bioRxiv preprint doi: https://doi.org/10.1101/2025.04.09.647871; this version posted April 10, 2025. The copyright holder for this preprint (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was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.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 made
available under aCC-BY 4.0 International license.

ReScale4DL: Balancing Pixel and Contextual
Information for Enhanced Bioimage
Segmentation
Mariana G. Ferreira1,2,3 , Bruno M. Saraiva1,2,3 , António D. Brito1,4 , Mariana G. Pinho4 , Ricardo Henriques1,2,5, , and Estibaliz
Gómez-de-Mariscal1,2,3,
1

AI-driven Optical Biology Laboratory, Instituto de Tecnologia Química e Biológica António Xavier, Universidade Nova de Lisboa, Oeiras, Portugal
2
Optical Cell Biology Laboratory, Instituto Gulbenkian de Ciência, Oeiras, Portugal
3
Optical Cell Biology Laboratory, Gulbenkian Institute for Molecular Medicine, Oeiras, Portugal
4
Bacterial Cell Biology Laboratory, Instituto de Tecnologia Química e Biológica António Xavier, Universidade Nova de Lisboa, Oeiras, Portugal
5
UCL Laboratory for Molecular Cell Biology, University College London, London, United Kingdom

Deep learning has established itself as the state-of-the-art approach for segmentation in bioimage analysis. However, these
powerful algorithms present an intriguing paradox regarding
image resolution: contrary to intuition, lower-resolution images can yield superior performance for specific image analysis carried out by deep learning. This phenomenon is particularly significant in microscopy imaging, where high-resolution
acquisitions come with substantial costs in throughput, data
storage requirements, and potential photodamage to specimens.
Through systematic experimentation, we evaluate how varying
image resolution impacts deep learning performance in cellular image segmentation tasks. We trained popular architectures on datasets downsampled to 6-50% of their original resolution, mimicking acquisitions at lower image magnification, and
compared their performance against models trained on nativeresolution images. Our results show that segmentation accuracy
either improves (by up to 25% of mean Intersection over Union
(IoU)) or experiences only minimal degradation (< 5% of mean
IoU) when using images downsampled by up to a factor of 4
(25% of the original resolution). This downsampling proportionally increases information throughput while reducing data
storage requirements and inference time. With these findings,
we contribute systematic guidelines to deep learning practitioners in creating efficient experimental pipelines for image-driven
discoveries. This approach improves the sustainability and costeffectiveness of bioimaging studies by reducing data and computing needs while optimising microscopy techniques.
deep learning | microscopy | image segmentation | receptive field | throughput
| image resolution
Correspondence: r.henriques@itqb.unl.pt | estibaliz.gomez@itqb.unl.pt

Introduction
Modern microscopy technologies have driven an unprecedented expansion in data volume. High-resolution techniques, particularly super-resolution microscopy (1), generate substantially larger datasets than traditional methods
when imaging equivalent sample areas. A single 3D multichannel cell image from a spinning disk confocal requires
approximately 1GB of storage (considering a size of 4MB
for each slice of a three-channel and 100 z-slices volume).
In contrast, the same field acquired with a super-resolution
laser scanning microscope would be approximately 60 times
larger (estimated from images acquired in-house). To ad-

Fig. 1. The Image Resolution Paradox. The counterintuitive relationship between
image resolution and deep learning performance emerges from the mismatch between pixel size and the network’s receptive field. a) Decreasing image resolution
(increasing pixel size) causes pixelation and limits observable details, yet may improve contextual information. b) The area observable within a convolutional neural
network’s (CNN) receptive field (magenta squares) determines the context available for predicting the value of a single pixel. c) Suboptimal pixel sizes lead to
impaired predictions—either fragmented segmentations caused by non-continuous
edges and inner pixels misclassified as background, referred to as false negatives
(too high resolution) or over-generalisation at edges (too low resolution). d) Optimal image resolution for CNN-based processing depends on object size rather than
maximal achievable microscopy resolution, and proper calibration can simultaneously enhance segmentation accuracy and experimental throughput.

dress the challenges in acquisition, storage, and processing
posed by these volumes of data, researchers increasingly rely
on deep learning approaches that automate analytical tasks
while delivering reliable and reproducible results. For this,
the bioimage analysis community has developed numerous
G. Ferreira et al.

|

bioRχiv

|

April 9, 2025

|

1–8


----!@#$NewPage!@#$----
bioRxiv preprint doi: https://doi.org/10.1101/2025.04.09.647871; this version posted April 10, 2025. The copyright holder for this preprint (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 made
available under aCC-BY 4.0 International license.

Fig. 2. ReScale4DL pipeline. Researchers can optimise the pixel size for a chosen deep learning architecture using their newly acquired and annotated images or on
existing datasets, which can be used as a starting point for the optimisation. A series of image rescaling factors is applied to the data, a deep learning model is trained for
each rescaled dataset, and different accuracy metrics and features are computed. Once an ideal target metric is chosen, such as the accuracy of the image processing
performance, the recovery of an image feature, or a good compromise between the throughput and accuracy, researchers can identify an adequate pixel size and launch their
imaging experiments with optimised image resolution. The results presented in this manuscript were obtained by following the ReScale4DL pipeline.

user-friendly solutions (2, 3), including standalone software
and platform-integrated plugins (4–11), enabling researchers
to leverage these powerful methodologies.
Yet, bioimage analysis faces unique efficiency challenges
compared to other deep learning domains. The most significant bottlenecks include: limited computational resources
and storage capacity, particularly problematic for time-lapse
volumetric data; scarcity of large annotated datasets essential to train robust models and transfer learning (12, 13); and
insufficient understanding of the relationship between image quality metrics and computational performance requirements (14, 15).
In computer vision, image quality often equates to an algorithm’s efficiency in extracting relevant information under
hardware constraints. Counterintuitively, a low-resolution
image containing sufficient detail to perform the computational task accurately may be optimal, as it maximises the
information-to-resource ratio. For instance, cell nuclei appear more homogeneous at lower resolutions, enabling neural
networks to generalise more effectively. Conversely, at high
resolution, nuclei images can display heterogeneous structures, introducing noise during segmentation and potentially
increasing false negatives. This computational perspective
contrasts with traditional microscopy practices, where image
quality has been equated with higher resolution and magnification, assuming finer details provide more informative
content (1). However, while visually appealing, the wealth
of information in high-resolution images can be irrelevant
2

|

bioRχiv

or even detrimental to computational analysis, as our results
here demonstrate.
We present this phenomenon as the image resolution paradox: While high-resolution microscopy captures more detail,
this additional data may contribute minimally—or even negatively—to a certain analysis task while requiring substantially
greater resources. This paradox is particularly evident in deep
learning-based pipelines (Fig. 1). The limitation stems from
convolutional neural network’s (CNN) receptive fields—the
spatial region influencing a single output pixel. When image resolution exceeds the network’s architectural optimum,
pixels lack sufficient contextual information, compromising
the model’s performance. Although practitioners often address this by downsampling during pre-processing, incorporating resolution considerations directly into the experimental
design would enable acquisition at optimal resolutions from
the outset, simultaneously increasing throughput and reducing computational demands.
While this paradox is familiar to computer vision experts,
few tools explicitly account for it. Cellpose (16) represents
an exception, implementing automated re-sampling to optimise pixel-to-cell-diameter ratios. However, most users remain unaware of this critical consideration, leading to suboptimal application of deep learning methods—particularly
with architectures like vanilla U-Net—and unnecessary prioritisation of high-resolution acquisition despite its potential
drawbacks.
In this study, we quantitatively demonstrate this paradox and
G. Ferreira et al.

|

Resolution-Optimised Deep Learning


----!@#$NewPage!@#$----
bioRxiv preprint doi: https://doi.org/10.1101/2025.04.09.647871; this version posted April 10, 2025. The copyright holder for this preprint (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 made
available under aCC-BY 4.0 International license.

its impact on segmentation performance in bioimage analysis. For clarity, we define image resolution as the information content per pixel, separate from optical resolution
considerations. We systematically evaluate resolution effects
on model accuracy using a canonical 2D U-Net (17) for semantic segmentation and StarDist (9) for instance segmentation. Our experiments employ diverse specimens, including
high-magnification Caenorhabditis elegans (C. elegans) (18)
and brightfield Escherichia coli (E. coli) (19) images. Additionally, we analyse Staphylococcus aureus (S. aureus) imaged with Structural Illumination Microscopy (SIM), demonstrating that critical phenotypic changes remain detectable
even when reducing resolution by 16×, enabling proportional
throughput improvements. Beyond these examples, we provide a generalisable framework, ReScale4DL (https://
github.com/HenriquesLab/ReScale4DL), that researchers can implement to determine optimal pixel sizes for
their specific contexts, maximising both computational efficiency and information extraction (Fig. 2).

Results
High resolution and segmentation accuracy:
a
counter-intuitive relationship. We trained a 2D U-Net ar-

chitecture to semantically segment different cellular regions
- inner area, edges, and background - using two distinct
datasets: C. elegans captured with brightfield microscopy
and E. coli imaged via phase contrast microscopy. For each
dataset, we systematically rescaled the images to multiple
resolutions and trained separate U-Net models to evaluate
performance across these conditions (Fig. 2). Since segmentation accuracy depends on the relationship between image
resolution and object size, we quantified the performance
relative to the percentage of an object’s diameter covered by
a single pixel. Our experiments revealed a counterintuitive
relationship between resolution and segmentation quality
(Fig. 3). Contrary to conventional expectations, highresolution images often produced fragmented segmentation
masks with unexpected discontinuities. In the C. elegans
dataset, the inner masks produced at the original resolution
contained spurious holes where pixels were misclassified
as background (Fig. 3a)). This occurred because the model
lacked sufficient contextual information to confidently
identify inner and outer objects’ areas and one-pixel width
edges were under-represented for the given receptive field
and resolution. Namely, the Intersection over Union (IoU)
of semantic segmentations decreased by approximately 40%
compared to binary segmentation, which remained relatively
high (> 0.9) despite these deficiencies (Fig 3b)). Notably,
downsampling the images improved accuracy metrics for
both binary and semantic segmentation approaches, providing clear evidence that maximum resolution does not
necessarily yield optimal segmentation results. The E. coli
dataset analysis corroborated these findings (Fig. 3c-e)).
To ensure consistent comparison, we maintained identical UNet architecture and hyperparameters across all experiments
in Fig. 3, enabling direct assessment of resolution’s impact
G. Ferreira et al.

|

Resolution-Optimised Deep Learning

Fig. 3. Impact of image resolution on deep learning segmentation accuracy.
A 2D U-Net architecture was employed for semantic segmentation of cellular regions (inner area and boundary) in microscopy images. Images were systematically
resized to simulate varying acquisition resolutions, with a separate U-Net model
trained for each resolution factor. a) Segmentation results for Caenorhabditis elegans (C. elegans) using original images (pixel covering ∼1% of worm diameter) and
images downsampled by factor 23 (pixel covering ∼8% of worm diameter). Scale
bars: 100 and 25 pixels, respectively. b) (Boxplots) Intersection over Union (IoU)
distributions across downsampling factors ranging from ×1 to ×24 , showing how
segmentation accuracy varies with pixel size relative to object diameter. IoU was
calculated both for semantic segmentation (average IoU of edge and inner regions)
and binary masks (merged labels). (Purple line in logarithmic scale) Theoretical
throughput was calculated as the number of objects observable by a unit of time (τ )
given a certain resolution. τ is specific to each microscopy acquisition modality and
preserved across the different resolutions. c-e) Equivalent analysis for Escherichia
coli (E. coli) segmentation in phase contrast microscopy. c) Examples showing
fragmented results (2% diameter coverage), optimal segmentation (8% coverage),
and edge over-segmentation. Scale bars: 100, 25, and 10 pixels, respectively.
d) IoU distribution (boxplots) and theoretical throughput (purple line in logarithmic
scale) across varying bacteria diameter coverage ratios. e) Comparison between
diameter distributions estimated from U-Net segmentation versus ground truth, with
closest correspondence at 4% diameter coverage. Vertical dashed lines in panels
b), d), and e) indicate the original dataset resolution.

on model performance. Both datasets’ accuracy curves exhibited similar quadratic relationships between image resolution, object size, and IoU scores. For C. elegans, binary and
semantic segmentation achieved peak IoU when pixels covered 2-4% and 4-8% of worm diameter respectively, demonstrating that optimal resolution differed from the original microscopy acquisition (1% coverage). Similarly, the E. coli
dataset exhibited the highest IoU at comparable proportional
resolutions: 4-16% diameter coverage per pixel for binary
segmentation and 8-16% for semantic segmentation. This
consistency across different biological specimens reveals a
fundamental principle: our U-Net’s 50 × 50 pixel receptive
field imposes practical constraints. Objects with diameters
exceeding 25 pixels (corresponding to < 4% diameter coverage per pixel) become problematic as their size exceeds the
receptive field, while objects smaller than 6 pixels (> 16% diameter coverage) risk information dilution and loss through
the network’s operations. As exemplified, for any given CNN
bioRχiv

|

3


----!@#$NewPage!@#$----
bioRxiv preprint doi: https://doi.org/10.1101/2025.04.09.647871; this version posted April 10, 2025. The copyright holder for this preprint (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 made
available under aCC-BY 4.0 International license.

Fig. 4. Deep learning-driven resolution adjustment enables phenotype distinction while increasing imaging throughput. Staphylococcus aureus (S. aureus) bacteria were treated with FtsZ inhibitor PC190723, which prevents cell division and results in larger bacteria phenotypes (20). S. aureus (wild type (WT) and
treated) were then imaged using Structured Illumination Microscopy (SIM) and segmented with StarDist. a) Representative images of S. aureus with ground truth (GT)
segmentation and StarDist results at the original resolution (WT) and after 23 -factor
downsampling (both WT and treated conditions). Scale bars: 100, 25, and 25 pixels, respectively. b) (Boxplots) Distribution of Intersection over Union (IoU) values
of StarDist models trained and tested only with WT bacteria images across multiple resolution factors (upsampling: 2−2 to 20 ; downsampling: 21 to 24 ), showing
the relationship between segmentation accuracy and the percentage of bacterial
diameter covered by a single pixel. (Purple line in logarithmic scale) Theoretical
throughput was calculated as the number of objects observable by a unit of time (τ )
given a certain resolution. τ is specific to each microscopy acquisition modality and
preserved across the different resolutions. c) Diameter distributions across resolution scales estimated from GT and the StarDist results for WT bacteria images in
b). d) Comparison of the diameter distributions across various resolution scales for
WT and treated bacteria. StarDist was trained with a mixed dataset containing WT
and PC190723-treated S. aureus images in this case.

architecture, there exists an optimal resolution that maximises segmentation accuracy—and importantly, this resolution is typically significantly lower than what human observers prefer, allowing imaging pipelines with substantially
improved throughput.
Optimising image resolution preserves phenotypic discrimination while maximising experimental
throughput. Building upon our findings, we demonstrate

how optimising image resolution preserves segmentation
performance while enhancing experimental throughput. We
tested this approach using Staphylococcus aureus (S. aureus)
bacteria treated with antibiotic PC190723 and imaged via
SIM. PC190723, a FtsZ inhibitor, prevents cell division and
produces enlarged bacteria (20). To quantitatively assess
this morphological effect, we employed StarDist—a deep
learning approach optimised for instance segmentation of
star-convex objects like round bacteria—to segment individual cells and measure diameter distributions (Fig. 4a)).
As in our previous experiments, IoU values across different
resolution factors revealed an optimal range where segmentation accuracy peaks. Notably, even at substantially reduced
resolutions, StarDist maintained robust performance: with
up to 11% of bacterial diameter covered by a single pixel,
IoU values exceeded 0.8, and even with 23% coverage,
IoU remained above 0.75 (Fig. 4b)). Analysis of wild-type
(WT) diameter distributions confirmed that measurements
derived from ground truth and StarDist predictions remained
4

|

bioRχiv

consistent across resolutions ranging from 0.75% to 6% of
diameter per pixel. Only at coverage ratios exceeding 11%
(downsampling factor > 23 ) did diameter estimates begin to
diverge, likely because objects were reduced to just a few
pixels in width (Fig. 4c)). Most importantly, when training
a StarDist model with WT and antibiotic-treated mixed
bacteria populations, the diameter differences remained
distinguishable across all sampling ratios, demonstrating that
phenotypic discrimination is preserved even at significantly
lower resolutions than typically employed (Fig. 4d)). Thus,
by strategically reducing acquisition resolution, researchers
can substantially increase imaging throughput and decrease
acquisition times (Figs. 3b), 3d) and 4b)), enabling more
comprehensive dataset collection for statistically robust
experimental analyses.

Discussion
Our systematic investigation of the resolution paradox in
deep learning-driven bioimage analysis reveals a fundamental trade-off between spatial resolution and computational
performance. Our results demonstrate that popular methods
like U-Net and StarDist (which often use a U-Net architecture) achieve optimal accuracy when image resolution is calibrated to match the network’s receptive field with biological
object dimensions, rather than maximising resolution. This
finding challenges conventional microscopy practices while
offering significant practical benefits for experimental workflows.
The paradox emerges from fundamental neural network characteristics. When object features exceed the receptive field
(high resolution), networks lack sufficient context for consistent boundary prediction, resulting in fragmented masks with
spurious discontinuities. Conversely, excessively downsampled images (low resolution) compress critical morphological features beyond recognition. Our quantitative analysis
reveals a predictable quadratic relationship between object
diameter coverage per pixel and segmentation accuracy, with
optimal performance for the 2D U-Net consistently occurring
when pixels cover 4-16% of object diameter—irrespective of
the type of specimen or imaging modality. This optimal range
represents a "sweet spot" where networks balance local detail
with global context.
Most significantly, we demonstrated that models trained on
optimally downsampled images maintain phenotypic discrimination capabilities. In the S. aureus antibiotic response
experiment, critical morphological differences remained detectable even at 16× lower resolution than commonly used
(Fig. 4). This finding has profound implications for experimental design: researchers can substantially increase
imaging throughput, reduce photobleaching and phototoxicity, minimise data storage requirements, and accelerate
analysis pipelines—all without compromising biological insights (21, 22).
Although these principles are established in computer vision
communities, they remain underappreciated in life sciences,
where imaging protocols continue to emphasise the maximum achievable resolution regardless of computational analG. Ferreira et al.

|

Resolution-Optimised Deep Learning


----!@#$NewPage!@#$----
bioRxiv preprint doi: https://doi.org/10.1101/2025.04.09.647871; this version posted April 10, 2025. The copyright holder for this preprint (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 made
available under aCC-BY 4.0 International license.

ysis requirements. Our work provides a practical framework
for biologists to determine optimal resolution parameters for
their specific experimental contexts. By conducting preliminary resolution calibration studies before full-scale data acquisition, researchers can design more efficient imaging protocols that balance biological information content with computational performance.
The paradox persists in emerging transformer-based architectures despite their theoretical multi-scale capabilities. As
demonstrated through MicroSAM (23) foundation model
evaluations, fixed context windows in pre-trained models remain susceptible to resolution mismatches (Note 2, Fig. 5).
However, the growing ecosystem of interactive tools like
BioImage Model Zoo (6), MicroSAM (23) and DinoSim
(24) creates opportunities for rapid empirical optimisation,
turning this challenge into an accessible experimental design
parameter.
Democratising AI in bioimaging requires deeper integration
of data physics with model architectures. Current trends
favour pre-trained "smart" solutions over adaptable systems,
risking suboptimal application to novel experimental contexts. Future developments should prioritise dynamic acquisition systems that adjust resolution in real-time based on ongoing analysis, coupled with neural architectures explicitly
encoding multi-scale relationships. By embedding these considerations into standard workflows, the field can achieve reproducible, resource-efficient microscopy pipelines that balance information content with computational sustainability.

Methods
Biological image datasets. Publicly available high-

resolution Caenorhabditis elegans (C. elegans) minimum
projection brightfield microscopy images and instance segmentation masks were obtained from (18). Publicly available Escherichia coli (E. coli) brightfield microscopy images
and instance segmentation masks dataset was obtained from
(19), which is available on Zenodo 10.5281/zenodo.
5550934. In-house acquired super-resolution Staphylococcus aureus (S. aureus) cell wall structured illumination microscopy (SIM) fluorescence images and annotated instance
segmentation masks dataset available on Zenodo 10.5281/
zenodo.15169017.
Dataset

re-sampling. ReScale4DL
Python
library
(https://github.com/HenriquesLab/
ReScale4DL) was used to resize raw images and their
respective masks. For C. elegans and E. coli, the instance masks were first rescaled and then, the edges
and inner side were estimated using the ImageJ macro
available in DeepBacs GitHub repository (19) (https:
//github.com/HenriquesLab/DeepBacs/). Upsampling was computed with catmull-rom interpolation
and nearest-neighbor interpolation methods for the raw
and the mask images, respectively (all available within
the NanoPyx Python package (25)). Downsampling was
computed with binning and transform.rescale from
scikit-image for the raw images and the masks respecG. Ferreira et al.

|

Resolution-Optimised Deep Learning

tively. C. elegans dataset was downsampled by factors 2, 22 ,
23 , and 24 . E. coli was upsampled by factors 2 and 22 and
downsampled by factors 2 and 22 . S. aureus was upsampled
by factors 2 and 22 and downsampled by factors 2, 22 , 23 ,
and 24 .
Segmentation networks. The 2D Multilabel U-Net and 2D

StarDist (9) ZeroCostDL4Mic (2) notebooks were used with
DL4MicEverywhere (11) with the hyperparameter configuration in Table 1.
Evaluation. The Intersection over Union (IoU) scores visu-

alised in Fig.s 3 and 4 were computed for the results of all
the models specified in Table 1, and for the binary segmentation (C. elegans and E. coli), semantic segmentation (C.
elegans and E. coli) and instance segmentation (S. aureus)
results. For each object j in an image i, the diameter Dij
of non-rounded shapes (C. elegans and E. coli datasets) was
computed as the median of the Euclidean Distance Transform
(EDT) values contained in the object’s skeleton. For S. aureus dataset, Dij of each bacteria was estimated by solving
the equation of the circle area, given as
r
Aij
(1)
Dij = 2 ·
π
where Aij is the area of the object j in the image i, measured
as the sum of all the pixels in the object. To compute the
average portion of diameter covered by one pixel, we first
computed the average diameter for each scaling factor, Ds as


Oi
N
h
i
1 X 1 X
Ds =
Dij  , s ∈ 2Z[−2,4]
(2)
N
Oi
i=1

j=1

where Oi is the total number of objects in the image i, N is
the total number of images in the dataset and s is the rescaling factor. Then, Ds was inverted and multiplied by 100 to
compute the portion of diameter covered by one pixel.
To measure the throughput achievable for each rescaling factor s, we first estimated the average area in pixels of each
object in an image in pixels (As ). Then, we consider the size
of the achievable field of view (FOV) for a camera as the area
in pixels of the images in the original dataset (CAMF OV ).
Considering the throughput as the amount of information
recordable per unit of time, we computed the throughout as
h
i
CAMF OV
1
T Hs =
× , s ∈ 2Z[−2,4]
(3)
As
τ
where τ is the time required to acquire a F OV . In other
words, T Hs provides an estimate of the number of objects
that can be captured by a unit of time, yet, using Equation 3
with preliminary information about the sample’s morphology
and microscopy camera features, it is possible to compute
this value easily.
Code availability. The code for rescaling the images

and computing the metrics is available athttps://
github.com/HenriquesLab/ReScale4DL.
The
bioRχiv

|

5


----!@#$NewPage!@#$----
bioRxiv preprint doi: https://doi.org/10.1101/2025.04.09.647871; this version posted April 10, 2025. The copyright holder for this preprint (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 made
available under aCC-BY 4.0 International license.

notebooks to train, test and deploy inference are available through the DL4MicEvrywhere (https://github.
com/HenriquesLab/DL4MicEverywhere) platform.
availability. WT and antibiotic-treated superresolution Staphylococcus aureus (S. aureus) cell
wall SIM images together with instance segmentation annotations are available on Zenodo https:
//zenodo.org/records/15169018.

11.

Data

ACKNOWLEDGEMENTS
M.G.F., B.S., E.G.M., and R.H. acknowledge the support of the Gulbenkian
Foundation (Fundação Calouste Gulbenkian), the European Research Council
(ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No. 101001332) (to R.H.) and funding from the European Union through the Horizon Europe program (AI4LIFE project with grant
agreement 101057970-AI4LIFE and RT-SuperES project with grant agreement
101099654-RTSuperES to R.H.). Funded by the European Union. However,
the views and opinions expressed are those of the authors only and do not
necessarily reflect those of the European Union. Neither the European Union
nor the granting authority can be held responsible for them. This work was
also supported by a European Molecular Biology Organization (EMBO) installation grant (EMBO-2020-IG-4734 to R.H.), an EMBO postdoctoral fellowship
(EMBO ALTF 174-2022 to E.G.M.), a Chan Zuckerberg Initiative Visual Proteomics
Grant (vpi-0000000044 with https://doi.org/10.37921/743590vtudfp to R.H.) and a
Chan Zuckerberg Initiative Essential Open Source Software for Science (EOSS60000000260). E.G.M. acknowledges funding by Fundação para a Ciência e Tecnologia, Portugal (2023.09182.CEECIND/CP2854/CT0004). A.D.B. acknowledges
the FCT 2021.06849.BD fellowship.

12.

13.

14.

15.

16.

Bibliography
1. Lothar Schermelleh, Alexia Ferrand, Thomas Huser, Christian Eggeling, Markus Sauer,
Oliver Biehlmaier, and Gregor P. C. Drummen. Super-resolution microscopy demystified. Nature Cell Biology, 21(1):72–84, January 2019. ISSN 1476-4679. doi: 10.1038/
s41556-018-0251-8. Number: 1 Publisher: Nature Publishing Group.
2. Lucas von Chamier, Romain F. Laine, Johanna Jukkala, Christoph Spahn, Daniel Krentzel,
Elias Nehme, Martina Lerche, Sara Hernández-Pérez, Pieta K. Mattila, Eleni Karinou, Séamus Holden, Ahmet Can Solak, Alexander Krull, Tim-Oliver Buchholz, Martin L. Jones,
Loïc A. Royer, Christophe Leterrier, Yoav Shechtman, Florian Jug, Mike Heilemann, Guillaume Jacquemet, and Ricardo Henriques. Democratising deep learning for microscopy
with ZeroCostDL4Mic. Nature Communications, 12(1):2276, April 2021. ISSN 2041-1723.
doi: 10.1038/s41467-021-22518-0. Publisher: Nature Publishing Group.
3. Joanna W. Pylvänäinen, Estibaliz Gómez-de Mariscal, Ricardo Henriques, and Guillaume
Jacquemet. Live-cell imaging in the deep learning era. Current Opinion in Cell Biology, 85:
102271, December 2023. ISSN 0955-0674. doi: 10.1016/j.ceb.2023.102271.
4. Estibaliz Gómez-de Mariscal, Carlos García-López-de Haro, Wei Ouyang, Laurène Donati,
Emma Lundberg, Michael Unser, Arrate Muñoz-Barrutia, and Daniel Sage. DeepImageJ: A
user-friendly environment to run deep learning models in ImageJ. Nature Methods, 18(10):
1192–1195, October 2021. ISSN 1548-7105. doi: 10.1038/s41592-021-01262-9. Publisher:
Nature Publishing Group.
5. Stuart Berg, Dominik Kutra, Thorben Kroeger, Christoph N. Straehle, Bernhard X. Kausler,
Carsten Haubold, Martin Schiegg, Janez Ales, Thorsten Beier, Markus Rudy, Kemal Eren,
Jaime I. Cervantes, Buote Xu, Fynn Beuttenmueller, Adrian Wolny, Chong Zhang, Ullrich
Koethe, Fred A. Hamprecht, and Anna Kreshuk. ilastik: interactive machine learning for
(bio)image analysis. Nature Methods, 16(12):1226–1232, December 2019. ISSN 15487105. doi: 10.1038/s41592-019-0582-9. Publisher: Nature Publishing Group.
6. Wei Ouyang, Fynn Beuttenmueller, Estibaliz Gómez-de Mariscal, Constantin Pape, Tom
Burke, Carlos Garcia-López-de Haro, Craig Russell, Lucía Moya-Sans, Cristina de-la
Torre-Gutiérrez, Deborah Schmidt, Dominik Kutra, Maksim Novikov, Martin Weigert, Uwe
Schmidt, Peter Bankhead, Guillaume Jacquemet, Daniel Sage, Ricardo Henriques, Arrate
Muñoz-Barrutia, Emma Lundberg, Florian Jug, and Anna Kreshuk. BioImage Model Zoo:
A Community-Driven Resource for Accessible Deep Learning in BioImage Analysis, June
2022. Pages: 2022.06.07.495102 Section: New Results.
7. Anastasia Mavropoulos, Chassidy Johnson, Vivian Lu, Jordan Nieto, Emilie C. Schneider,
Kiran Saini, Michael L. Phelan, Linda X. Hsie, Maggie J. Wang, Janifer Cruz, Jeanette Mei,
Julie J. Kim, Zhouyang Lian, Nianzhen Li, Stephane C. Boutet, Amy Y. Wong-Thai, Weibo
Yu, Qing-Yi Lu, Teresa Kim, Yipeng Geng, Maddison (Mahdokht) Masaeli, Thomas D. Lee,
and Jianyu Rao. Artificial Intelligence-Driven Morphology-Based Enrichment of Malignant
Cells from Body Fluid. Modern Pathology, 36(8), August 2023. ISSN 0893-3952, 15300285. doi: 10.1016/j.modpat.2023.100195. Publisher: Elsevier.
8. Benjamin Midtvedt, Saga Helgadottir, Aykut Argun, Jesús Pineda, Daniel Midtvedt, and Giovanni Volpe. Quantitative digital microscopy with deep learning. Applied Physics Reviews,
8(1):011310, February 2021. ISSN 1931-9401. doi: 10.1063/5.0034891.
9. Uwe Schmidt, Martin Weigert, Coleman Broaddus, and Gene Myers. Cell detection with
star-convex polygons. In Medical Image Computing and Computer Assisted Intervention
- MICCAI 2018 - 21st International Conference, Granada, Spain, September 16-20, 2018,
Proceedings, Part II, pages 265–273, 2018. doi: 10.1007/978-3-030-00934-2_30.
10. Daniel Franco-Barranco, Jesús A. Andrés-San Román, Ivan Hidalgo-Cenalmor, Lenka
Backová, Aitor González-Marfil, Clément Caporal, Anatole Chessel, Pedro Gómez-Gálvez,

6

|

bioRχiv

17.

18.

19.

20.

21.

22.

23.

24.

25.

26.

Luis M. Escudero, Donglai Wei, Arrate Muñoz-Barrutia, and Ignacio Arganda-Carreras. BiaPy: Accessible deep learning on bioimages, January 2025. Pages: 2024.02.03.576026
Section: New Results.
Iván Hidalgo-Cenalmor, Joanna W. Pylvänäinen, Mariana G. Ferreira, Craig T. Russell, Alon
Saguy, Ignacio Arganda-Carreras, Yoav Shechtman, Guillaume Jacquemet, Ricardo Henriques, and Estibaliz Gómez-de Mariscal. DL4MicEverywhere: deep learning for microscopy
made flexible, shareable and reproducible. Nature Methods, 21(6):925–927, June 2024.
ISSN 1548-7105. doi: 10.1038/s41592-024-02295-6. Publisher: Nature Publishing Group.
Damian Dalle Nogare, Matthew Hartley, Joran Deschamps, Jan Ellenberg, and Florian
Jug. Using AI in bioimage analysis to elevate the rate of scientific discovery as a community. Nature Methods, 20(7):973–975, July 2023. ISSN 1548-7105. doi: 10.1038/
s41592-023-01929-5. Publisher: Nature Publishing Group.
Christopher Schmied, Michael S. Nelson, Sergiy Avilov, Gert-Jan Bakker, Cristina Bertocchi, Johanna Bischof, Ulrike Boehm, Jan Brocher, Mariana T. Carvalho, Catalin Chiritescu,
Jana Christopher, Beth A. Cimini, Eduardo Conde-Sousa, Michael Ebner, Rupert Ecker,
Kevin Eliceiri, Julia Fernandez-Rodriguez, Nathalie Gaudreault, Laurent Gelman, David
Grunwald, Tingting Gu, Nadia Halidi, Mathias Hammer, Matthew Hartley, Marie Held, Florian Jug, Varun Kapoor, Ayse Aslihan Koksoy, Judith Lacoste, Sylvia Le Dévédec, Sylvie
Le Guyader, Penghuan Liu, Gabriel G. Martins, Aastha Mathur, Kota Miura, Paula Montero Llopis, Roland Nitschke, Alison North, Adam C. Parslow, Alex Payne-Dwyer, Laure
Plantard, Rizwan Ali, Britta Schroth-Diez, Lucas Schütz, Ryan T. Scott, Arne Seitz, Olaf Selchow, Ved P. Sharma, Martin Spitaler, Sathya Srinivasan, Caterina Strambio-De-Castillia,
Douglas Taatjes, Christian Tischer, and Helena Klara Jambor. Community-developed
checklists for publishing images and image analyses. Nature Methods, 21(2):170–181,
February 2024. ISSN 1548-7105. doi: 10.1038/s41592-023-01987-9. Publisher: Nature
Publishing Group.
Virginie Uhlmann, Laurène Donati, and Daniel Sage. A Practical Guide to Supervised Deep
Learning for Bioimage Analysis: Challenges and good practices. IEEE Signal Processing
Magazine, 39(2):73–86, March 2022. ISSN 1558-0792. doi: 10.1109/MSP.2021.3123589.
Conference Name: IEEE Signal Processing Magazine.
Siân Culley, Alicia Cuber Caballero, Jemima J Burden, and Virginie Uhlmann. Made to
measure: An introduction to quantifying microscopy data in the life sciences. Journal
of Microscopy, 295(1):61–82, 2024. ISSN 1365-2818. doi: 10.1111/jmi.13208. _eprint:
https://onlinelibrary.wiley.com/doi/pdf/10.1111/jmi.13208.
Carsen Stringer, Tim Wang, Michalis Michaelos, and Marius Pachitariu. Cellpose: a generalist algorithm for cellular segmentation. Nature Methods, 18(1):100–106, January 2021.
ISSN 1548-7105. doi: 10.1038/s41592-020-01018-x. Publisher: Nature Publishing Group.
Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for
Biomedical Image Segmentation. In Nassir Navab, Joachim Hornegger, William M. Wells,
and Alejandro F. Frangi, editors, Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, pages 234–241, Cham, 2015. Springer International Publishing.
ISBN 978-3-319-24574-4. doi: 10.1007/978-3-319-24574-4_28.
Kevin J. Cutler, Carsen Stringer, Teresa W. Lo, Luca Rappez, Nicholas Stroustrup,
S. Brook Peterson, Paul A. Wiggins, and Joseph D. Mougous. Omnipose: a high-precision
morphology-independent solution for bacterial cell segmentation. Nature Methods, 19(11):
1438–1448, November 2022. ISSN 1548-7105. doi: 10.1038/s41592-022-01639-4. Publisher: Nature Publishing Group.
Christoph Spahn, Estibaliz Gómez-de Mariscal, Romain F. Laine, Pedro M. Pereira, Lucas
von Chamier, Mia Conduit, Mariana G. Pinho, Guillaume Jacquemet, Séamus Holden, Mike
Heilemann, and Ricardo Henriques. DeepBacs for multi-task bacterial image analysis using
open-source deep learning approaches. Communications Biology, 5(1):1–18, July 2022.
ISSN 2399-3642. doi: 10.1038/s42003-022-03634-z. Publisher: Nature Publishing Group.
David J. Haydon, Neil R. Stokes, Rebecca Ure, Greta Galbraith, James M. Bennett,
David R. Brown, Patrick J. Baker, Vladimir V. Barynin, David W. Rice, Sveta E. Sedelnikova,
Jonathan R. Heal, Joseph M. Sheridan, Sachin T. Aiwale, Pramod K. Chauhan, Anil Srivastava, Amit Taneja, Ian Collins, Jeff Errington, and Lloyd G. Czaplewski. An Inhibitor
of FtsZ with Potent and Selective Anti-Staphylococcal Activity. Science, 321(5896):1673–
1675, September 2008. doi: 10.1126/science.1159961. Publisher: American Association for
the Advancement of Science.
Guillaume Jacquemet, Alexandre F. Carisey, Hellyeh Hamidi, Ricardo Henriques, and
Christophe Leterrier. The cell biologist’s guide to super-resolution microscopy. Journal
of Cell Science, 133(11), June 2020. ISSN 14779137. doi: 10.1242/JCS.240713.
Estibaliz Gómez-de Mariscal, Mario Del Rosario, Joanna W. Pylvänäinen, Guillaume
Jacquemet, and Ricardo Henriques. Harnessing artificial intelligence to reduce phototoxicity in live imaging. Journal of Cell Science, 137(3):jcs261545, February 2024. ISSN
0021-9533, 1477-9137. doi: 10.1242/jcs.261545.
Anwai Archit, Luca Freckmann, Sushmita Nair, Nabeel Khalid, Paul Hilt, Vikas Rajashekar, Marei Freitag, Carolin Teuber, Genevieve Buckley, Sebastian von Haaren, Sagnik Gupta, Andreas Dengel, Sheraz Ahmed, and Constantin Pape. Segment Anything
for Microscopy. Nature Methods, 22(3):579–591, March 2025. ISSN 1548-7105. doi:
10.1038/s41592-024-02580-4. Publisher: Nature Publishing Group.
Aitor González-Marfil, Estibaliz Gómez-de Mariscal, and Ignacio Arganda-Carreras. DINOSim: Zero-Shot Object Detection and Semantic Segmentation on Electron Microscopy
Images, March 2025. Pages: 2025.03.09.642092 Section: New Results.
Bruno M. Saraiva, Inês Cunha, António D. Brito, Gautier Follain, Raquel Portela, Robert
Haase, Pedro M. Pereira, Guillaume Jacquemet, and Ricardo Henriques. Efficiently
accelerated bioimage analysis with NanoPyx, a Liquid Engine-powered Python framework. Nature Methods, 22(2):283–286, February 2025. ISSN 1548-7105. doi: 10.1038/
s41592-024-02562-6. Publisher: Nature Publishing Group.
Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson,
Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross
Girshick. Segment Anything. April 2023.

G. Ferreira et al.

|

Resolution-Optimised Deep Learning


----!@#$NewPage!@#$----
bioRxiv preprint doi: https://doi.org/10.1101/2025.04.09.647871; this version posted April 10, 2025. The copyright holder for this preprint (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 made
available under aCC-BY 4.0 International license.

Supplementary Note 1: Deep learning model training hyperparameters
C. elegans
Network
2D U-Net
Epochs
1000
Patch size
512 (x2−2,−1,0 )
(per resizing
256 (x2)
factor)
128 (x22 )
64 (x23 )
32 (x24 )
Batch size
5
Validation [%]
10
Learning rate
0.0001
Pooling steps
2
Minimal fraction
0.05

E. coli
Network
Epochs
Patch size
(per resizing
factor)

2D U-Net
1000
512 (x2−2,−1,0 )
256 (x2)
128 (x22 )

Batch size
Validation [%]
Learning rate
Pooling steps
Minimal fraction

5
10
0.0001
2
0.05

S. aureus
Network
2D StarDist
Epochs
1000
Patch size
2048 (x2−2,−1,0 )
(per resizing
1184 (x2)
factor)
576 (x22 )
272 (x23 )
128 (x24 )
Batch size
2
Validation [%]
10
Learning rate
0.0003
Grid parameter
2
Number of rays
32

Table 1. The network architecture and the training parameters for each dataset were preserved. The patch size for the input during
training was dynamically changed with the image resizing to maintain the same field of view, except for the upsampling due to limited
GPU capabilities. The values 2−2,−1 correspond to image upsamplings, 20 is the original image resolution and 21,...,4 correspond to
image downsampling.

Supplementary Note 2: Optimising the pixel size for foundation models
Here we provide a visual example of segmentation performance using foundation models for images with different pixel sizes.
We performed instance segmentation of Escherichia coli (E. coli), Caenorhabditis elegans (C. elegans), and Staphylococcus
aureus (S. aureus) samples using MicroSAM (23) software, with the same rescaling factors used for the Figs. 2 and 4. For
simplicity, we used their 2D Annotator plugin in Napari to deploy interactive segmentation with a single-point prompt-based
approach, followed by the automatic instance segmentation. Rescaled images were individually segmented by manually pointing to the same structural landmark each time. For all experiments, we chose the Vision Transformer (ViT) Large backbone
architecture, a fine-tuned version of the original Segment Anything Model (SAM) (26) for cellular and nuclear segmentation in
light microscopy (ViT-l-lm). Importantly, the segmentation results provided here may not be optimal and only aim to show
the variability of the model according to the image resolution under the same conditions. Indeed, the results could be improved
by changing the prompt or by increasing their number.

G. Ferreira et al.

|

Resolution-Optimised Deep Learning

bioRχiv

|

7


----!@#$NewPage!@#$----
bioRxiv preprint doi: https://doi.org/10.1101/2025.04.09.647871; this version posted April 10, 2025. The copyright holder for this preprint (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 made
available under aCC-BY 4.0 International license.

Fig. 5. Segmentation results of Segment Anything for Microscopy on images with different pixel sizes using a single point
prompting. MicroSAM (23) was used to obtain the instance segmentation of a) S. aureus, b) E. coli and c) C. elegans example images.
For simplicity, we followed a single point-wise prompting strategy, pointing to the same object on each rescaled image (i.e., the manual
prompt was the same on each panel), and run the automatic instance segmentation pipeline. All the images matched model input
dimensions (512 × 512 pixels) to avoid internal tiling or image re-scaling. The images that had smaller dimensions were padded with
zeros. Scale bars correspond to 25 µm for a), and 50 µm for b-c). Square boxes correspond to the same physical area on each image.

8

|

bioRχiv

G. Ferreira et al.

|

Resolution-Optimised Deep Learning