nature methodshttps://doi.org/10.1038/s41592-024-02295-6 Correspondence DL4MicEverywhere: deep learning for microscopy made flexible, shareable and reproducible Deep learning enables the trans - formative analysis of large multi- dimensional microscopy datasets, but barriers remain in implement - ing these advanced techniques1,2. Many researchers lack access to annotated data, high performance computing (HPC) resources and expertise to develop, train and deploy deep learning models. In recent years, several approaches have been developed to democratize deep learning usage in micros - copy2. Tools such as the BioImage Model Zoo facilitate sharing and reuse of pretrained models, distributing them as one-click image processing solutions3,4. Yet often, deep learn - ing models need to be trained or fine-tuned on the end-user dataset to perform well2,3,5. We previously released ZeroCostDL4Mic6, an online platform relying on Google Colab that helped democratize deep learning by provid - ing a zero-code interface to train and evaluate models capable of performing various bioim - age processing tasks, such as segmentation, object detection, denoising, super-resolution microscopy and image-to-image translation. Here, we introduce DL4MicEverywhere, an advancement of the ZeroCostDL4Mic 6 framework (Fig. 1 ). DL4MicEverywhere is a platform that lets users train and implement their models in different computational environments. These environments include Google Colab, personal computational resources such as a desktop or laptop, and HPC systems. This flexibility is achieved by encapsulating each deep learning technique in an interactive Jupyter Notebook within a Docker container, enabling others to replicate analyses consist-ently across multiple platforms. DL4MicEve-rywhere (https://github.com/HenriquesLab/ DL4MicEverywhere) enables users to install and interact with a large offering of standard - ized, user-friendly deep learning workflows, away from the limitations of proprietary plat- forms such as Google Colab and in a secure computational environment with controlled data privacy and resources. DL4MicEvery - where can be launched graphically, via X11 forwarding, or directly through a command line (headless mode), supporting HPC usage. This cross-platform containerization technol - ogy boosts the long-term platform’s sustain- ability and reproducibility, enhancing user convenience7. DL4MicEverywhere features a zero-code interface that handles all the behind-the- scenes complexities, so users no longer need to deal with Docker configuration and deploy - ment through a terminal. The intuitive inter- face abstracts away these technical details while providing a standardized Docker encap - sulation for executing advanced techniques reliably. Researchers can select a notebook, choose computing resources and run the cor - responding deep learning-powered analysis with just a few clicks (Fig. 1c–e). This allows users to train and apply models on various computing resources they control, eliminat - ing reliance on third-party platforms. Further - more, researchers can launch a notebook on local or remote systems with GPU acceleration whenever available, without worrying about complex software dependencies, Docker con - tainer management, or loss of access to deep learning frameworks (Fig. 1f–h). Compared to ZeroCostDL4Mic, DL4MicEverywhere dou - bles the number of deep learning approaches and provides new bioimaging analysis tasks, such as semantic segmentation, interactive instance segmentation, image registration, 3D single molecule localization microscopy, temporal and spatial upsampling, and image generation. The platform is designed to encourage the sharing and reuse of deep learn - ing workflows provided as Jupyter Notebooks, which are then integrated into the BioImage Model Zoo. DL4MicEverywhere is strength - ened by automated build pipelines8 that allow tracked versioning of ZeroCostDL4Mic note- books and the seamless integration of new trainable models contributed by the commu - nity as user-friendly notebooks independently of the original ZeroCostDL4Mic framework (Fig. 1b). DL4MicEverywhere handles the cor - responding testing and building of fully docu - mented and open-source containers, making it easy for researchers to share not just the latest method, but the full software environ - ment required to run it reliably. DL4MicEverywhere is an open-source ini - tiative that aims to make deep learning acces - sible to everyone by providing a flexible and community-driven platform. Encapsulating software in Docker containers makes it pos - sible to integrate new methods without wor - rying about complex installation procedures, enriching the microscopy community through support to developers to easily contribute new pipelines, encouraging participatory innova - tion. Users can rely on shared techniques while customizing models across diverse hardware, retaining control over data and analysis. The platform sets a baseline for the development and use of cutting-edge foundation models 9. By bundling these sophisticated models into shareable containers, researchers can easily exploit them in their microscopy applica - tions. It is noteworthy that containerization approaches can increase local storage usage. Compared to proprietary platforms, which are not universally accessible, DL4MicEve - rywhere simplifies complex deep learning workflows through open, easy-to-use graphi- cal user interfaces and automated pipelines. It leverages local computational resources, HPC and cloud-based solutions, which pro - vides flexibility for sensitive biomedical data where privacy risks may limit reliance on pub - lic cloud platforms. It also helps with continu - ously scaling data, such as high-throughput high-content imaging data, whose storage, dissemination and access often rely on institutional infrastructures with specific data-sharing protocols. The containerization of notebooks is secure, as Jupyter Notebook ports are virtualized, private and protected with tokens. DL4MicEverywhere also adheres to FAIR (findability, accessibility, interoper - ability and reusability) principles, enhanc - ing data-driven scientific discoverability10. We expect DL4MicEverywhere to represent an important step toward reliable, transpar - ent and participatory artificial intelligence in microscopy. Check for updates ----!@#$NewPage!@#$---- nature methods Correspondence#ZeroCostDL4Mic Developers Easy-to-use interface to train, evaluate and use modelsEasy to deploy,share, showcaseand benchmark Local, HPC or cloudDL container imagesModels Docker hubReproducible Transferable Transparent FAIRStandards Users Communitya Engagement b Notebooks ModelsImagesNotebookSearch or Build Identify Local Remoted Automated containerization Local or remote host Load dataDependencies Train modelRun container Define Docker image PredictQuality control e Run DL4MicEverywhere Notebook Run CloseImages folder Output folder GPUSelect: c Interface Original CellPose pix2pix Original Widefield DeepSTORMf Super-resolution h Segmentation g Artificial labelingDeep learning bioimage analysisBioImage.IO PullContinuousintegration External notebooksDL4MicEverywhere-compatibleBespoke notebooksHosted in DL4MicEverywhereInspired by Zero Cost DL4MicAutomatic testing Docker hub Fig. 1 | The DL4MicEverywhere platform. a, DL4MicEverywhere eases deep learning workflow sharing, deployment and showcasing by providing a user-friendly interactive environment to train and use models. Cross-platform compatibility ensures reproducible deep learning model training. DL4MicEverywhere contributes to deep learning standardization in bioimage analysis by promoting transferable, FAIR and transparent pipelines. The platform exports models compatible with the BioImage Model Zoo 3 and populates free and open source container images in Docker Hub for developers to reuse. b , DL4MicEverywhere accepts three types of notebook contribution: ZeroCostDL4Mic 6 notebooks, bespoke notebooks inspired by ZeroCostDL4Mic6, and notebooks hosted in external repositories that are compliant with our format. The requirements and format of these contributions are automatically tested. c, In the DL4MicEverywhere graphical user interface, the user chooses a notebook, images and output folder, and chooses a GPU-running model if possible. d, DL4MicEverywhere automatically identifies the system architecture and requirements, checks whether the corresponding Docker image is available in Docker hub to download, and it builds it otherwise. This image is used to create a Docker container: a functional instance of the image that gathers the code environment to use the chosen notebook. e , A Jupyter lab session is launched inside the Docker container to train, evaluate or use the chosen model within an interactive notebook, equivalent to ZeroCostDL4Mic 6 notebooks. f–h, DL4MicEverywhere enables the use of the same notebooks for super- resolution (f ), artificial labeling (g ) or segmentation (h ) pipelines, among many others, in different local or remote infrastructures such as workstations, the cloud or HPC clusters. ----!@#$NewPage!@#$---- nature methods CorrespondenceCode availability The source code, documentation and tuto - rials for DL4MicEverywhere are available at https://github.com/HenriquesLab/DL4MicE - verywhere under a Creative Commons CC-BY- 4.0 license. Iván Hidalgo-Cenalmor   1, Joanna W. Pylvänäinen   2, Mariana G. Ferreira   1, Craig T . Russell   3, Alon Saguy4, Ignacio Arganda-Carreras5,6,7,8, Yoav Shechtman   4,9, AI4Life Horizon Europe Program Consortium*, Guillaume Jacquemet   2,10,11,12 , Ricardo Henriques   1,13 & Estibaliz Gómez-de-Mariscal   1 1Optical Cell Biology Group, Instituto Gulbenkian de Ciência, Oeiras, Portugal. 2Faculty of Science and Engineering, Cell Biology, Åbo Akademi University, Turku, Finland. 3European Molecular Biology Laboratory, European Bioinformatics Institute (EMBL-EBI), Wellcome Genome Campus, Cambridge, UK. 4Department of Biomedical Engineering, Technion – Israel Institute of Technology, Haifa, Israel. 5Department of Computer Science and Artificial Intelligence, University of the Basque Country (UPV/EHU), San Sebastián, Spain. 6IKERBASQUE, Basque Foundation for Science, Bilbao, Spain. 7Donostia International Physics Center (DIPC), San Sebastián, Spain. 8Biofisika Institute (CSIC-UPV/EHU), Leioa, Spain. 9Department of Mechanical Engineering, University of Texas at Austin, Austin, TX, USA. 10Turku Bioscience Centre, University of Turku and Åbo Akademi University, Turku, Finland. 11Turku Bioimaging, University of Turku and Åbo Akademi University, Turku, Finland. 12InFLAMES Research Flagship Center, Åbo Akademi University, Turku, Finland. 13UCL Laboratory for Molecular Cell Biology, University College London, London, UK. *A list of members and affiliations appears at the end of the paper.  e-mail: guillaume.jacquemet@abo.fi; rjhenriques@igc.gulbenkian.pt; egomez@igc.gulbenkian.pt Published online: xx xx xxxx References 1. Moen, E. et al. Nat. Methods 16, 1233–1246 (2019). 2. Pylvänäinen, J. W., Gómez-de-Mariscal, E., Henriques, R. & Jacquemet, G. Curr. Opin. Cell Biol. 85, 102271 (2023). 3. Wei, O. et al. Preprint at bioRxiv https://doi. org/10.1101/2022.06.07.495102 (2022). 4. Gómez-de-Mariscal, E. et al. Nat. Methods 18, 1192–1195 (2021). 5. Laine, R. F., Arganda-Carreras, I., Henriques, R. & Jacquemet, G. Nat. Methods 18, 1136–1144 (2021). 6. von Chamier, L. et al. Nat. Commun. 12, 2276 (2021). 7. Moreau, D., Wiebels, K. & Boettiger, C. Nat. Rev. Methods Primers 3, 1–16 (2023). 8. Beaulieu-Jones, B. K. & Greene, C. S. Nat. Biotechnol. 35, 342–346 (2017). 9. Bommasani, R. et al. Preprint at arXiv https://doi. org/10.48550/arXiv.2108.07258 (2022). 10. Wang, H. et al. Nature 620, 47–60 (2023). Acknowledgements We thank Amin Rezaei, Ainhoa Serrano, Pablo Alonso, Urtzi Beorlegui, Andoni Rodriguez, Erlantz Calvo, Soham Mandal and Virginie Uhlmann for their contributions to the ZeroCostDL4Mic notebook collection. I.H.-C., M.G.F., C.T.R., R.H. and E.G.-d.-M. received 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.). I.H.-C., M.G.F., E.G.-d.-M. and R.H. also acknowledge the support of the Gulbenkian Foundation (Fundação Calouste Gulbenkian) and the European Research Council (ERC) under the European Union’s Horizon 2020 Research and Innovation Programme (grant agreement 101001332 to R.H.). Funded by the European Union. Views and opinions expressed are, however, 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.-d.-M.), a Chan Zuckerberg Initiative Visual Proteomics Grant (vpi- 0000000044 with https://doi.org/10.37921/743590vtudfp to R.H.). R.H. also acknowledges the support of LS4FUTURE Associated Laboratory (LA/P/0087/2020). This work is partially supported by grant GIU19/027 (to I.A.-C.) funded by the University of the Basque Country (UPV/ EHU), grant PID2021-126701OB-I00 (to I.A.-C.) funded by the Ministerio de Ciencia, Innovación y Universidades, MICIU/AEI/10.13039/501100011033, and “ERDF A way of making Europe” (to I.A.-C.). This study was also supported by the Academy of Finland (338537 to G.J.), the Sigrid Juselius Foundation (to G.J.), the Cancer Society of Finland (Syöpäjärjestöt; to G.J.), and Solutions for Health strategic funding to Åbo Akademi University (to G.J.). This research was supported by the InFLAMES Flagship Programme of the Academy of Finland (decision number 337531). Author contributions I.H.-C., G.J., R.H. and E.G.-d.-M. conceived, designed and wrote the source code of the project with contributions from all co-authors; I.H.-C., J.P.W., M.G.F., C.R., A.S., Y.S., G.J., R.H. and E.G.-d.-M. tested the platform; I.H.-C., J.P.W., M.G.F., G.J., R.H. and E.G.-d.-M. wrote the user documentation; I.H.-C., G.J., R.H. and E.G.-d.-M. wrote the paper with input from all co-authors. F.J. (florian.jug@fht.org) serves as a contact for the consortium. Competing interests The authors declare no competing interests. Additional information Peer review information Nature Methods thanks Eugene Katrukha, Nils Körber and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. AI4Life Horizon Europe Program Consortium Arrate Muñoz-Barrutia14,15, Beatriz Serrano-Solano16, Caterina Fuster Barcelo14,15, Constantin Pape17, Craig T . Russell3, Emma Lundberg18,19,20, Estibaliz Gómez-de-Mariscal1, Florian Jug21, Joran Deschamps21, Iván Hidalgo-Cenalmor1, Mariana G. Ferreira1, Matthew Hartley3, Mehdi Seifi21, Ricardo Henriques1,13, Teresa Zulueta-Coarasa3, Vera Galinova21 & Wei Ouyang22 14Bioengineering Department, Universidad Carlos III de Madrid, Leganes, Spain. 15Instituto de Investigación Sanitaria Gregorio Marañón, Madrid, Spain. 16Euro-BioImaging ERIC Bio-Hub, European Molecular Biology Laboratory (EMBL) Heidelberg, Heidelberg, Germany. 17Georg-August-University Göttingen, Institute of Computer Science, Göttingen, Germany. 18Department of Pathology, Stanford University School of Medicine, Stanford, CA, USA. 19Science for Life Laboratory, School of Engineering Sciences in Chemistry, Biotechnology and Health, KTH Royal Institute of Technology, Stockholm, Sweden. 20Department of Bioengineering, Stanford University, Stanford, CA, USA. 21Fondazione Human Technopole, Milan, Italy. 22Department of Applied Physics, Science for Life Laboratory, KTH Royal Institute of Technology, Stockholm, Sweden.