool for the automated identification and analysis of exocytosis events marked by a pH-sensitive probe. The identification algorithm was designed as a fully user-configurable processing pipeline, making it suitable for the detection and analysis of various forms of fusion events independently from the imaging condition. Built-in options were implemented for visualization and manual curation of event features.

We demonstrate that ExoJ can monitor fusion events reported by TSPAN– and v-SNARE–pHluorin proteins with high accuracy. Reporting of spatiotemporal features provided quantitative insights with unique details between and among subpopulations of vesicle exocytosis. Using simulated data covering a wide range of feature data, performance assessment underscored ExoJ as a robust and reliable identification tool. As it is, ExoJ usage has no specific constraints on vesicle exocytosis features including normalized intensity ΔF/F₀, duration of signal decay τ and apparent size. In response to changing conditions in imaging, the sensitivity of event detection could be adjusted by lowering the detection threshold value kσ right from the first step (Fig. 1B). This will increase the number of captured spots along with the processing time in the subsequent steps (up to 12 min per movie on average for +σ_G condition; Fig. 3A). We derived the duration of fusion events by evaluating the mean lifetime τ of fluorescence peak intensity profiles F fitted with single mono-exponential decay functions (Fig. 1A). However, fusion events can also exhibit fluorescence decay patterns that deviate from a single exponential function, as shown in, for instance, epithelial cells (Mahmood et al., 2023) and neurons (Hiester et al., 2017; Jullie et al., 2014; Roman-Vendrell et al., 2014). To address this limitation, ExoJ allows users to export the raw data to facilitate manual curation. This enables users to fit alternative mathematical models to their data. In contrast to fusion intensity and duration, the apparent size measurements reported by FWHM were, however, limited due to the resolution of TIRFM (Fig. 2D). This limitation can lead to underestimations of the actual size of fusion events due to several factors. Events with low fluorescence intensity or an incomplete capture within the field of view will have a smaller FWHM, underestimating their actual size. Additionally, elongated or irregularly shaped events can appear smaller in one dimension, and background noise and camera imperfections can introduce further inaccuracies in FWHM measurements. The advent of fast-imaging structured illumination microscopy and novel fluorescent markers offer promising avenues for overcoming these limitations (Huang et al., 2018; Li et al., 2015; Liu et al., 2021; Roth et al., 2020). Such advancements could enable the recording of the structural dynamics of exocytosis with unprecedented resolution, providing more accurate information about the size of fusion events. In combination with ExoJ, we are confident that these developments will certainly help providing insights into exocytosis-associated protein dynamics, and exploring machinery involved in EV biogenesis and secretion (Verweij et al., 2022).

Numerous bioinformatics tools on single or cross-platforms have previously been developed for specific applications which make fair comparison difficult (Bebelman et al., 2020; Diaz et al., 2010; Huang et al., 2007; Jullie et al., 2014; Mahmood et al., 2023; Moro et al., 2021a; Sebastian et al., 2006; Urbina and Gupton, 2021; Wang et al., 2018; Yuan et al., 2015). In addition, some image formats are not accepted, a file size limit could hinder the time of analysis and the feature extraction approach could vary. Our goal in developing ExoJ was to provide a common, yet robust computer-vision assisted solution regardless of the experimental condition. To this end, we designed a GUI-based tool as a plugin for Fiji/ImageJ2, a well-established platform for biological image analysis (Schindelin et al., 2012). As a result, ExoJ accepts all type of image formats and bit depths recognized by Fiji/ImageJ2. Prior to running ExoJ, if necessary, experimental data could be pre-processed using built-in filter toolboxes (see Materials and Methods for more details). In the last decade, supervised methods approaches have achieved great success in providing solutions for detecting subcellular structures in fluorescent microscopy images (Boland and Murphy, 2001; Hu et al., 2010; Johnson et al., 2015). In particular, machine learning approaches were successfully applied for spot detection upon a training phase (Jiang et al., 2007; Lin et al., 2019). Careful selection and label of training datasets is, however, an essential prerequisite, either manually or using other detection methods. Instead, we opted for a feature-based approach. This method relies on simple user-configurable features, such as fluorescently patterned vesicular spots, to achieve good performance, eliminating the need for extensive labeled training data (Basset et al., 2015; Smal et al., 2010). We illustrated the capacity of ExoJ to detect distinct fusion events, illustrated here by TSPAN– and SNARE–pHluorin, using a limited number of biological parameters (Fig. 1B–D). We based ExoJ detection on a proven wavelet transform algorithm (Lagache et al., 2018; Olivo-Marin, 2002; Püspöki et al., 2016; Ruusuvuori et al., 2010; Toonen et al., 2006; Yuan et al., 2015). We also introduced the statistical measure MAD to hone the capture of candidate fusion events, which could partly explain the performance difference with the wavelet-based tool from Yuan and colleagues (Yuan et al., 2015). Our detection algorithm also explored a user-set time window centered around fusion events, whereas previously published tools had it fixed or hardly modifiable (Mahmood et al., 2023; Urbina et al., 2018; Yuan et al., 2015). In our study, we directly compared ExoJ to AMvBE-assisted manual detection using simulated movies. Manual detection proved to be time-consuming and substantially less accurate than ExoJ (Fig. 3B). Even attempts to improve capture rate by relaxing detection criteria with AMvBE resulted in a further decline in accuracy. ExoJ offers functionalities to facilitate the identification of user-defined genuine events as well as implementing options for data visualization, manual curation and export (Fig. S2). Throughout the workflow, users can go back and forth to readjust parameters and preview results (Table 1). In addition to our study, ExoJ has recently demonstrated its capability in providing quantitative insights on the role of SNARE protein SNAP29 in CD63–pHluorin-labeled EVs in PC-3 cells imaged with a confocal spinning-disk microscope (Hessvik et al., 2023). We are convinced that ExoJ could become a standard tool for quantitative review of comparative studies of vesicle exocytosis in an unbiased manner.

ExoJ was developed as an ImageJ2 and Fiji (Schindelin et al., 2012) plugin for the automated detection and analysis of fusion events in 8- or 16-bit grayscale fluorescent image series. The plugin is published under the GPLv3 license. It uses ImageJ2 and Fiji capabilities to open a wide range of image formats using the plugin Bio-Formats (Linkert et al., 2010). The latest version of ExoJ as well as a tutorial can be found on the following website https://www.project-exoj.com. The source codes for ExoJ and for the generation of simulated movies was deposited in a GitHub repository: https://github.com/zs6e/excytosis-analyzer-plugin. To install the plugin, place the .jar file in the ImageJ2 or Fiji plugin directory. After a restart, ExoJ will be available in the ImageJ2 or Fiji plugin menu. ExoJ is compatible with ImageJ v1.53t or newer versions, and has been tested on Windows and MacOS platforms. To ensure smooth operation, ExoJ required a 64-bit operating system and Java 8. Although 4 GB of RAM is the minimum, that might limit ExoJ performance when working with large datasets (image files and/or time series). For optimal performance, users should consider increasing the allocated memory for Fiji within its settings. This option is accessible by navigating to Edit>Options>Memory & Threads… in the menu. To avoid potential plugin errors, we strongly recommend setting ‘.’ as the decimal separator. This can be changed in the ‘Region and Language’ format options for both MacOS and Windows platforms.

A series of built-in options are implemented to visualize and manage single and/or the population of fusing vesicles (Fig. S2). For each analyzed fusion event, users can visualize the corresponding sequence of cropped images as a movie (button ‘Movie’) or a montage of stills (button ‘Montage’) along with associated fluorescence peak intensity F [button Peak Intensity (F) Plot], first-order differential intensity dF [button (dF) Plot] and estimated apparent size (button ‘Size’) fitting plots. Note that a visual smooth effect could persist while generating montages of still frames. To avoid such effect, users can rely on the built-in ‘Make Montage’ function in ImageJ. We also record the fluorescence spreading over (button Spatial dynamics; Fig. S2A) as initially proposed by Bowser and Khakh (Bowser and Khakh, 2007). We allow the possibility to discard detected events using the appropriate button in the result table (Remove; Fig. S2A). Similarly, candidate fusion events can be manually added using the built-in drawing selection tools in ImageJ (button ‘Add’; Fig. S2B). Upon region selection around the potential event, the plugin will display the fluorescence peak intensity profile F for a user-defined timeframe. Right after pressing the button ‘Add’ in the profile window, the newly added event is immediately reviewed. The exact size and shape of region selection are not crucial for adding events. However, we strongly recommend drawing a precise region around the candidate event. This helps avoid capturing other nearby potential events that might occur within the same timeframe. Three options are made available to generate histograms for the temporal occurrence (button ‘Event Counts’), apparent size (button ‘Size Stat.’) and mean lifetime (button ‘Tau Stat.’) of fusion events. A spatiotemporal map distribution of fusion events is available upon selection in the result table (Fig. S2C), and can further be exported as an RGB image. We designed three action buttons for exporting the result table (Export current table), fluorescence peak intensity profile F (Export all intensity plots) and feature table (Export all Table) for all individual events.

We used TSPAN and v-SNARE proteins as fusion event reporters. Specifically, we focused on the TSPANs CD9, CD63 and CD81 (Escola et al., 1998; Kowal et al., 2016; Théry et al., 2018; Yáñez-Mó et al., 2015), and v-SNARE vesicle-associated membrane protein 2 (VAMP2; also known as synaptobrevin2) and vesicle-associated membrane protein 7 [VAMP7; also known as Tetanus neurotoxin-insensitive vesicle-associated membrane protein (7TI-VAMP)] (Chaineau et al., 2009; Galli et al., 1998) proteins coupled to pHluorin (Miesenböck et al., 1998). CD9–pHluorin, CD63–pHluorin and CD81–pHluorin plasmids were constructed as described previously (Verweij et al., 2018). VAMP2–pHluorin was a kind gift from Dr Timothy A. Ryan (Cornell University, USA). VAMP7–pHluorin construction corresponds to an improved version (Chaineau et al., 2008) previously described (Martinez-Arca et al., 2000) and further characterized (Burgo et al., 2012; Wang et al., 2018). We used the HeLa Kyoto cell line, originally obtained from ATCC and gifted to us by Dr Bruno Goud laboratory (CNRS UMR144, Institut Curie, Paris, France); these cells were cultured in DMEM (Gibco, Thermo Fisher Scientific) supplemented with 10% FBS (Perbio Sciences; HyClone), 100 U/ml penicillin G and 100 mg/ml streptomycin sulfate and 2 mM glutamax (Invitrogen, Thermo Fisher Scientific). Cells at 50–60% confluence were transfected using Lipofectamine 2000 reagent (Invitrogen) on either 35-mm glass bottom Petri dishes (Ibidi) scaled with 500 ng of TSPAN plasmids or 18-mm round glass coverslips deposited in 12-well plates scaled with 1 mg of VAMP plasmid. Glass coverslips were further mounted on a Chamlide EC magnetic chamber (LCI).

Prior to imaging, cell medium was replaced with homemade HEPES-buffered Krebs-Ringer as previously used by Danglot and colleagues (2012) or Leibovitz's L-15 solution (Gibco). HeLa cells were imaged 24 h after transfection on an inverted microscope (Axio Observer 7, Zeiss) equipped with a TIRF module and a top stage imaging chamber (Tokai Hit STX-CO2) ensuring a constant temperature at 37°C. All imaging experiments were carried out with a 100×1.46 NA oil objective (Zeiss), an air-cooled 488 nm laser line at 0.4–0.6% power, with a TIRF angle set for 300–400-nm penetration depth and an additional optovar 1.6× to reach a pixel size of 100 nm. We opted for a slightly deeper penetration depth to capture events potentially occurring slightly beyond the immediate membrane surface. Images were acquired with Zen Black (SP2.3, Zeiss) onto an electron-multiplying charge-coupled device camera (iXon, Andor) at a frame rate of 5 Hz. Fluorescent timeseries were saved as tiff or czi files. Fusion activity was defined as the number of detected events throughout the cell surface over the course of a time-lapse experiment, which was typically 3 min (Fig. 2B). Thus, experimental data in this study consisted in single fluorescent timeseries of 901 images. Cell surface measurement was carried out using ImageJ or Fiji segmentation tools (Schindelin et al., 2012). The number of analyzed cells (N) per labeled population is indicated.

The detection parameters are adapted to the different types of fusion event. To find optimal parameters, we analyzed three time series per subpopulation. We stopped the optimization process when the number of detected events reached a plateau, taking ∼45 min for each TSPAN and 1 h for each v-SNARE reporter. Once optimized, these parameters were saved (‘Save’ button; Table 1) and applied (‘Load’ button; Table 1) to all remaining time series within each subpopulation. This streamlines analysis, significantly reducing the average processing time per time series to ∼10 min. This time primarily reflects the computational demands of analysis, user review of detected events and verification of potentially missed events (see the ‘Built-in tools’ section for details; Fig. S2A,B). In detail, we set the detection threshold values kσ_wavelet to 4.0–5.0 (resp. 3.0) and dF/σ_dF to 4.0–5.0 (resp. 3.0–3.5) for fusion events reported by TSPAN–pHluorin (resp. v-SNARE–pHluorin). We imposed a minimal signal duration of two frames (Fig. 1D). We completed the detection of candidate events with additional frames corresponding to 1 s (resp. 3–4 s) before (resp. after) the maximal peak fluorescence at t₀ (Expanding frames; Fig. 1C and Table 4). For v-SNARE–pHluorin, we limited the temporal window to 0.6 s (resp. 2–3 s) before (resp. after) the maximal peak. The fluorescence profiles were fitted with a minimal number of five points and a minimal goodness-of-fit R² of 0.75 were set as the fitting threshold. The initial vesicle apparent size (range from 4 to 10 pixels; step 1; Fig. 1B) and the tracking parameters (step 2; Fig. 1C) do not vary between TSPAN and v-SNARE protein reporters. We reported the number of identified events n for each protein reporters as follows: CD9, n=129; CD63, n=200; CD81, n=152; VAMP2, n=232; VAMP7, n=83.

We first focused on three time series per condition and/or reporter proteins. The time taken for optimal parameter search depends on many factors such as memory allocation for ImageJ (see the ‘Installation and system requirements’ section), data size and image quality. Large datasets potentially contain more fusion events for analysis, which can extend the time for parameter search and analysis. Poorly resolved or low SNR images necessitate more relaxed detection threshold parameter (σ_wavelet, step 1), leading to longer computation times. To optimize the search of optimal detection parameter using ExoJ, users can leverage the extensive capabilities of ImageJ and Fiji for pre-processing fluorescent time series using. Existing plugins and built-in filters offer convenient options to address common image quality issues. Plugins like ‘rolling ball’ (Sternberg, 1983) and convoluted background subtraction (Peng et al., 2017) and bilateral filter (Chaudhury et al., 2011) can help reducing noise, background signal and non-uniform illumination. Additionally, the deep-learning tool CARE (Boothe et al., 2023) can be used for more advanced noise reduction. To account for photobleaching, ExoJ includes a custom-written tool (see the ‘Photobleaching correction’ section), which does not assume a specific bleaching behavior.

For highly motile cells, registering time series with plugins like TurboReg and StackReg helps mitigate spatial movement that could lead to detection errors (Thevenaz et al., 1998).

ExoJ offers increased flexibility in defining the analysis area. Users can either draw a custom ROI using the ImageJ or Fiji drawing tools, or select previously segmented areas saved in the ROI manager. This allows for focusing on specific cellular regions, excluding irrelevant spots like those on cell sides. Although defining the ROI may increase computation time during spot detection due to sorting and excluding irrelevant areas, it significantly speeds up subsequent steps (tracking and identification).

Users have the possibility to perform the spot detection on a first-order differential movie rather than the original one (dF movie; Fig. 1B, Table 2). To ensure that all pixels in the final dF movie have non-negative values, a constant offset is added, determined by finding the minimum intensity across all pixels after the difference is calculated.

To account for the limitations imposed by the diffraction limit on apparent size (FWHM) measurements, we strongly recommend measuring the point spread function (PSF) of the optical system. In our work, as described by Faklaris and colleagues (Faklaris et al., 2022), the PSF was measured to be 246±0.03 nm (mean±s.e.m.) at 488 nm. This information can be optionally entered in ExoJ workflow from step 1 (Theoretical resolution input box; Fig. 1B, Table 2). ExoJ will then compare the user-provided PSF value to the estimated size of each detected event. Events with an estimated FWHM smaller than the specified PSF value will be flagged in red within the ‘ExoJ: Detected exocytosis list’ table (Fig. S2A). This helps users identify potential limitations in size accuracy for these events.

ExoJ currently detects events with a single exponential decay pattern in their fluorescence intensity profiles (Fig. 1A; Fig. S1A,B). To account for events exhibiting fluorescence decay patterns that deviate from a single exponential function, we suggest lowering the min R² (Fig. 1D parameter and Table 4) input value, deactivating ‘the lower decay limit’ box (Fig. 1D parameter 5 and Table 4) and further exporting features tables for all detected events for manual curation (Export all table action button, Fig. S2A). These suggestions may allow ExoJ to detect a wider range of events but may also increase the number of false positive that require manual review.

To fairly assess the robustness of ExoJ, we generated 20 movies of 901 images. These movies simulated fusion events using randomly distributed gaussian-shaped pHluorin signal intensities on the cell surface, with a fixed pixel size of 100 nm, a frame rate of 5 Hz and no pre-processing filter applied. We simulated two groups of synthetic events, labeled as relevant and non-relevant according to the features we imposed. A series of parameters was defined to encompass distinct features of relevant fusion events including those approximated by users in this study as well as previous works (Persoon et al., 2019; Urbina et al., 2018; Verweij et al., 2018; Wang et al., 2018; Yuan et al., 2015). In detail, we imposed on the simulated events a random distribution of: (1) normalized fluorescence peak intensity ΔF/F₀ as low as 10% up to 200% above the local background (centered circle of 9-pixel radius) and (2) apparent size over a range of gaussian widths from 3 to 7 pixels. Events featured as relevant exhibited single exponential decay behavior with a mean lifetime τ from 1 to 50 timeframes (i.e. 0.2 s to 10 s). In contrast, we modified the transient nature of fusion events featured as non-relevant where the signal decay could be (1) 1-frame long, (2) long-lived and constant fluorescence peak intensity, or (3) a damped sine wave along with (4) a random spatial displacement. The maximal number of simulated relevant and non-relevant events was both fixed to 50, and randomly set for each movie. The simulated fusion activity eventually ranged from 1.2×10³ to 9.8×10³ µm⁻²·min⁻¹, which matches with the previously recorded frequency rate (Persoon et al., 2019; Urbina et al., 2018; Verweij et al., 2018; Wang et al., 2018; Yuan et al., 2015). The identification pipeline was optimized to capture simulated events labeled as relevant. To represent random noise, we added fluorescent intensity variations following a Gaussian distribution to the simulated movies. We achieved this using the built-in ImageJ function ‘Add Specified Noise’. We added noise intensity with standard deviations (σ_G) of 0.25, 0.5 or 1 times the standard deviation calculated for each frame of the simulated time series.

To compare ExoJ accuracy with manual detection, we used the AMvBE macro, a tool within Fiji for analyzing events in fluorescent time series (Bebelman et al., 2020). We optimized AMvBE identification parameters to maximize the capture of all relevant events. In details, we impose a time window for event correction of 5 frames, a duration of at least 1 frame after peak fluorescence, a minimal event size fixed to 0.2 mm, a rescue threshold fixed to 3 and a spot movement up to twice the FHWM. Notably, with increasing Gaussian noise added to the simulated movies, we relaxed the minimum goodness-of-fit R² value associated with the estimation of mean lifetime. This parameter can be adjusted within the macro itself. We lowered it as low as 0.5 for the noisiest condition (σ_G=1).

We performed statistical analysis using Prism 9.0 (GraphPad). For multiple comparisons between populations of labeled vesicles, a Kruskal–Wallis non-parametric one-way ANOVA was used to determine significance, followed by Dunn's post hoc test for multiple comparisons between populations of labeled vesicles and simulated movies with increasing Gaussian noise intensity. Statistical significance was considered for α=0.05. Data are reported as box and whisker plots. Boxes show the median±interquartile range and the whiskers go down to the minimum and up to the maximal value in each dataset. Median values are displayed in brackets below each box and whisker plot unless otherwise specified in the figure caption (e.g. mean values are used in Fig. 3B–D and Fig. S3B).

The authors thank the NeurImag imaging facility of the Institute of Psychiatry and Neurosciences of Paris (Inserm U1266) where the imaging experiments were carried out. NeurImag is part of GIS IBISA and the national infrastructure France BioImaging supported by the French National Research Agency (ANR-10-INBS-04). We thank Dr Liangyi Chen and Dr Yongdeng Zhang for sharing their MATLAB executable tool for exocytosis detection. We are grateful to Julie N'guyen for maintaining and providing HeLa cells. We thank Dr Sebastien Nola, Dr Cédric Blouin and Brett Davis for their insightful comments on the manuscript. Zeiss Confocal LSM 880 Elyra PS.1 equipped with a TIRF module was purchased with a Région Ile-de-France grant awarded to Lydia Danglot DIM Cerveau et Pensée project. Guillaume Van Niel would like to acknowledge Institut National du Cancer, Fondation ARC pour la Recherche sur le Cancer and the ANR for their financial support.

The peer review history is available online at https://journals.biologists.com/jcs/lookup/doi/10.1242/jcs.261938.reviewer-comments.pdf