DBlink: dynamic localization microscopy in super spatiotemporal resolution via deep learning

Single-molecule localization microscopy (SMLM) has revolutionized biological imaging, improving the spatial resolution of traditional microscopes by an order of magnitude. However, SMLM techniques require long acquisition times, typically a few minutes, to yield a single super-resolved image, because they depend on accumulation of many localizations over thousands of recorded frames. Hence, the capability of SMLM to observe dynamics at high temporal resolution has always been limited. In this work, we present DBlink, a deep-learning-based method for super spatiotemporal resolution reconstruction from SMLM data. The input to DBlink is a recorded video of SMLM data and the output is a super spatiotemporal resolution video reconstruction. We use a convolutional neural network combined with a bidirectional long short-term memory network architecture, designed for capturing long-term dependencies between different input frames. We demonstrate DBlink performance on simulated filaments and mitochondria-like structures, on experimental SMLM data under controlled motion conditions and on live-cell dynamic SMLM. DBlink’s spatiotemporal interpolation constitutes an important advance in super-resolution imaging of dynamic processes in live cells. DBlink uses deep learning to capture long-term dependencies between different frames in single-molecule localization microscopy data, yielding super spatiotemporal resolution videos of fast dynamic processes in living cells.

Single-molecule localization microscopy (SMLM) has revolutionized biological imaging, improving the spatial resolution of traditional microscopes by an order of magnitude.However, SMLM techniques require long acquisition times, typically a few minutes, to yield a single super-resolved image, because they depend on accumulation of many localizations over thousands of recorded frames.Hence, the capability of SMLM to observe dynamics at high temporal resolution has always been limited.In this work, we present DBlink, a deep-learning-based method for super spatiotemporal resolution reconstruction from SMLM data.The input to DBlink is a recorded video of SMLM data and the output is a super spatiotemporal resolution video reconstruction.We use a convolutional neural network combined with a bidirectional long short-term memory network architecture, designed for capturing long-term dependencies between different input frames.We demonstrate DBlink performance on simulated filaments and mitochondria-like structures, on experimental SMLM data under controlled motion conditions and on live-cell dynamic SMLM.DBlink's spatiotemporal interpolation constitutes an important advance in super-resolution imaging of dynamic processes in live cells.
Spatial resolution in standard optical microscopes is bounded by the diffraction limit at about half the wavelength of light, corresponding, in the visible range, to ~200-300 nm.Super-resolution microscopy (SRM) methods overcome this limitation and enable higher resolution.Notable methods of this family include stimulated emission depletion (STED) 1 , structured-illumination microscopy (SIM) 2 as well as SMLM 3 .Prominent variants of SMLM include photoactivated localization microscopy 4 (PALM), stochastic optical reconstruction microscopy 5 (STORM), points accumulation for imaging in nanoscale topography 6 (PAINT) and DNA-PAINT 7 .SMLM variants differ in their experimental conditions; however, they share a similar overall pipeline.First, fluorescent molecules are used to label structures in a specimen.Then, a sequence of frames is captured, in which only a sparse, random subset of molecules emit light per frame.Subsequently, each emission event is detected and fit to a model of the system point spread function (PSF), allowing highly precise determination of the emitting fluorophore position.Finally, by accumulating the localizations of thousands of emitters, the output of SMLM is a single super-resolved image of the structure, typically with an order of magnitude resolution improvement compared with the diffraction limit.
An inherent limitation in SMLM is its temporal resolution.Accumulating a large enough number (typically millions) of single-molecule emission events to generate a continuous image takes a long time.Moreover, densely labeled structures present a challenge in the detection Article https://doi.org/10.1038/s41592-023-01966-0Recently, methods utilizing both spatial and temporal information have emerged [18][19][20] .Super-resolution radial fluctuation (SRRF) 19 measures the image intensity gradients in subpixel accuracy and detects the imaged structure by searching gradient convergence points.Since single molecules often blink over multiple frames, additional pixelwise cross-correlation analysis helps distinguish adjacent fluorophores.Recently, enhanced SRRF (eSRRF) 21 has enabled live-cell three-dimensional (3D) video reconstruction at spatial resolution of ~70 nm and temporal resolution of ~1 reconstructed volume (20 × 20 × 3.6 μm 3 ) per second.Although eSRRF achieves a substantial improvement in the spatiotemporal resolution, it does not exploit long-term temporal correlations, which contain valuable structural information.
In this paper, we present DBlink, a new method that increases the spatiotemporal resolution in the reconstruction of live-cell dynamic SMLM data.We use a bidirectional convolutional neural network (CNN) combined with long short-term memory (CNN-LSTM) that receives as input a video containing super-resolved localization maps and outputs a video of a dynamic super-resolved structure (Fig. 1).The super-resolved localization maps can be obtained by using existing methods, for example, Deep-STORM 12 or ThunderSTORM 22 .To perform spatiotemporal interpolation, DBlink relies on long-term interframe structural-correlation, and on previous information regarding the imaged sample-namely, its type (for example, microtubules, mitochondria, and so on).We first demonstrate the ability of DBlink to of single emitters, which limits the maximal number of localizations per frame.Indeed, the typical temporal resolution in SMLM is on the order of minutes, while resolutions of tens of seconds have also been reported 8,9 .A possible way to increase temporal resolution in SMLM is by increasing the illumination intensity while decreasing frame acquisition time 10 .However, high illumination intensity is inherently incompatible with live-cell imaging due to phototoxicity 11 .Recent advances in deep-learning algorithms have yielded computational algorithms that further improve the capabilities of SMLM.ANNA-PALM 8 significantly reduces the number of frames needed for super-resolution reconstruction.Deep-STORM 12,13 as well as DECODE 14 enable researchers to analyze densely labeled SMLM experiments by training a neural network to perform multi-emitter fitting in super-resolution.Importantly, while these algorithms perform exceptionally well in visualization of nanoscale structures, they are still applicable mostly to the analysis of static data or processes with very slow dynamics.
Existing methods are useful mostly for static data because the typical localization-based reconstruction process does not exploit structural-correlations over long periods of time (longer than the temporal window being reconstructed).Non-SMLM-based super-resolution methods [15][16][17] , which exhibit naturally high temporal resolution, either compromise on spatial resolution, or do not exploit the abundant super-resolved information of single-molecule emission events.A method that combines both spatial and long-range temporal interpolation would be optimal.

Results
Our goal is to extend the temporal resolution of SMLM beyond its inherent limitation, dictated by the trade-off between emitter density and localization ability, while maintaining high spatial resolution.Conceptually, the problem at hand is spatiotemporal interpolation of a 3D manifold (two-dimensional (2D) + time) that describes the continuous movement of a 2D object, from noisy discrete samples in space (2D localizations) and time (due to the camera acquisition rate).Clearly, there is insufficient information per frame, and there are multiple valid solutions, mathematically.The strategy we chose here is to train a neural network with realistic regularization based on previous knowledge on the imaged sample.
Injecting relevant previous knowledge to our network requires thousands of videos containing ground truth information at super spatiotemporal resolution.Experimental acquisition of this amount of data is extremely challenging; furthermore, ground truth information is not available at high spatiotemporal resolution.Therefore, we took the approach of simulating the training videos based on relatively simple static biological models and applying to them time-varying affine transformations.The challenge in simulation-based training is to generate simulations that resemble experimental data; in our case, the biological models facilitated this step, and we only had to tune several experimental parameters (for example, blinking density, motion velocities, and so on) to generate realistic simulated videos.To capture long-term dependencies between different video frames, we employ an LSTM network architecture, which has previously proven itself as a good solution for this task 23,24 .LSTM networks are suitable for sequential analysis tasks since each layer is designed to carry information from previous frames throughout the sequential analysis of the video.Because our method analyzes experiments in retrospect, we also possess information from future frames; therefore, we use a bidirectional LSTM network, and concatenate the forward pass and the backward pass.As expected, this strategy boosts the performance in comparison with a one-directional LSTM network architecture (LSTM evaluation; Supplementary Video 15).Each LSTM cell consists of a convolutional layer aiming to capture local spatial correlations between adjacent pixels.Each output element of the CNN-LSTM network is analyzed by an additional CNN that provides the final reconstructed frame (Supple mentary Fig. 1).In the following example applications, we show that this approach is feasible and produces high quality results.Article https://doi.org/10.1038/s41592-023-01966-0 In addition, the recurrent part of the network is not sensitive to different temporal blinking patterns; hence, we obtain clean and uniform structures in the reconstructions.First, we tested our approach on simulated data (Fig. 2).To do so, we generated simulated filaments according to the model of Shariff et al. 25 (Methods); then, we shifted and rotated them randomly, while maintaining structural smoothness and continuity in time (Filament simulation).The neural network was able to reconstruct random shifting structures over time with high accuracy (Supplementary Video 1), namely, 90% of the simulated binary map matched the predicted structure, and only 1.3% of the predicted structure was hallucinated (Reconstruction accuracy).In this simulation, the temporal resolution corresponds to one reconstructed frame per ten simulated blinking frames.
To quantify the spatial resolution of the reconstruction, we performed Fourier ring correlation (FRC) analysis 26 as well as decorrelation analysis 27 between the network reconstruction and a STORM reconstruction on a static sample.The network result was consistent with standard STORM reconstruction using Deep-STORM 12 up to a resolution of 29 nm according to FRC and up to 30 nm according to decorrelation analysis (Spatial resolution quantification).Next, as a first validation of our method on experimental data, we reconstructed a static structure that was shifting laterally over time.For this, we captured a STORM experiment of fixed microtubules exhibiting naturally occurring lateral sample-drift.We estimated the drift using the Deep-STORM drift correction mechanism, which is based on cross-correlation, and obtained a total shift of 240 nm in the y direction and 400 nm in the Fluorescent beads (red arrows) served as fiducial markers reporting on the rotation.At some point, rotation was stopped, and a static STORM video was captured for 15,000 more frames.This video was used to produce a ground truth static structure via Deep-STORM.The static structure was then compared with each frame in the dynamic reconstructed video, rotated appropriately.Scale bar, 2.5 μm.

Article
https://doi.org/10.1038/s41592-023-01966-0 x direction (Fig. 3a).Next, we used the localization maps provided by Deep-STORM and summed them over windows of 100 frames with 50 ms acquisition time.Finally, we input the summed localization maps into our network and obtained a super spatiotemporal resolution reconstruction of the shifting data at a temporal resolution of 0.2 frames per second (fps) (Supplementary Video 2).We predicted the drift according to the cross-correlation between the first reconstructed frame of our network and every other frame in the reconstructed video.The mean distance between our drift prediction and Deep-STORM prediction over the course of the experiment was 38 nm (Supplementary Fig. 3).Notably, the network did not have any previous knowledge that the sample was static and drifting, namely, that the only motion was a global shift; rather, the network treated this data the same as general dynamic data.A global motion prior would improve the performance significantly, at the cost of a less generalized solution.
To demonstrate DBlink's performance on a more complex type of motion than lateral shift, while still possessing knowledge of the sample structure to serve as validation, we captured a STORM video of static microtubules while rotating the camera manually (Fig. 3b).
We added to the sample fluorescent beads to serve as fiducial markers reporting on sample rotation.At the end of the experiment, we stopped rotating the camera and let the blinking continue for ~15,000 more frames, from which the ground truth structure was obtained using Deep-STORM.To test reconstruction performance, we computationally rotated back the predicted structure in each frame according to the calculated rotation angle.Then, we compared the computationally rotated video with the static reconstruction obtained by Deep-STORM (Supplementary Video 3).To quantify the prediction error, we measured the consistency between the reconstructed video frames based on the cross-correlation between every two frames in the reconstructed video, achieving a mean consistency score of 0.91, which indicated a highly consistent reconstruction (Reconstruction accuracy quantification).In this experiment, we used a window size of 40 summed frames to generate the reconstructed video, with acquisition time of 20 ms per frame, resulting in temporal resolution of 1.25 fps.
For the validation of DBlink on experimental data that contained dynamic morphological changes, and for which we could also possess ground truth information at super-resolution, we reconstructed Localizations within the temporally moving window serve as the input to DBlink, which provides as output the reconstruction of the dynamic simulated filaments at super spatiotemporal resolution.c, Overlay of the static structure reconstruction (blue) and the dynein motor dynamic reconstruction (white).The edges of the temporally moving window are marked in yellow.Temporal resolution, 50 s (25 frames with intervals of 2 s).

Article
https://doi.org/10.1038/s41592-023-01966-0dynamic motion of dynein motors moving on static microtubules 28 (Supplementary Video 4).The static microtubule reconstruction served as the structural ground truth in super-resolution for the traces that the dynein motors could track.To simulate moving filaments, instead of blinking molecules, we summed the localization of dynein motors over short spatiotemporal windows.For independent reconstruction validation, we used different algorithms for the static localization (ThunderSTORM) and for the dynamic DBlink localization-input (Deep-STORM).DBlink reconstruction of the dynein dynamics agreed with the static reconstruction of ThunderSTORM spatially (Fig. 4 and Supplementary Video 5).Quantitatively, the hallucination percentage of DBlink's reconstruction relative to ThunderSTORM's reconstruction was 0.1 % (see the Reconstruction accuracy quantification section in the Supplementary Information).To validate the temporal aspect of the reconstruction, we marked the edges of the temporal window containing the input localizations for DBlink.
Additional quantification of performance as a function of deviations between training data and testing data is reported in Additional performance quantification in the Supplementary Information.
Next, we tracked microtubule and ER dynamics in live cells.Because ground truth information is not available in this case, we compared our network reconstructions to four alternative solutions: (1) Deep-STORM reconstructions based on short temporal windows (Supplementary Video 6); (2) Deep-STORM reconstructions combined with a previously reported blind inpainting algorithm 9 ; (3) DECODE 14 reconstruction (Supplementary Video 7); (4) eSRRF 21 reconstruction (Supplementary Video 8).The input to DBlink was the sum of localizations over windows of 40 frames for the ER experiment and 20 frames for the microtubule experiment.Notably, the reconstructed video was at the same temporal resolution as the input video, which was achieved by using overlapping windows with one frame shifts.While this comes with a price of longer inference time ranging from a few minutes up to a few hours, the result was a super-resolved video at the temporal resolution of a single blinking frame.Importantly, the same exact network was used to recover both the ER data and the microtubule data, with no retraining; this serves as a demonstration of the generalizability of DBlink beyond its training data.The decision on the window sizes for each reconstruction algorithm is described in the Window size optimization section in the Supplementary Information.Qualitatively, DBlink reconstructions consistently outperformed the other methods.Although blind inpainting has managed to filter most of the noise in the Deep-STORM data, it performed poorly in densely labeled areas.Furthermore, rapid dynamics caused motion blur in temporally windowed Deep-STORM and DECODE reconstructions (Fig. 5), while DBlink provided a more stable reconstruction in areas exhibiting rapid motion.Moreover, eSRRF managed to overcome motion blur and provide a stable reconstruction; nonetheless, achieving the same temporal resolution as DBlink with eSRRF entailed losing some structural information in the reconstruction (Supplementary Video 8).Notably, the temporal resolution achieved by DBlink in these experiments was 15 ms (66.6 fps), compared with eSRRF temporal resolution of 1.5 s and 3.75 s (0.66 and 0.26 fps) for the microtubule and the ER experiments, respectively.The spatial resolution of DBlink reconstructions measured by decorrelation analysis was 30 nm.
To enable the evaluation of the quality of the reconstruction, we supply DBlink's confidence measure as an additional explainable output (Fig. 5c).The confidence map highlights areas in which the network reconstructions are more likely to capture or miss the correct structure.To validate the usefulness of the confidence map, we measured the hallucination percentage in low confidence areas in comparison with high confidence areas.For more details, see Confidence hallucination calculation in the Supplementary Information.While the hallucination level is very low in both cases, the mean hallucination percentage was 1.5% in low confidence areas and 0.1% in high confidence areas, indicating that high confidence regions indeed correlate with fewer errors in the reconstructed frame.

Article
https://doi.org/10.1038/s41592-023-01966-0 Next, we reconstructed the dynamics of mitochondria in live cells from high-density single-molecule data recorded using cells labelled for the mitochondrial protein COX8.We extended the observation time in live-cell imaging by using a HaloTag7 fusion in combination with a noncovalent, weak-affinity fluorophore tag that binds to and unbinds from the target and acts as an exchangeable fluorophore label 29 (Supplementary Videos 9 and 10).In this case, training required a model for mitochondrion size and shape, labeling density, motion type and speed, and so on.For this purpose, we developed a dynamic simplified-mitochondria simulator (Mitochondria simulation).After training the neural network, we analyzed a SMLM video of live-cell imaging data of mitochondria labeled with HaloTag7 (Methods) and followed the structural dynamics (Fig. 6).We could clearly visualize morphological changes of mitochondria, including fusion, fission and drift, at different velocities.Similarly to the filament experiments, the reconstructions of DBlink and eSRRF agreed spatially; however, DBlink's superior temporal resolution resulted in substantially less motion blur compared to eSRRF.The temporal resolution we achieved in this experiment was 20 fps and the spatial resolution was 75 nm, determined by decorrelation analysis 27 .
DBlink managed to reproduce mitochondrial dynamics at high fidelity (Fig. 6), despite the fact that the training data contained simple pixelated-structures and motions (Supplementary Video 11).This validates the generalizability and applicability of our network for the analysis of various biological samples, contingent on appropriate training.The structural changes observed were similar to previously published work on mitochondrial dynamics 30 , extended by faster temporal resolution, longer observation time and additional subdiffraction structural information.Previous work on live-cell STED microscopy of mitochondrial dynamics has shown the potential of long-term observations of mitochondria, yet at the expense of extended time periods for fluorophore recovery that limits the temporal resolution 31 .DBlink in combination with exchangeable fluorophores increases the temporal resolution while not requiring time for signal recovery.This opens the door to study the relationship between ultrastructural organization and dynamics of mitochondria, which are key processes in the life cycle of cells and are tightly regulated in health and disease 32,33 .So far, many of the underlying mechanisms remain elusive, because of their inherent heterogeneity and because of a limited spatiotemporal resolution 34 .

Discussion
In this paper, we present a method for super spatiotemporal resolution reconstruction of dynamic SMLM data.Our solution utilizes two main assumptions: (1) the imaged sample class is known (for example, filaments, mitochondria), and (2) dynamically varying objects maintain some degree of structural similarity over time, allowing the network to exploit this information.In other words, the information used by our network to recover a SMLM video with a temporal window of 20 frames is not contained in these 20 frames alone-but rather also in a window of hundreds of frames around it.Notably, both assumptions are necessary to achieve the spatiotemporal resolutions demonstrated in this work and must hold for accurate reconstructions.
To overcome the challenge of verifying the network results, we tested several cases in which estimation of ground truth position was possible, including numerical simulation, whole-sample-motion, that is, sample-drift and camera rotation, and motion of dynein motors on static microtubules.In simulations, the network could reconstruct nanoscale rapid movement of simulated filaments with 90% of the structure correctly classified per video, assuming certain SMLM conditions (for example ~20 nm localization precision, ~1 emitter μm -2 ); naturally, prediction accuracy varies as a function of fluorophore density, motion speed and other experimental parameters.
Ultimately, a main goal of our method is to enable live SMLM.For this purpose, we have analyzed SMLM videos of live-cell microtubule dynamics, provided by R. Tachibana et al. 35 as well as videos of ER dynamics previously analyzed by DECODE 14 .Since no ground truth structure is available in such an experiment, we have qualitatively assessed the reconstruction accuracy by comparison with other state-of-the-art solutions for dynamic SMLM.While state-of-the-art methods suffered from loss of structural information due to the analysis of short temporal windows, DBlink reconstructions presented a more complete description of the entire structure.
As in all model-based neural-net reconstruction algorithms, the network's ability to generalize will always be limited by the training data, and caution should be exercised when applying the method; specifically, training data must resemble the experimental structure to avoid hallucinations [36][37][38] .To guide DBlink users in choosing the appropriate SMLM training data and in the interpretation of the reconstructions, we have reported multiple ablation studies, the confidence map and further discussion in the sections Additional performance quantification and Confidence hallucination calculation in the Supplementary Information.Moreover, to validate the applicability of DBlink to structures with higher structural complexity than filaments, we tracked mitochondrial dynamics.First, we trained the neural network on simulated mitochondria-like structures, drifting and wobbling in time.Then, we used weak-affinity, noncovalent fluorophore labels that allow extended observation time 39 .The combination of these dyes with our high spatiotemporal reconstruction enables tracking dynamics in live cells at high spatiotemporal resolution in SMLM imaging over long observation times.
Future work can include extensions to other structures in live cells, expanding the types of motion in the training data to simulate elongation, contraction, wobbling and more complex dynamics, and systematic parameter optimization, for example, optimal sample densities, sample-motion-rate to acquisition rate ratio and more.Additionally, although in this work we used the same network architecture to reconstruct different types of samples, this is not always the optimal solution.For example, in case that the imaged sample is relatively large, increased receptive fields in the convolutional layers might be beneficial.Moreover, reducing the inference time of our network would also be very useful; new self-attention-based neural network architectures such as Transformers 40,41 might be exploited for this task.Also, expanding the ability of the network to analyze 3D information is also a desirable extension for this work.This would require good four-dimensional (3D+time) models of different biological structures to generate the training data.
Finally, our LSTM-based framework is versatile in terms of its input data.Here, we have used single-molecule localization maps as inputs; however, other inputs may provide even better performance.For example, DBlink can be used in combination with eSRRF, receiving as input radial gradient maps, which is suitable for analysis of densely labeled samples and may further improve performance.Our neural network based spatiotemporal interpolation could enable higher quality observation and ultimately facilitate discovery in various applications including cellular dynamics 42 , colocalization of nanoparticles with organelles 43,44 , synthetic materials 45 and more.

Fig. 1 |
Fig. 1 | DBlink concept.a, Low-resolution frames containing stochastic blinking events are analyzed by a localization method, in our case, Deep-STORM 12 , which generates super-resolved localization maps for each input frame.The localization maps serve as input to a CNN-LSTM network that provides as output

10 ]Fig. 2 |
Fig. 2 | Generation and analysis of simulated filament data.a, We simulated a random number of filaments in the field-of-view (FOV) according to the model of Shariff et al.25 .b, Then, we applied gradually increasing affine transformations over a predefined video length of N frames.c, Next, we generated random

Fig. 3 |
Fig. 3 | Static structure reconstruction during global motion.a, A STORM experiment containing undesirable drift was captured over 10,000 frames.We used Deep-STORM 12 to obtain a super-resolution reconstruction of the microtubule structure.Then, we used Deep-STORM's drift correction tool to predict the drift over the course of the experiment and acquire a single reconstructed frame in super resolution.The same localizations, without drift correction, were analyzed using DBlink.Scale bar, 2.5 μm.b, A STORM movie

Fig. 4 |
Fig. 4 | Tracking dynein motors dynamically moving on microtubules.a, Microtubules are scattered in the FOV.Dynein motors labeled with HALO Alexa 488 are moving on the microtubules.The blinking events of the entire experiment are localized using ThunderSTORM to generate the ground truth structure in super-resolution.b, Then, dynein motor localization is performed using Deep-STORM; localization maps are multiplied by a temporally changing window to simulate the movement of a short filament-like structure.

Fig. 5 |
Fig. 5 | Qualitative comparison of DBlink with other state-of-the-art methods.a, Left to right, single diffraction limited frame of microtubules in live-cell experiment (region of interest (ROI) marked by a red dashed rectangle); Deep-STORM reconstruction based on 300 frames; application of blind inpainting algorithm on Deep-STORM temporal window; DECODE reconstruction based on 500 frames; eSRRF reconstruction based on 100 frames; DBlink reconstruction at temporal resolution of a single blinking frame.b, Left to right, single diffraction limited frame of ER in live-cell experiment (ROI marked by a red dashed rectangle); Deep-STORM reconstruction based on 300 frames; application of blind inpainting on temporally windowed Deep-STORM; DECODE reconstruction based on 300 frames; eSRRF reconstruction based on 250 frames; DBlink reconstruction at temporal resolution of a single blinking frame.c, Left, confidence map of a reconstructed ROI; blue colored pixel intensities are increased for better visualization.DBlink has higher confidence in red colored pixels than in blue colored pixels.Right, DBlink reconstructed image.

Fig. 6 |
Fig. 6 | Reconstruction of a 12.5 min long video of mitochondria dynamics in a live cell.a, eSRRF 21 reconstruction of mitochondrial dynamics using temporal windows of ten frames.b, eSRRF reconstruction using temporal windows of 250 frames.c, DBlink reconstruction at the temporal resolution of the blinking data, namely, 20 fps.Each column represents a different timepoint.Yellow arrow marks the rapid motion of a circular mitochondria structure that could not be clearly observed by analyzing long temporal windows, due to motion blur.Red arrows mark missing structure not resolved by eSRRF due to the short analyzed temporal window.Scale bar, 3 μm.d, ROI 1, DBlink reconstruction enables tracking of mitochondrial fusion and fission dynamics in high spatiotemporal resolution.Scale bar, 1 μm.e, ROI 2, DBlink reconstruction containing previously reported mitochondria thinning before a fission event 47,48 .Scale bar, 1 μm.