Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
1
DeepCristae, a CNN for the restoration of mitochondria cristae in live
microscopy images
Salomé Papereux1,2,&, Ludovic Leconte1,2,&, Cesar Augusto Valades-Cruz1,2,&, Tianyan
Liu3, Julien Dumont4, Zhixing Chen3, Jean Salamero1,2, Charles Kervrann1,2, Anaïs
Badoual1,2,*
1 SERPICO Project Team, Centre Inria de l’Université de Rennes, F-35042 Rennes, France
2 SERPICO Project Team, UMR144 CNRS Institut Curie, PSL Research University, F-75005,
Paris, France
3 College of Future Technology, Institute of Molecular Medicine, National Biomedical Imaging
Center, Beijing Key Laboratory of Cardiometabolic Molecular Medicine, Peking-Tsinghua
Center for Life Science, Academy for Advanced Interdisciplinary Studies, Peking University,
Beijing 100871, China
4 CIRB Microscopy facility, Collège de France, UMR 7241 CNRS, Inserm U1050, Paris,
75005, France
& these authors contributed equally to the work
* corresponding authors contact: anais.badoual@inria.fr
Abstract
Mitochondria play an essential role in the life cycle of eukaryotic cells. However, we still don't
know how their ultrastructure, like the cristae of the inner membrane, dynamically evolves to
regulate these fundamental functions, in response to external conditions or during interaction
with other cell components. Although high-resolution fluorescent microscopy coupled with
recently developed innovative probes can reveal this structural organization, their long-term,
fast and live 3D imaging remains challenging. To address this problem, we have developed a
convolutional neural network (CNN), called DeepCristae, to restore mitochondrial cristae in
low spatial resolution microscopy images. Our CNN is trained from 2D STED images using a
novel loss specifically designed for cristae restoration. Random sampling centered on
mitochondrial areas was also developed to improve training efficiency. Quantitative
assessments were carried out using metrics we derived to give a meaningful measure of
cristae restoration. Depending on the conditions of use indicated, DeepCristae works well on
broad microscopy modalities (STED, Live-SR, AiryScan and LLSM). It is ultimately applied in
the context of mitochondrial network dynamics during interaction with endo/lysosomes
membranes.
Keywords: Convolution neural network, image restoration, mitochondria cristae, fluorescence
microscopy, super resolution, live imaging
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
2
Introduction
The study of certain pathologies has shown the importance of mitochondria, which above all,
ensure ATP production within cells and are central in many biological functions (e.g., metabolic
pathways, ion homeostasis, apoptosis, autophagy, epigenetics…)1,2. Mitochondrial energetic
adaptations to environmental constraints encompass a plethora of processes that maintain
cell survival. An alteration of these processes generally leads to serious diseases such as
cancer, neurodegenerative and cardiovascular disorders3. Although much attention has been
paid to the role of mitochondria, the precise niche the organelle plays in cell life and death still
remains unclear. The lack of in-depth knowledge about the ultrastructural evolution of
mitochondria in live cells, under normal and stressful conditions, might be one of the blind
spots. In particular, the cristae formed by the inner membrane of mitochondria that concentrate
ATP production in a defined area, their dynamic behavior, sublocation or density have been
poorly related to the various functionalities or dynamic processes (e.g., fusion, fission) that
mitochondria undergo. The challenge we address, lies in imaging mitochondria cristae, which
measure between 30 and 50 nm wide4, at a high spatial and temporal resolution so that their
structural dynamics and interactions can be accurately studied over time for several dozens of
milliseconds to a few seconds. However, this is starting to be possible with the recent
development of high-resolution imaging approaches5.
Stimulated emission depletion (STED) microscopy, which allows for sub-diffraction resolution
(xy: 30-50 nm), is one of the very few techniques6,7 able to decipher dynamics of mitochondria
cristae in live cells4. However, their observation in 3D and in fast time is limited by the
acquisition frame rate capacity (1 plane ≈1 to 10 s). In addition, depletion STED, which is the
principle that achieves nanoscopic resolution, induces local heat by high illumination intensity8
to which mitochondria are known to be particularly sensitive9,10. This can affect their overall
physiology and potentially lead to apoptosis and mitophagy. A number of new fluorescent
probes that are more photostable with less saturation intensity and that allow cristae
decoration, have been developed the very last years7,9,11,12. Yet, the application of a dark
recovery step (≈30 s) after STED imaging is still necessary, again at the expense of temporal
resolution. This could be improved by applying a partial STED depletion protocol, leading to
an intermediate quality resolution (xy≈100 nm)13, but insufficient to spatially resolve
mitochondria cristae and not solving the frame rate limitation (4-5 s in average).
In this context, one solution to study the dynamics of mitochondria cristae is to collect as much
temporal information with minimal phototoxicity using an appropriate microscope, and then
restore the spatial dimension using computational methods. Indeed, the development of image
restoration algorithms has become increasingly popular in recent years with the need for
nanoscale analysis14–22. At the heart of fluorescence microscopy have been actively developed
denoising algorithms23–28, dedicated to images corrupted by a mixed Poisson-Gaussian noise,
as well as deconvolution algorithms29–31, designed to remove the blur induced by the limited
aperture of the microscope objective. Some methods combine the two approaches32.
However, these conventional restoration methods usually rely on general assumptions, such
as the nature and level of noise and spatial regularity, which hampers their effectiveness on
the diversity of structures and level of degradation in microscopy images. Over the years, the
literature on image restoration has evolved considerably due to deep learning and the rapid
growth of convolutional neural networks (CNNs). These methods have the advantage of
making assumptions based on image content, resulting in state-of-the-art performance in
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
3
denoising33,34 and deblurring14,19,22,35–38 fluorescence microscopy images. However, these
methods have two major drawbacks. First, these CNNs often require a training step based on
a large ground truth dataset that is generally not available in microscopy. Second, they focus
on restoring the entire image, while sometimes little information is worth restoring within it,
especially in the dark background. This is the case with mitochondria cristae, which have a
sparse number of pixels in the image compared to the background. Therefore, CNNs that have
been previously applied to mitochondria microscopy images21,22,38,39 provide good global
restoration of the background and mitochondria but fail to accurately restore fine details as
cristae, especially in very low spatial resolution images. To circumvent this, new conventional
methods have been proposed to enhance resolution and suppress artifacts in high-resolution
techniques, including Hessian-SIM17. However, as explained above, the denoising results are
limited when dealing with low signal-to-noise ratio images and Hessian deconvolution
assumes that the unknown image is smooth in some sense and sparse. A hybrid solution has
been proposed in TDV-SIM40, which combines the strengths of conventional physical model-
based algorithms with deep learning-based algorithms. Another hybrid solution, rdLSIM21,
incorporates the deterministic physical model of specific microscopy into network training and
inference. Nevertheless, the effectiveness of these methods, along with conventional
restoration algorithms, relies on the careful selection of optimal parameters or on prior
knowledge of illumination patterns, respectively.
Instead of developing an additional generic image restoration method that may not
satisfactorily enhance certain sparse but informative pixels in the image, we present
DeepCristae, a CNN specifically developed to restore mitochondrial cristae in low spatial
resolution microscopy images. DeepCristae was applied to several microscopy modalities and
different biological scenarios capturing live mitochondria at high speed with low illumination
and thus low phototoxicity. DeepCristae allows long-term/fast dynamic observation of cristae
behavior and organization. The main challenge was to handle the low number of cristae pixels
compared to the background in the acquired images. Therefore, the main contributions of this
work are 1) the design of a new training loss dedicated to the restoration of specific pixels of
interest, 2) a random patch sampling focusing on areas of mitochondria, and 3) the building of
metrics for objective assessment of cristae restoration.
Results
Overview of DeepCristae
DeepCristae aims to restore mitochondrial cristae in intermediate to low spatial resolution
microscopy images (Fig. 1). DeepCristae mainly consists of a U-Net trained on a dedicated
dataset built from real high-resolution 2D STED images (Fig.1 and Supplementary Table 1)
and using a novel training loss we specifically designed for cristae restoration (Methods, Eq.
(1)). Although the term is not fully appropriate, for simplification we refer to this dataset as
“synthetic” . A pipeline for random patch sampling focusing on regions of mitochondria in
the acquired data was also developed (Supplementary Fig. 1). DeepCristae image restoration
network was implemented in Python (TensorFlow version 2.11) and is freely available as an
open-source software (see code availability). DeepCristae is also integrated into BioImageIT41,
an open-source platform with existing software for microscopy.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
4
DeepCristae quantitatively outperforms state-of-the-art algorithms on the synthetic
dataset
Our method was quantitatively compared to existing both conventional and deep learning
algorithms for image restoration. It includes denoising methods (ND-SAFIR23, 2D median filter
and Noise2Void33), deconvolution algorithms (Richardson-Lucy29,30 and Wiener31), and
approaches combining both, such as SPITFIR(e)32, CARE19, RCAN37 and SRResNet35 (details
are in Supplementary Note 2.2.2). All methods were applied to 9 test images of
(Supplementary Note 1.2.1). Note that in this dataset, the ground truth images were obtained
by deconvolving the high-resolution 2D STED images with a Richardson-Lucy (RL) algorithm
to enhance mitochondria cristae. Iterative RL algorithm is suitable for signal-dependent noise
removal. Nevertheless, after several iterations it may create unwanted artifacts such as a
"night sky" pattern in the image. To generate ground truth images from the input 2D STED
images, restored intermediate images with little “sky night” pattern, less blurred and noisy than
the input images, were carefully selected. All deep learning methods were trained from the
same patches extracted from .
First, to evaluate the performance of the different methods, we used current metrics, namely
NRMSE, PSNR and SSIM (Supplementary Note 2.1). However, these measures are relevant
to the image as a whole, but insufficient in the context of mitochondrial cristae restoration.
Indeed, the images contain only few pixels of cristae and thus have too little impact in those
metrics unlike the many background pixels. To overcome this issue, we encouraged the
evaluation metrics to focus exclusively on mitochondria pixels. To solve this problem,
evaluation metrics are focused exclusively on mitochondrial pixels. We call these mitochondrial
metrics , and . To go one step beyond and accurately assess
cristae restoration, we also introduced the cristae metrics , and
. These metrics are computed over mitochondria cristae pixels only, obtained from
manual annotations (Supplementary Note 2.1). Each competing algorithm was evaluated over
nine test images, for the nine aforementioned metrics (Fig. 2b). For all measurements focusing
on mitochondria or cristae, DeepCristae ranks first, and is either first or second otherwise.
Conventional methods behave worse than deep learning approaches, CARE appearing
DeepCristae's most competitive method. However, as confirmed by the values of the metrics
, and , CARE restores mitochondria cristae with less
sharpness, especially for mitochondria with low contrast or in a noisy background (Fig. 2a,
CARE white arrows). It should be noted that if we had considered conventional NRMSE, PSNR
and SSIM measurements, we would not have been able to demonstrate the benefits of
restoring more informative pixels. In terms of visual assessment, SRResNet and RCAN amplify
the background noise, resulting in less accurate restoration of cristae and unrealistic
reconstructed structures in the background or in mitochondria (Fig. 2a). DeepCristae removes
noise background while restoring most of the cristae details.
Second, a test image of was restored with DeepCristae (Fig. 3a). We selected four
regions of interest (ROIs) to better appreciate the restoration results (Fig. 3b-e). For each ROI,
a comparison of normalized intensity profiles between the input image, DeepCristae restored
image and the ground truth image is performed. It shows that DeepCristae restores spatial
information by revealing mitochondria cristae while improving signal to noise ratio. The line
profiles of DeepCristae are sometimes smoother than the ones of the ground truth, especially
in areas where the image background is noisier (Fig. 3d, e). This is due to a residual “night
sky” effect in the ground truth that does not appear with DeepCristae. This surprising result
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
5
can be explained in two ways. First, we reinject information from the input into the network at
the end and second, our new Similarity Component Prioritization (SCoP) loss (see Methods)
especially focuses on the mitochondria and cristae pixels where there are the fewest artifacts
with RL.
Robustness and stability of DeepCristae.
We have shown that DeepCristae performs well on 2D STED images and outperforms state-
of-the-art algorithms. However, it is important to verify the reliability of DeepCristae more
widely. DeepCristae has been trained on a dedicated dataset acquired with specific
microscope settings and mitochondria properties (e.g., fluorescence markers, width in pixels
of the mitochondria in the images). Any change in these settings is expected to alter the quality
of the restoration results. To evaluate the influence of changes in these parameters on the
results, three experiments were performed. Our model was trained from the training images of
depicting mitochondria of width 15.64 ± 4.04 pixels on average. We started by studying
the quality of the predictions as a function of the mitochondria width in pixels in the input
images. To that end, the 9 test images of were rescaled 11 times in order to contain
mitochondria of specific widths (in pixels) on average. It thus results on 11 test datasets on
which our trained DeepCristae model was applied (Supplementary Fig. 2a, b) and the metrics
(Supplementary Note 2.1) were computed. The evolution of the metrics as a function of
average mitochondrial width shows that the closer you get to the training parameters (i.e. an
average width of 15.64 pixels), the better the quality of the restorations. In fact, if the
mitochondria are too small, few cristae are restored and the mitochondria are thin. On the
contrary, if the size is too large, DeepCristae tends to create artifacts looking like cristae
patterns. Next, our model was trained on images obtained with specific parameters that mimic
microscope settings. Real images are assumed to be corrupted by mixed Poisson-Gaussian
noise (with standard deviation = 4) and the point spread function of the microscope is
approximated by an isotropic Gaussian function of standard deviation = 3.25 pixels. We
performed two experiments, similar to the one described above, to examine the assessment
of DeepCristae image restoration as a function of and (Supplementary Fig. 3). For
both experiments, visual results show that the quality of the restoration decreases as
and increase (Supplementary Fig. 3b, d). The higher the or values are, the
blurriest the mitochondria's boundaries and their cristae. This is also confirmed by the evolution
of the metrics as a function of (Supplementary Fig. 3a) and by the evolution of
and as a function of (Supplementary Fig. 3c). Surprisingly, the evolution of
the PSNR and NRMSE as a function of have a bell-shape with a maximum and a
minimum, respectively, for values of close to 3.25 pixels. We thus recommend using
DeepCristae on microscopy images with blur and noise levels at worst equal to our training
conditions ( = 4 and = 3.25 pixels). Beyond this, the quality of the restoration can
drastically decrease, especially for blur, and for too high values of these parameters
DeepCristae could reveal hallucinated cristae shapes.
Moreover, it is important to guarantee that under well-controlled conditions of use
(mitochondria width, image blur and noise), DeepCristae is hallucination-free and stable. By
stable we mean that different training leads to consistent predictions. To demonstrate that
these requirements have been met, we performed three experiments. First, we trained 10
DeepCristae neural networks with different training data, each generated with our patch
generation method applied to the 24 training images of (Fig. 4g-l). For each training, the
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
6
resulting model was applied to the 9 test images of and the aforementioned metrics were
computed. The average metrics obtained over the 10 trainings are close to the ones obtained
with our model and the standard deviations are very low, showing consistency between
predictions (Fig. 4l). By visually analyzing the predictions, the color map of the standard
deviation (Fig. 4g) as well as looking at normalized intensity line profiles along mitochondria
(Fig. 4h-k), we observe that the 10 trainings agree overall on the presence or absence of
cristae but diverge in their intensity and their precise boundaries. In this experiment, all
networks were initialized with the same weights, confirming that our patch sampling method is
robust and leads to homogeneous learning. A second similar experiment was carried out. Ten
trainings were performed from one dataset but with 10 different weights’ initializations
(Supplementary Fig. 2c-h). This experiment indicates that the same dataset leads to
homogeneous learning, meaning that the randomness of initialization does not play a key role
in the learning process. Finally, to qualitatively assess the performance of DeepCristae on real
data, we acquired 5 pairs of real 2D STED images. Each pair contains one low-resolution (LR)
and one high-resolution (HR) STED image, acquired as quickly as possible (~30 s), to
minimize the displacements and deformations of mitochondria between the two acquisitions
(Supplementary Note 1.2.1). The HR STED images were deconvolved using the RL algorithm
to enhance mitochondria cristae and are considered as “ground truths” (GTs). The LR STED
images were resized to have an average mitochondrial width of 15.64 pixels (391 nm), in line
with the conclusions drawn above, and were then given as input to DeepCristae. The obtained
predictions were qualitatively compared to the GTs to control their consistency. Four ROIs,
from two of the five pairs of real STED images and corresponding predictions, were selected
in regions where small mitochondrial displacements were observed to better appreciate the
restoration (Fig. 4a-f). For each ROI, a comparison of normalized intensity profiles between
the input LR image, DeepCristae restored image and the ground truth image is performed. A
consistency between the cristae restored by DeepCristae and the ones present in the ground
truths is observed overall, and no meaningful “hallucination” is observed.
DeepCristae enables to restore 2D+time STED images
While 2D STED nanoscopy enables to resolve mitochondria cristae and was here helpful to
develop DeepCristae, live STED acquisition encompasses a number of hurdles. It includes
relatively long-time frames between images, even when a photostable probe was used (Fig.
5a, b), limiting the temporal overview of the mitochondrial dynamics in the same plane.
Moreover, STED imaging may rapidly induce photo-bleaching, which makes ultrastructural
details progressively dimmed. More problematic, repeated STED imaging rapidly induces
morphological deterioration of live mitochondria, illustrated by their swelling in the latest time
points (Fig. 5a, b). This swelling effect was quantified here for HR STED by measuring the
lateral widths of 7 distinct mitochondria over the 10 time points (Fig. 5d) in the image series
(Fig. 5a, b). The swelling appears between the 5th and the 7th frames.
In order to improve the frame rate of 2D+time STED imaging while limiting the photodamages
on mitochondria, one may adjust the STED imaging protocols (Supplementary Note 1).
Accordingly, we built another live STED dataset, first to give an indication on how long, with
STED acquisition in lower depletion conditions, we can image before the mitochondria are
damaged and second, to control the efficiency of DeepCristae restoration over time. In what
follows, these low-resolution (LR) STED images were referred to as Fast STED images. This
goes at the depend of xy resolution (Fig. 5c, left bottom triangles in image series) for both the
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
7
lateral width of mitochondria and the cristae width (Fig. 5d, Fast STED and Fig. 5e, RAW,
respectively). Applying DeepCristae restoration on these latest series clearly revealed cristae
morphology (Fig. 5c right top triangle in image series). As expected, LR (or Fast) 2D+time
STED images show little changes in mitochondria lateral widths in time, in contrast to HR
STED (early and late time points in Fig. 5d) but a degraded resolution in the cristae widths
(from a mean () of ~90nm in HR STED to ~120 nm in Fast STED with = ± 47 nm). Applying
DeepCristae allows recovery of a resolution lower than 100 nm and drastically reduces the
variability of the measurement. The mean crista-to-crista distance, measured as peak-to-peak
intervals (Fig. 5f), widely depends on the cristae density along the mitochondria network. Here,
in RPE1 cell, it varies from 50 nm to more than 500 nm in early time points in HR STED (
=319 nm, = ± 247 nm) while the heterogeneity increases in late time points (from 130 nm up
to 1.6 µm), consistently with the observable swelling of the mitochondria. In Fast STED, the
cristae intervals measurements were non-significant. However, after DeepCristae restoration
the mean crista-to-crista distance was estimated at ~142 nm ( = ± 46 nm) (Fig. 5f, g).
Differences in these crista-to-crista measurements with similar studies on HeLa or Cos7 cells
for instance, will be further discussed. DeepCristae restored the individual cristae at 81 nm of
resolution ( = ± 9 nm) at FWHM (Full Width at Half Maximum), as compared to the
approximately 50 nm obtained in other studies4. DeepCristae provides a useful way to improve
live STED nanoscopy by improving the resolution and decreasing the frame rates (3 to 6s
versus 13s), yet with no observable photodamage as illustrated here by measuring the swelling
of mitochondria.
DeepCristae restore 3D+time images of mitochondria cristae by using intermediate
high-resolution and diffraction limited microscopy
STED nanoscopy is not the only high-resolution microscopy adapted to resolve internal
mitochondria ultrastructures in live cells. Indeed, a number of works using adaptation of SIM
approaches have been published over the last few years16,42, some combined with
conventional deep learning methods20,22,39. Yet, the best compromise between Fast and 3D
imaging still remains an issue. We next investigated the performance of DeepCristae
prediction on intermediate HR microscopies chosen for their 3D optical slicing performance.
Spinning disk confocal equipped with a Live-SR module (or SDSRM for Spinning Disk Super
Resolution Microscopy) is one of those well-disseminated systems equivalent to SIM. It
improves the xy resolution by a factor of ~2 (~120-130 nm at 488 nm, ~140 nm at 561 nm)43
while giving access to the depth (z-axis) of the sample and the live imaging of mitochondria
(time t) without severe photo-bleaching and phototoxicity. The use of the Live-SR is therefore
motivated here by both the study of these four dimensions and the ability of our model to
correctly perform cristae reconstruction via multiple microscope imaging modalities.
DeepCristae efficiently revealed cristae organization in single 2D Live-SR images acquired
within 30 ms (Fig. 6a, upper images) and thus in 3D (Fig. 6a, lower images, MIP on 14 planes,
with a stack time ~800 ms), giving access to the overall mitochondria network in the live cell
at a fast rate. In this respect, it outperforms HR STED imaging and even Fast (LR) 2D-STED
imaging after DeepCristae restoration (compared to Fig. 5c). Cristae width comparative
estimation (Fig. 6b) shows the improvement in resolution obtained after DeepCristae
restoration on single plane Live-SR images (Raw =149 ± 64 nm when measurable;
DeepCristae =87 ± 11 nm). These results are close to the expected widths of circumvoluted
cristae tubules (50 to 100 nm) obtained by other methods derived from SIM imaging22. We
then tested DeepCristae restoration on LLSM44 imaging which is not an HR microscopy by
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
8
itself (at least in the dithered mode) but gives the best compromise in terms of fast 3D
acquisition with minimal photon dose illumination and consequently low photo-damage of the
mitochondria over time. Surprisingly, although with intrinsic limited and non-isotropic
resolutions (PSF xy=300 nm and z=600-700 nm in our system) and a particular geometrical
acquisition mode, cristae were however detectable in some mitochondria after realignment
and a Richardson Lucy (RL) deconvolution. The resulting images were here considered as
“Raw” data (Fig. 6c, 2D upper panel, left image). Applying DeepCristae to them (Fig 6c, 2D
upper panel, right image and composite zoomed area for comparison) improves the cristae
resolution (Raw = 339 ± 248 nm, when measurable; DeepCristae= 94 ± 15 nm) and strongly
reduces the variance of paired measurements (Fig. 6d). One of the obvious advantages of
LLSM over confocal imaging is to allow continuity between single image planes over large
stacks coupled to an extended depth of focus, as illustrated here by the 3D rendering as an
oblique projection (Fig. 6c, 3D). Moreover, LLSM is particularly adapted to long range/high
frequency imaging on whole living cells, which, coupled to low photon dose illumination, makes
it one of the best imaging systems, if not the best, for the highly light-sensitive organelles that
are the mitochondria. Applying DeepCristae adds information on cristae ultrastructural
organization in the whole mitochondria network of the cell. Finally, Fast 3D Live-SR and LLSM
time series (Fig 6e, f) were treated for DeepCristae restoration. Cristae ultrastructural features
can be observed, while the mitochondrion network undergoes well known dynamic
modifications such as fusion or fission processes (Fig. 6e, f, panels of composite zoomed area
in both time series; left “RAW '' and right “DeepCristae”; Movies S1 and S2). Images are of
better quality after restoration of Live-SR compared to LLSM images. However, it should be
noted the gain in acquisition parameters for the latter in these experiments, with 75 slices per
stack and a double channel stack time=1.3 s versus 14 planes per stack and double channel
stack time= 3.9 s for Live-SR. DeepCristae restoration was also tested with an AiryScan 5
LSM 980. It provided similar improvements, although for a 15 planes stack time of about 30s
and with more artifacts appearing after DeepCristae, the nature of which most probably lies in
the way the reconstruction of the AiryScan images was carried out from the values determined
automatically by the commercial software (Supplementary Fig. 5).
DeepCristae restoration allows to decipher mitochondria cristae morphodynamics
during inter organelles interactions
The most documented membrane-membrane interactions involving mitochondria are the
endoplasmic reticulum (ER)–mitochondria contacts, whose functions have been continuously
expanded since the 1990s45,46. In addition to the ER, mitochondria contact
vacuoles/lysosomes, peroxisomes, lipid droplets, endosomes, the Golgi, the plasma
membrane (PM) and melanosomes47. The number of these interactions as well as their
duration drastically vary from one type to the other, as they depend on the respective
membrane surface of the specific organelles within the cell and their contact time48. Their
detection may thus require fast and/or long-range 3D imaging. As already mentioned, even
high-resolution approaches which are well adapted to decipher ultrastructural features of
mitochondria such as cristae, generally fail to capture their dynamic evolution in the 3D space
of the whole cell at multiple time scales. This can be critical, if one wants to study inter-
organelle membrane interactions and their effects. We next initiate the investigation of
endosome/lysosome-mitochondria dynamic interactions by addressing specifically the
ultrastructural behavior of the cristae during these contacts. This was done by imaging multiple
3D+time double fluorescence series in Live-SR (represented as a single stack MIP in Fig. 7a,
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
9
top left) or LLSM (represented as a single stack MIP in Fig. 7c, top left), where the membranes
of the endo-lysosomal pathway were continuously labeled (Supplementary Note 1.1).
DeepCristae restoration was applied on both MIP datasets (respectively, Fig. 7 a, c, bottom
left). A number of mitochondria dynamic events correlated with endosomal structure behaviors
were captured. Only a few of them are here extracted as thumbnail time series (Fig. 7b, d) of
zoomed area (colored insets in Fig. 7a, c) from the Live-SR and LLSM acquisitions,
respectively. Among others, the formation of endo-lysosomes contacts sites with mitochondria
(Fig. 7b, blue and green), positioning of endo-lysosomes relative to fission sites of
mitochondria (Fig. 7b, blue and red), very long confinement of endo-lysosomes within the
mitochondria network (Fig. 7b, orange) and image series of endo-lysosomes appearing to pull
a small mitochondria from one to another elongated tubules of mitochondria (Fig. 7b, red).
DeepCristae restoration on the space-time localization of these events can also be evaluated
dynamically (Movie S3). Similar events are followed with LLSM, such as the fission of
mitochondria at a contact site with an endo/lysosome vesicle (Fig. 7d, orange) and long
confinement of an endo/lysosome vesicle within the mitochondria network (Fig. 7d, green).
The main advantage of the LLS modality (fast frame rate, low photon illumination of the sample
coupled to whole cell 3D acquisition) is the improvement of the time resolution of the data
series (or long-range acquisition). Consequently, fast events involving endo-lysosome
contacts with mitochondria are easier to capture and these dynamics are precisely deciphered.
For instance, one may extract first (Fig. 7c, blue inset), probably a fusion process (Fig. 7d,
from time point 105 to time point 119; =13 ), and second a fission process (Fig. 7d, from
time point 192 to time point 198; = 6 ; Movie S4). At each time point, DeepCristae-restored
mitochondria and denoised/deconvoluted endo-lysosomes double-labeled images
(Supplementary Note 1) are paired to non-treated images (right and left panel, respectively, of
thumbnails time series in Fig. 7b, d). While cristae resolution in LLSM does not reach that
obtained with Live-SR, DeepCristae restoration brings values closer together (Fig. 6b, d). In
all situations and for both intermediate HR (Live-SR) and diffracted limited (LLSM) imaging
modalities, DeepCristae restoration provides ultrastructural information on the positioning,
density, and dynamics of mitochondria cristae. We then wanted to quantitatively assess how
the dynamic architecture of the mitochondria internal membrane during endo/lysosomes-
mitochondria interaction could be revealed with DeepCristae. We focused on the fission
process. To do so, we first selected 21 distinct 3D+time image series from the Live-SR
datasets, in which mitochondria fission was monitored. Intensity line plots were measured
along mitochondria on some time points framing the fission event. This was done on both
DeepCristae-restored and unrestored individual time points in a “blind” manner, meaning
without looking in the second channel depicting the location of endo-lysosomes
(Supplementary Fig. 6a). Measurements of “peak to peak” intervals between cristae, were only
possible in the DeepCristae restored images and show an increased density after fission
occurs (Supplementary Fig. 6b, dark circles). Interestingly, by overlaying the second channel
in a second step, 60% (13 over 20) of these selected time series showed proximity if not direct
contact at the site where mitochondria fission is observed (Supplementary Fig. 6b, red circles;
6c for statistics). While still preliminary and not deciphering the exact nature of the endosomal
compartments involved (i.e., PMDR labels the overall endo-lysosomal pathway), this illustrates
how DeepCristae would represent an asset to quantitatively study the dynamic architecture of
the mitochondria internal membrane during diverse dynamic processes or in particular
physiological or constrained conditions.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
10
Discussion
Mitochondrial membrane architecture is essential for the many functions of mitochondria. In
particular, mitochondria cristae are the main site of energy production and are dynamic
ultrastructures remodeling upon various cellular stimuli and natural processes (apoptosis1;
aging49). Therefore, understanding the structure and dynamics of cristae is vital for
comprehending mitochondrial function and its implications in cellular physiology and diseases.
High-resolution microscopy coupled with robust mitochondrial probes7,9 are key recent
developments that started to reveal the fine details of mitochondrial cristae structure and
organization, overcoming the limitations of conventional microscopy. However, imaging at high
spatial and temporal resolution remains a challenge.
DeepCristae exploits the power of deep learning to reveal cristae in images taken with low
photon illumination, enabling clearer visualization and analysis of mitochondria cristae in living
cells without interfering with the natural behavior of mitochondria. While it has been trained on
a dedicated dataset that was created from real high-resolution 2D STED images, we have
shown that it operates for a wide range of optical resolutions, from diffraction-limited to
intermediate high-resolution microscopy, providing researchers with a powerful tool to study
cristae dynamics without compromising their structural integrity or functionality.
While there are other deep learning approaches available for revealing cristae
ultrastructure21,22,38,39, DeepCristae offers unprecedented advantages. First, thanks to a well -
defined training loss dedicated to the restoration of mitochondria signals, it outperforms state-
of-the-art methods. Secondly, it not only makes it possible to visualize and restore cristae
dynamics in 2D STED nanoscopy with minimal illumination and without damaging
mitochondria but more importantly, it extends these capabilities to other high-resolution
imaging techniques such as Live-SR, SIM and AiryScan, more suited to such 3D dynamics.
Finally, DeepCristae can be applied to advanced microscopy techniques such as LLSM,
enabling fast and long-duration 3D+time acquisitions within the diffraction-limited range. This
versatility makes DeepCristae a unique and valuable solution for studying cristae dynamics
across a range of spatial and temporal scales.
Overall, our results show that fluorescence microscopy combined with DeepCristae enables
long-term/fast dynamic observation of cristae behavior and organization with high quality. To
illustrate the contribution of our approach to biological phenomena that are likely to involve the
functional structure of mitochondria, we have chosen to focus on inter-organelle interactions
and their consequences. While mitochondria-associated ER membranes, the biochemical
composition of the contact sites and diverse physiological and disease-related functions have
been extensively studied over the decades50,51, it is increasingly recognized that other
organelle contacts have a vital role in diverse cellular functions52. More recently, there has
been growing interest in quantifying other membrane interactions with mitochondria and their
cell distribution in space and time48, in particular within the endo-lysosomal pathway and their
contribution to the fission/fusion process of the mitochondria network53. Here, while confirming
the coincidence of contacts between the endo-lysosomal membrane and mitochondria, we
enlightened the change of cristae density during fission (Supplementary Fig. 6). This density
as well as complex cristae arrangements depends on cell types and metabolic activities4,54,
not talking of obvious modifications induced by environmental conditions. Until now, to provide
a dynamic view of individuals and groups of cristae required 3D nanoscopy or linear SIM16,
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
11
which are not always compatible with the time frame required to capture the event of interest.
In this respect, DeepCristae might be an asset to compare the cristae dynamics in different
cell types and in these various conditions.
However, as with any image restoration method, scientists may be concerned by the
robustness of DeepCristae to accurately restore mitochondria cristae and not hallucinate them.
This is why we investigated the stability and limits of our method. First, we consider the RL
algorithm in our proof-of-concept as it may be considered as the baseline algorithm for image
deconvolution in fluorescence microscopy. The behavior of this algorithm is well established,
including the “night sky” artifacts when iterated. The GTs shown in the manuscript may be
improved if we consider more sophisticated deconvolution algorithms. We also worked out
different conditions of use to be respected to guarantee good quality and truthfulness of the
results. It is important to feed DeepCristae with images containing mitochondria whose
average width in pixels is close to the one seen during the training. Concerning the microscope
settings, it is better to ensure that the level of noise and blurring in the input images is
equivalent to or better than the one present in the training data (which was quite high in our
training). Under these conditions of use, across all our experiments and through different
microscopy modalities, no hallucination was observed: a consistency between line profiles
along mitochondria between raw and restored data was always observed.
Like cytoskeletal elements, the mitochondrial ultrastructure is a key element for comparing the
performance of new super-resolution microscopy techniques. In terms of applications,
DeepCristae makes it possible to track the evolution of mitochondrial cristae morphology over
time, during interactions with other membrane components of the cell, or under extracellular
conditions that mimic various pathological or stress situations.
Methods
In this section, we present the main features of DeepCristae. We first present the dataset we
created from real 2D STED images to train the network. Then, we overview our network
architecture and present the novel learning loss function, which prioritizes the restoration of
specific pixels. We finally detail the image patch-sampling steps, which is a crucial stage in the
model pipeline.
Generation of the 2D STED dataset - . As mitochondria are living organelles, mostly
organized as a quite fast-moving network in RPE1 cells (Supplementary Note 1.1), the
acquisition of a pair of high and low-resolution images at the exact same time point is
impossible. To train and quantitatively validate DeepCristae, we thus created a dataset, called
, from 33 acquired high resolution (HR) 2D STED images (25 x 25 nm) that we denote
(Fig. 1 a-d). More information on the acquisition of the images are available in
Supplementary Note 1.2.1.
First, we degraded the images to obtain low resolution (LR) images of mitochondria,
denoted , that will serve as input to the neural network. To that end, we first applied a
Gaussian filter of standard deviation = 3.25 pixels to the images in order to
approximate the blurring effect due to the point spread function of the microscope. Then, we
added a Poisson-Gaussian noise ( = 4.0 ), consistently with noises observed in
real STED images. The parameters and were set to create pertinent input data
that mimic real LR STED images (Supplementary Note 1.2.1).
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
12
Second, as our goal is to learn how to restore LR images to better reveal the mitochondria
cristae, we paired the LR STED images with their restored counterpart . These
ground truths are obtained by deconvolving the real HR images with the Richardson-Lucy
algorithm29,30, enhancing the mitochondria cristae. Other deconvolution algorithms were tried,
such as SPITFIR(e)32 or Wiener31, but the results obtained after training were not as good.
Finally, the dataset is made of 24 training images and 9 test images. To further increase
the size of the training set, data augmentation and patch sampling (later described in Methods)
are performed on the pair of LR images and ground truths . The dataset is first
augmented by applying three different rotations to the images (90°, 180° and 270°). Then, a
shrink transform, and horizontal and vertical flips are successively applied to 25% of the
augmented dataset, randomly selected. The final training set is made of 1824 patches of size
128×128 pixels, whose 20% are used for the validation set (summary in Supplementary Table
1).
Network architecture. We used the network proposed by Weigert et al.19 as the backbone of
the CNN architecture, itself built upon the U-Net55. It has a contracting path and an expansive
path, each one consisting of 3 sequential downsampling and upsampling blocks, respectively.
Each block of the first path is skip-connected to the associated one of the expansive paths.
The contracting path consists of two successive 3×3 convolutions, each followed by a Rectified
Linear Unit (ReLU), and a 2×2 max pooling operation with stride 2 for downsampling. Every
depth in the expansive path consists of a 2×2 up-sampling of the feature map, concatenated
with the corresponding feature map from the contracting path, followed by two 3×3
convolutions with a ReLU activation function. At the final layer, one 1×1 convolution is used.
The output results from an additive assembly between the input of the neural network and the
last layer's output. The network (Fig. 1f) outputs the same size restored images.
Design of the training loss. We present our new loss, the Similarity Component Prioritization
(SCoP) loss, that has been designed to better restore mitochondria cristae. Most losses and
metrics used to train networks or to evaluate the quality of restorations compute the score on
the whole image, giving the same weight to any pixel. For example, the MAE computes the
mean absolute error between the prediction and the target image, while the SSIM, despite not
basing the calculation on pixels-to-pixels difference, computes the similarity among all the
pixels of both images. Instead, our purpose is to focus on informative pixels corresponding to
target structures in images. Indeed, the dark and noisy background occupies most of
fluorescence images, which alters the learning. To overcome this issue, we introduce SCoP,
a novel loss which adaptively prioritizes the restoration of mitochondria pixels.
Our SCoP loss is built upon the structural dissimilarity (DSSIM) measure. Consider (,) the
spatial coordinates of a given pixel and a patch of size (,). The loss formula between a
target image and its prediction is given by
(,)=1
1,
(,)
2
,
, (1)
Where ,
is the map of the local structural similarity (SSIM) values for corresponding
pixels between the images and . Each SSIM value ranges in [-1,1], where -1 (1,
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
13
respectively) testifies of a bad (very good similarity, respectively) between (,) and (,).
The parameter , prioritizes the restoration of specific regions of interest. In our case, we
chose ,= 1 if the pixel (,) belongs to a mitochondrion, 4 otherwise. In this way, we
encourage the network to focus on restoring mitochondria pixels and reduce the influence of
a poorly restored background on the loss. Determining whether a pixel belongs to the
background or to a mitochondrion can be performed automatically (using our method
described in Methods “Image patch sampling for the training step - Thresholding”) or manually
by using any binary segmentation provided by the user.
Data normalization. Our training images of have different ranges of intensity values. To
homogenize them, we normalized the input data and their corresponding ground truth to a
common distribution of intensity values with the percentile normalizer. This normalization also
has the advantage to exclude outliers, which are very frequent in microscopy imaging due to
noise and luminance. The percentile normalization of an image is defined as
=
(,
)
, (,) , (2)
where (,) is the p-th percentile of . We used = 2 and =99.8. This step is
also performed during the inference step on any input data.
Image patch sampling for the training step. Our model is trained on the dataset
containing 24 images (96 after data augmentation) of different sizes. In order to homogenize
and increase the training dataset, we performed patch sampling. We sampled each input
training image ×, defined over the grid of size ×, within =
patches
of size 128×128. As our images contain more background pixels than mitochondria pixels, grid
or simple random patch sampling will end in too many empty patches. This can degrade the
training of our model. Instead, we perform a random sampling focusing on the regions of
interest, the mitochondria. Our pipeline (Supplementary Fig. 1) is described as follows.
● Anscombe transform. To detect the areas of interest, we need to enhance the
mitochondria signal with respect to the noise. To do this, we first remove the Poisson-Gaussian
noise in STED images. This is achieved by applying an Anscombe transform, which enables to
stabilize noise variance and to approximately convert Poisson-Gaussian noise into white Gaussian
noise with a constant variance. The Anscombe transform of an image is given by
(,)= 23
8+(,),(,) . (3)
● Z-score. Then, we compute the Z-score map defined as
(,)=
(,)
,(,) , (4)
where and are the estimated mean and standard deviation of the Gaussian noise ,
respectively. Since most of the pixels in belong to the background, we consider =
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
14
{(,)}(,). For , we use a robust estimator derived from the Median Absolute
Deviation (MAD) such that = 1.4826 { |(,)|}(,), where (,) =
(,) (,) (,)
, (,), are the pseudo-residuals. In fact, under the
hypothesis of having a white Gaussian noise and that the noise-free image is piecewise
smooth in a local neighborhood, we have that =[(,)].
● Thresholding. The higher the Z-score in Eq. (4), the higher the pixel value is above the
mean of the measured noise and therefore the pixel (,) is considered as a pixel of interest.
We apply a threshold c, in a way that any pixel (,) such that (,) > is considered as
a mitochondria pixel. We denote this set as . The threshold is automatically adapted for
each training image. Starting from a fixed high value of 30, while does not contain a
minimum of 10% mitochondrial information (i.e., # < 10% #, where # and #
denote the number of pixels in the sets and , respectively), we subtract 5 from the
threshold value. This creates a binary mask on which we apply a median to remove the
surrounding noise. This automatic procedure avoids cumbersome manual annotations. Note
that this mask can also be used to compute the parameter , in our loss (see Eq. (1)).
● ROIs selection. From , we randomly choose different pixels to be the center of
ROIs of size 128×128 pixels. Thus, the more pixels of mitochondria a ROI contains, the more
likely it is to be chosen. The following conditions have to be respected: i) the ROI centers
should not belong to the borders of the image; ii) to avoid redundancy, a minimum distance of
60 pixels is established between each pairwise ROI center. The resulting ROIs are finally used
to create the patches from the normalized training data (see above “Data normalization”).
Network evaluation. In addition to a quantitative comparison to the state-of-the-art methods
and experiments to show the reliability of our method (Results), an ablation study was also
performed (Supplementary Fig. 4; Supplementary Note 2.3) to highlight the individual
contribution of key components of our method. More details about the evaluation metrics and
the implementation details of DeepCristae are also given in Supplementary Note 2.1 and
Supplementary Note 2.2.1, respectively.
Other Methods and Materials
Cell culture and biological materials, fluorescence labeling, all used microscopy techniques,
image acquisition protocols and quantitative measurements are detailed in the Supplementary
Note 1. PKMITO dyes are commercially available at Spirochrome (Stein-am-Rhein,
Switzerland) and Genvivo Biotech (Nanjing, China)
Data availability
Data can be accessed in the following private link:
https://figshare.com/s/39819e2b1a84a7afdb7d
Code availability
DeepCristae source code used in this publication is open-source and published under the BSD
3-Clause "Original" or "Old" License. Source code will be available shortly through GitHub.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
15
References
1. Bock, F. J. & Tait, S. W. G. Mitochondria as multifaceted regulators of cell death. Nat
Rev Mol Cell Biol 21, 85–100 (2020).
2. Stimpfel, M., Jancar, N. & Virant-Klun, I. New Challenge: Mitochondrial Epigenetics?
Stem Cell Rev Rep 14, 13–26 (2018).
3. Chakrabarty, S. et al. Mitochondria in health and disease. Mitochondrion 43, 25–29
(2018).
4. Stephan, T., Roesch, A., Riedel, D. & Jakobs, S. Live-cell STED nanoscopy of
mitochondrial cristae. Sci Rep 9, 12419 (2019).
5. Jakobs, S., Stephan, T., Ilgen, P. & Brüser, C. Light Microscopy of Mitochondria at the
Nanoscale. Annu Rev Biophys 49, 289–308 (2020).
6. Ishigaki, M. et al. STED super-resolution imaging of mitochondria labeled with TMRM
in living cells. Mitochondrion 28, 79–87 (2016).
7. Yang, Z. et al. Cyclooctatetraene-conjugated cyanine mitochondrial probes minimize
phototoxicity in fluorescence and nanoscopic imaging. Chem Sci 11, 8506–8516
(2020).
8. Vicidomini, G., Bianchini, P. & Diaspro, A. STED super-resolved microscopy. Nat
Methods 15, 173–182 (2018).
9. Yang, X. et al. Mitochondrial dynamics quantitatively revealed by STED nanoscopy
with an enhanced squaraine variant probe. Nat Commun 11, 3699 (2020).
10. Kilian, N. et al. Assessing photodamage in live-cell STED microscopy. Nat Methods
15, 755–756 (2018).
11. Liu, T. et al. Multi-color live-cell STED nanoscopy of mitochondria with a gentle inner
membrane stain. Proceedings of the National Academy of Sciences 119,
e2215799119 (2022).
12. Wang, C. et al. A photostable fluorescent marker for the superresolution live imaging
of the dynamic structure of the mitochondrial cristae. Proceedings of the National
Academy of Sciences 116, 15817–15822 (2019).
13. Zhang, J. et al. Low-Power Two-Color Stimulated Emission Depletion Microscopy for
Live Cell Imaging. Biosensors (Basel) 11, (2021).
14. Guo, M. et al. Rapid image deconvolution and multiview fusion for optical microscopy.
Nat Biotechnol 38, 1337–1346 (2020).
15. Koho, S. et al. Fourier ring correlation simplifies image restoration in fluorescence
microscopy. Nat Commun 10, 3103 (2019).
16. Huang, X. et al. Fast, long-term, super-resolution imaging with Hessian structured
illumination microscopy. Nat Biotechnol 36, 451–459 (2018).
17. Zhao, W. et al. Sparse deconvolution improves the resolution of live-cell super-
resolution fluorescence microscopy. Nat Biotechnol 40, 606–617 (2022).
18. Arigovindan, M. et al. High-resolution restoration of 3D structures from widefield
images with extreme low signal-to-noise-ratio. Proceedings of the National Academy
of Sciences 110, 17344–17349 (2013).
19. Weigert, M. et al. Content-aware image restoration: pushing the limits of fluorescence
microscopy. Nat Methods 15, 1090–1097 (2018).
20. Jin, L. et al. Deep learning enables structured illumination microscopy with low light
levels and enhanced speed. Nat Commun 11, 1934 (2020).
21. Qiao, C. et al. Rationalized deep learning super-resolution microscopy for sustained
live imaging of rapid subcellular processes. Nat Biotechnol 41, 367–377 (2023).
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
16
22. Li, Y. et al. Incorporating the image formation process into deep learning improves
network performance. Nat Methods 19, 1427–1437 (2022).
23. Boulanger, J. et al. Patch-Based Nonlocal Functional for Denoising Fluorescence
Microscopy Image Sequences. IEEE Trans Med Imaging 29, 442–454 (2010).
24. Kempen, G. M. P. van, Voort, H. T. M. van der, Bauman, J. G. J. & Strasters, K. C.
Comparing maximum likelihood estimation and constrained Tikhonov-Miller
restoration. IEEE Engineering in Medicine and Biology Magazine 15, 76–83 (1996).
25. Van Der Voort, H. T. M. & Strasters, K. C. Restoration of confocal images for
quantitative image analysis. J Microsc 178, 165–181 (1995).
26. Van Kempen, G. M. P., Van Vliet, L. J., Verveer, P. J. & Van Der Voort, H. T. M. A
quantitative comparison of image restoration methods for confocal microscopy. J
Microsc 185, 354–365 (1997).
27. Buades, A., Coll, B. & Morel, J. M. A non-local algorithm for image denoising. in 2005
IEEE Computer Society Conference on Computer Vision and Pattern Recognition
(CVPR’05) vol. 2 60–65 vol. 2 (2005).
28. Dabov, K., Foi, A., Katkovnik, V. & Egiazarian, K. Image Denoising by Sparse 3-D
Transform-Domain Collaborative Filtering. IEEE Transactions on Image Processing
16, 2080–2095 (2007).
29. Richardson, W. H. Bayesian-based iterative method of image restoration. JoSA 62,
55–59 (1972).
30. Lucy, L. B. An iterative technique for the rectification of observed distributions. Astron
J 79, 745 (1974).
31. Gonzalez, R. C. & Woods, R. E. Digital Image Processing. (Pearson Education, Inc.,
2003).
32. Prigent, S. et al. SPITFIR(e): a supermaneuverable algorithm for fast denoising and
deconvolution of 3D fluorescence microscopy images and videos. Sci Rep 13, 1489
(2023).
33. Krull, A., Buchholz, T.-O. & Jug, F. Noise2Void - Learning denoising from single noisy
images. in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition
(CVPR) 2124–2132 (2019).
34. Batson, J. & Royer, L. Noise2self: Blind denoising by self-supervision. in International
Conference on Machine Learning 524–533 (2019).
35. Ledig, C. et al. Photo-Realistic Single Image Super-Resolution Using a Generative
Adversarial Network. in 2017 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR) 105–114 (2017).
36. Zhang, K., Zuo, W., Chen, Y., Meng, D. & Zhang, L. Beyond a Gaussian Denoiser:
Residual Learning of Deep CNN for Image Denoising. IEEE Transactions on Image
Processing 26, 3142–3155 (2017).
37. Zhang, Y. et al. Image Super-Resolution Using Very Deep Residual Channel Attention
Networks. in Computer Vision – ECCV 2018 (eds. Ferrari, V., Hebert, M.,
Sminchisescu, C. & Weiss, Y.) 294–310 (Springer International Publishing, 2018).
38. Chen, J. et al. Three-dimensional residual channel attention networks denoise and
sharpen fluorescence microscopy image volumes. Nat Methods 18, 678–687 (2021).
39. Chen, X. et al. Self-supervised denoising for structured illumination microscopy
enables long-term super-resolution live-cell imaging. bioRxiv 2023.04.05.535684
(2023).
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
17
40. Wang, J., Fan, J., Zhou, B., Huang, X. & Chen, L. Hybrid reconstruction of the
physical model with the deep learning that improves structured illumination
microscopy. Advanced Photonics Nexus 2, 16012 (2023).
41. Prigent, S. et al. BioImageIT: Open-source framework for integration of image data
management with analysis. Nat Methods 19, 1328–1330 (2022).
42. Opstad, I. S., Wolfson, D. L., Øie, C. I. & Ahluwalia, B. S. Multi-color imaging of sub-
mitochondrial structures in living cells using structured illumination microscopy. 7,
935–947 (2018).
43. Hayashi, S. & Okada, Y. Ultrafast superresolution fluorescence imaging with spinning
disk confocal microscope optics. Mol Biol Cell 26, 1743–1751 (2015).
44. Chen, B.-C. et al. Lattice light-sheet microscopy: Imaging molecules to embryos at
high spatiotemporal resolution. Science (1979) 346, 1257998 (2014).
45. Lang, A., John Peter, A. T. & Kornmann, B. ER–mitochondria contact sites in yeast:
beyond the myths of ERMES. Curr Opin Cell Biol 35, 7–12 (2015).
46. Vance, J. E. Phospholipid synthesis in a membrane fraction associated with
mitochondria. Journal of Biological Chemistry 265, 7248–7256 (1990).
47. Lackner, L. L. The Expanding and Unexpected Functions of Mitochondria Contact
Sites. Trends Cell Biol 29, 580–590 (2019).
48. Valm, A. M. et al. Applying systems-level spectral imaging and analysis to reveal the
organelle interactome. Nature 546, 162–167 (2017).
49. Amorim, J. A. et al. Mitochondrial and metabolic dysfunction in ageing and age-related
diseases. Nat Rev Endocrinol 18, 243–258 (2022).
50. Vance, J. E. MAM (mitochondria-associated membranes) in mammalian cells: Lipids
and beyond. Biochimica et Biophysica Acta (BBA) - Molecular and Cell Biology of
Lipids 1841, 595–609 (2014).
51. van Vliet, A. R. & Agostinis, P. Mitochondria-Associated Membranes and ER Stress.
in Coordinating Organismal Physiology Through the Unfolded Protein Response (eds.
Wiseman, R. L. & Haynes, C. M.) 73–102 (Springer International Publishing, 2018).
52. Rowland, A. A., Chitwood, P. J., Phillips, M. J. & Voeltz, G. K. ER Contact Sites
Define the Position and Timing of Endosome Fission. Cell 159, 1027–1041 (2014).
53. Kleele, T. et al. Distinct fission signatures predict mitochondrial degradation or
biogenesis. Nature 593, 435–439 (2021).
54. Stoldt, S. et al. Spatial orchestration of mitochondrial translation and OXPHOS
complex assembly. Nat Cell Biol 20, 528–534 (2018).
55. Ronneberger, O., Fischer, P. & Brox, T. U-Net: Convolutional Networks for Biomedical
Image Segmentation. in Medical Image Computing and Computer-Assisted
Intervention – MICCAI 2015 234–241 (2015).
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
18
Acknowledgements
This work was supported by the France-BioImaging Infrastructure (French National Research
Agency, ANR-10-INBS-04-07, “Investments for the future”) and the Labex Cell(n)Scale (ANR-
11-LABX-0038) as part of the Idex PSL (ANR-10-IDEX-0001-02). We acknowledge the Cell
and Tissue Imaging (PICT IBiSA, Institut Curie) and the IMACHEM (Collège de France)
platforms, also members of the national infrastructure France-BioImaging (ANR-10-INBS-04-
01) for access to and maintaining the spinning-disk, Airyscan and STED microscopes. We also
wish to thank M. Maurin from Inserm U932 for his help in a preliminary study on STED image
acquisition.
Author Contributions
A.B., L.L., C-A.V-C and J.S. conceived the project. S.P. and A.B. designed the framework of
DeepCristae, and conducted benchmarks and every experiment relative to the reliability of the
network. S.P. implemented the code of DeepCristae and the Jupyter notebooks. L.L., J.S., and
C-A.V-C. designed the biological experiments. L.L., and J.S. prepared samples. L.L. with the
aid of J.D. performed the acquisitions. L.L. and C-A.V-C analyzed and organized the biological
data. L.L., S.P., J.S. and A.B. prepared figures and videos. T.L. and Z.C. provided the
mitochondrial dye. A.B., S.P., J.S., L.L. and C-A.V-C. wrote the manuscript. All authors
critically discussed the results and commented on the manuscript. A.B., J.S. and C.K.
supervised the research.
Competing interests
Z.C. is an inventor of the patent on the mitochondria dye described in this work. All other
authors have no competing interests to declare.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
19
Figures
Figure 1. Overview of DeepCristae. Training step: (a) Acquisition of 33 high-resolution (HR)
2D STED images of live RPE1 cells stained with PKMITO-Orange, for mitochondria. From
these HR images a dataset is obtained: (b) deconvolution of the HR images using a
Richardson-Lucy algorithm to enhance the mitochondria cristae. The resulting images
are used as ground truths to train the network; (c) resolution degradation of the images by
applying Gaussian filtering and by corrupting images with Poisson-Gaussian noise. The
resulting images are denoted . The obtained dataset is divided into 24 training
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
20
images and 9 test images. To increase the size of the training set, the images and
ground truth are then augmented (d) and sampled in patches of size 128x128 pixels
(e). We finally obtained 1824 pairs of ground truths (blue) and low-resolution input images
(orange) to train our network (f). The training is performed by minimizing our SCoP loss,
especially dedicated to restoring mitochondria pixels. Inference step: (g) Long-term and fast
acquisition with low illumination of live mitochondria. Note that if the training was performed on
degraded STED images, the inference can be made on other microscopy modalities (e.g.,
STED, Live-SR and LLSM). (h, i) Frame-by-frame restoration of the acquired sequence by our
previously trained DeepCristae network, allowing observation of the mitochondrial cristae
dynamics at high resolution.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
21
Figure 2. DeepCristae outperforms state-of-the-art methods for restoring mitochondria
cristae in low-resolution 2D STED images. (a) The image grid displays restoration results
of 3 test images from the dataset obtained with DeepCristae and three competitive deep
learning methods: RCAN37, SRResNet35 and CARE19. GT: ground truth. Pixel size: 25 nm.
Scale bar: 1 μm. White arrowheads indicate mitochondria with low contrast or in a noisy
background restored by CARE; to be compared with DeepCristae column (b) Quantitative
comparison of DeepCristae with conventional (Richardson-Lucy (RL)29,30, Wiener31,
SPITFIR(e)32, ND-SAFIR23, 2D median filter (Median 2D) and deep learning (Noise2Void33,
CARE19, RCAN37 and SRResNet35 ) image restoration methods. Metrics were computed on
the 9 test images of . Note that all deep learning methods were trained using the same
patches extracted from the training images of . Parameters used for conventional
methods are indicated in Supplementary Note 2.2.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
22
Figure 3. DeepCristae reveals mitochondria cristae from low resolution (LR) 2D STED
images. RPE1 cells were labeled with PKMITO-Orange prior to high resolution (HR) 2D STED
imaging. (a) Degradation of the HR image with a Gaussian filter ( = 3.25 pixels) and
additive Poisson-Gaussian noise ( = 4.0). The resulting image is part of the test dataset
of . Pixel size: 25 nm. Scale bar: 3 μm. (b-e) Zoom in on the colored insets depicted in
(a). Top, from left to right: thumbnails of the zoomed-in area on the LR image, the image
restored by DeepCristae and the ground truth (GT) image, respectively. The GT was obtained
by deconvolving the HR raw image with the RL algorithm to enhance mitochondria cristae.
Bottom: comparison of normalized intensity line profiles along a mitochondrion between the
three thumbnails. The yellow line indicated in the corresponding colored inset in (a) serves to
identify the fluorescence profile. Note that GT shows undesired high-intensity pixels in the
background due to noise amplification by the RL algorithm. Therefore, DeepCristae line
profiles are sometimes smoother than the one of the ground truth, especially in (d) and (e).
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
23
Figure 4. Reliability of image restoration by DeepCristae. First experiment (a-f):
assessment of the reliability of DeepCristae on real data by controlling the consistency
between DeepCristae restoration and a real “ground truth”. RPE1 cells were labeled with
PKMITO-Orange prior to low resolution (LR) and to high resolution (HR) 2D STED imaging,
acquired successively. (a, d) Two LR STED images (Inputs) were given as input to
DeepCristae for restoration and qualitatively compared to the corresponding HR STED
images. Notes that the HR data were deconvolved using the Richardson-Lucy algorithm to
enhance mitochondria cristae and are considered here as “ground truths” (GTs). Pixel size: 25
nm. Scale bar: 1 μm. (b, c) and (e,f) Zoom in on the colored insets depicted in (a) and (d),
respectively. Top, from left to right: thumbnails of the zoomed-in area on the LR image, the
image restored by DeepCristae and the ground truth (GT) image, respectively. Bottom:
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
24
comparison of normalized intensity line profiles along a mitochondrion between the three
thumbnails. The yellow line indicated in the corresponding colored inset in (a) and (d) serves
to identify the fluorescence profile. Note that the line profiles were taken in regions where small
mitochondrial displacements were observed. Despite an offset due to mitochondrial motion, a
consistency between the cristae restored by DeepCristae and the ones present in the ground
truths is observed overall. Second experiment (g-l): assessment of the reliability of
DeepCristae by studying the consistency between its predictions obtained with different
training. To that end, 10 DeepCristae neural networks were trained with different training data,
each one generated with our patch generation method applied to the 24 training images of
. Note that for this experiment, all networks were initialized with the same weights. (g)
From left to right: predictions of three DeepCristae networks on two test images of , the
average prediction over the 10 trainings and the corresponding pixel-wise normalized standard
deviation. Pixel size: 25 nm. Scale bar: 1 μm. (h-k) Comparison of normalized intensity line
profiles along a mitochondrion in (g) between the 10 trainings. The yellow line indicated in the
corresponding colored inset in (g) serves to identify the fluorescence profile. (l) Quantitative
comparison of the 10 DeepCristae models. Metrics were computed on the 9 test images of
.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
25
Figure 5. DeepCristae reveals mitochondria cristae from low resolution 2D live STED.
(a-c) 2D live STED imaging of RPE1 cells labeled with PKMITO-Orange. (a, b) Time series of
10 images (Δt~13s) using 2D high-resolution (HR) STED (Supplementary Note 1: Materials
and Methods). Pixel size is 50 x 50 nm in (a) and 25 x 25 nm in (b). Early and late time points
are shown in a, the 4 first time points are shown in b. Phototoxic damage as illustrated by the
swelling of the mitochondria. (c) Time series of 2D Fast STED, improving the time delay
between time points (Δt~5.9 s) and reducing mitochondrial damage. Each thumbnail is
diagonally divided for raw 2D Fast (LR) STED images (bottom-left part) and after DeepCristae
treatment (top-right). To match the training mitochondria settings (see Results “Reliability and
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
26
limits of image restoration by DeepCristae”), a rescaling of 1.87 is applied to the LR image
before DeepCristae inference. Scale Bars in a-c are: 2μm. (d) Lateral widths of 7 mitochondria
measured at each time point (series of 10 time points) in HR STED and in Fast (LR) STED,
before and after DeepCristae (line profiles as depicted as arrow heads in a t8 and b t2). Line
profiles were fitted to a Gaussian model and FWHM was measured, as indicated in
Supplementary Note 1: Materials and Methods. Data are expressed as mean ± SD. (e) Cristae
widths measured as in d for 30 cristae on 20 mitochondria (from 4 distinct image series) in
early and late time points of HR STED and 31 cristae in Fast (LR) STED, before (Raw; two
measures above the ordinate scale) and after DeepCristae (line profiles as illustrated in a t8,
b t4 and c t10). Data are expressed as mean ± SD. Student t-test: ** (p-val=0.0082), ***(p-
val=0.0003). (f) Distances between two cristae (Cristae intervals) measured in a peak-to-peak
intensity from plot profiles (RAW Fast STED was not possible). Early: N=56 from 2 series;
Late: N=80 in 2 series; DeepCristae: N=60 in 3 series, distributed in all time points. Data are
expressed as mean ± SD. F-test: ****(p-val<0.0001). (g) Table with complete statistics
including significance using Student- and Fisher-tests; ns= non-significant.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
27
Figure 6. DeepCristae restoration enhances cristae width resolution in 3D and 3D+live
imaging. (a) 2D plane (top) and 3D MIP (Maximum Intensity Projection of 14 planes) (bottom)
of cell labeled with PKMITO-Orange, acquired with a SD microscope with a Live-SR module,
before (left) and after DeepCristae (right). Thumbnails are zoomed areas corresponding to the
insets (red and blue) and are composites of raw and DeepCristae images. Color scale bar
indicates the bottom to top position of mitochondria (bottom right). (b) Cristae widths were
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
28
measured as in Fig. 5e; each individual measurement in DeepCristae restored images is
compared to its equivalent in raw images except for 10 cristae, not measurable in Raw (N=60
and N=60-10, respectively). Data are expressed as mean ± SD (DeepCristae=87 ± 1 nm; Raw
=149 ± 64 nm; Student t-test, [****] p<0.0001). (c) One section plane (top) and 3D
reconstructed MIP of 71 planes (bottom) of a cell acquired with a Lattice Light Sheet
microscope (LLSM) in a dithered mode, before (left) and after DeepCristae (right). Thumbnails
are zoomed areas corresponding to the insets (red and blue) and are composite of raw and
Deepcristae restored images. Color scale bar indicates the bottom to top position of
mitochondria (bottom right). (d) Cristae widths were measured as in Fig. 5e in DeepCristae
restored images and compared to its Raw equivalents, when possible (N=60 and N=60-19,
respectively). Data are expressed as mean ± SD (DeepCristae=94 ± 15 nm; Raw=339 ± 248
nm; Student t-test, [****] p<0.0001). (e, f) 3D+time imaging using Live-SR (e) or LLSM (f). MIP
of single time points are shown (left images). Insets indicated in red are zoomed in the
thumbnails (right image series) to illustrate fusion or fission dynamics of mitochondria. The
selected zoomed areas are shown at different time points before (left panel) and after (right
panel) DeepCristae restoration. Time frames between stacks are 5.6s and 1.31s in double
channel acquisition for Live-SR and LLSM, respectively. Scale bars = 5µm in full field image
and =1µm in zoomed thumbnails. Before DeepCristae restoration, rescaling of 2.6 and of 4.16
were first applied to each raw Live-SR and LLSM data, respectively (see Results “Reliability
and limits of image restoration by DeepCristae”).
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
29
Figure 7. DeepCristae reveals 3D+time cristae morphology during endo/lysosome
mitochondria interactions. RPE1 cells incubated 4 hours with Cell Mask PM Deep Red (red)
were labeled with PKMITO-Orange (green) for the last 15 minutes. (a) A Maximum Intensity
Projection (20 planes; Stack time=1.86s/channel, time point t1 over 60) image acquired with
Live-SR microscopy is shown before (top) and after (bottom) DeepCristae restoration of the
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
30
mitochondria (green channel) as well as after denoising (ND-SAFIR) and RL deconvolution of
the endo/lysosomes (red channel). Colored Insets indicate intracellular locations with dynamic
events of interest. (b) Thumbnails show selected time points of zoomed area as indicated by
insets in a. They are presented as paired images before (left panels) and after (right panels)
DeepCristae restoration. Both time points (left panels) and time frames in seconds (right
panels) are indicated for comparison with d. (c) 3D reconstruction (56 planes; stack
time=0.49s/channel, time point t1 over 200) of data acquired with LLSM after deskewing and
RL deconvolution is shown before (top) and after (bottom) DeepCristae restoration (green
channel). Colored Insets indicate intracellular locations with a focus on dynamic events of
interest. (d) shows thumbnails of zoomed areas indicated by insets in c. Note that for the “blue”
area, an early time (T105-T119) and a late (T190-T200) time series are represented, showing
mitochondria fusion and fission events, respectively. Orange arrows focus on local interactions
between endo/lysosomes and mitochondria. Scale bars in (a, c) and (b, d) are equal to 5µm
and 1µm, respectively. Note that in these experiments, before DeepCristae restoration,
rescalings of 2.6 and of 4.16 were first applied to each raw Live-SR and LLSM data,
respectively, to match the training mitochondria settings (see Results “Reliability and limits of
image restoration by DeepCristae”.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted July 5, 2023. ; https://doi.org/10.1101/2023.07.05.547594doi: bioRxiv preprint
