scieee AI-readable full text Open interactive document viewer

How LLMs impact BioImage Data Science

Haase, Robert

Abstract

In this slide-deck we dive into large language models for bioimage data science, focusing on code generation for bio image analysis. The slides also give an outlook on how AI-systems may change data analysis in the future through data analysis code generation.

Full text

CENTER FOR SCALABLE DATA ANALYTICS AND ARTIFICIAL INTELLIGENCE NATIONAL RESEARCH DATA INFRASTRUCTURE FOR MICROSCOPY AND BIOIMAGE ANALYSIS How LLMs impact BioImage Data Science Robert Haase These slides can be reused under the terms of the CC-BY 4.0 license unless mentioned otherwise. Funded by https://doi.org/10.5281/zenodo.17669681 LLMs Bio-Image Analysis, NFDI4DS Conference, Robert Haase, November 25th 2025 2 The consortium NFDI4BIOIMAGE 10/04/24 2 2 2 11 co-applicant institutions (6 Task Areas) Lead institution: Heinrich Heine University Düsseldorf Speaker: Prof. Dr. Stefanie Weidtkamp-Peters 12 participating institutions community use cases data stewards (DaSts) & research software engineers IT IT IT IT infrastructure (storage & playground: @Uni Münster & @Uni Freiburg) Community-oriented services (implemented or planned) •Help Desk for use community support •Training portfolio •Bioimage RDM playground infrastructure •Supporting reproducible image analysis But not: a central data archive for all bioimaging data Modified after: Stefanie Weidtkamp-Peters. (2023, Juni 22). NFDI4BIOIMAGE - National Research Data Infrastructure for Microscopy and BioImage Analysis - Online Kick-Off 2023. Zenodo. https://doi.org/10.5281/zenodo.8070038 LLMs Bio-Image Analysis, NFDI4DS Conference, Robert Haase, November 25th 2025 3 NFDI4BIOIMAGE Objectives 310/04/24 Modified after: Stefanie Weidtkamp-Peters. (2023, Juni 22). NFDI4BIOIMAGE - National Research Data Infrastructure for Microscopy and BioImage Analysis - Online Kick-Off 2023. Zenodo. https://doi.org/10.5281/zenodo.8070038 Out of scope: Medical, clinical, and preclinical imaging (MRT, PET, ultrasound, etc.) LLMs Bio-Image Analysis, NFDI4DS Conference, Robert Haase, November 25th 2025 4 Bio-image Analysis Biologist Domain-specialist (focused on real-world problems) Computer Scientist Method + infrastructure specialist (algorithm-centered) Bio-image Analyst  Generalist → (data-driven, service-oriented) Kubernetes cluster Many computers working together Tribolium castaneum Flour beetle LLMs Bio-Image Analysis, NFDI4DS Conference, Robert Haase, November 25th 2025 5 Bio-image Analysis •My job … LLMs Bio-Image Analysis, NFDI4DS Conference, Robert Haase, November 25th 2025 6 Bio-image Analysis •My job is changing, since we have ChatGPT Prompt: “Please segment the nuclei in the blue channel of the image” LLMs Bio-Image Analysis, NFDI4DS Conference, Robert Haase, November 25th 2025 7 Bio-image analysis beyond image segmentation •Crossing this border seems challenging “Frontier of perfectionism” Feature extraction Dimensionality reduction Classification Clustering Statistics Visualization Background removal Denoising Deconvolution Segmentation Insight Wide variety of [computational] skills required to achieve all this LLMs Bio-Image Analysis, NFDI4DS Conference, Robert Haase, November 25th 2025 8 Consulting LLMs LLMs Bio-Image Analysis, NFDI4DS Conference, Robert Haase, November 25th 2025 9 Consulting LLMs LLMs Bio-Image Analysis, NFDI4DS Conference, Robert Haase, November 25th 2025 26 Collaborative working with AI assistants •New AI-assisted coding solution appear daily •Human-human-AI interaction •Transparency •Knowledge exchange https://github.com/haesleinhuepf/git-bob/ https://www.nature.com/articles/s43588-025-00781-1 H Domain expert H Data analyst Github / Gitlab CI Hi! I have data and would like to measure x. I propose to try algorithm alg. git-bob comment on how to do this with Python AI git-bob LLM service provider 1 LLM service provider 2 Certainly! Try this: from algorithms import alg x = alg(data) boxplot(x) For result visualization let’s better use a violinplot. git-bob ask claude to try this! I created a notebook for using alg and drawing a violinplot: LLM service provider 3 LLM gpt-4o LLM claude LLM gemini That looks ok/correct! Nice! Thanks! LLMs Bio-Image Analysis, NFDI4DS Conference, Robert Haase, November 25th 2025 28 Summary & Outlook •Using AI / LLMs in [bioimage] data analysis seems more and more unavoidable •Maintaining good scientific practice is challenging (esp. human-AI peer-review) •Relying on institutional infrastructure exclusively is important but hard •Learn how to use AI responsibly (-> train the trainers) •We are on the edge towards AI-agents analyzing image data autonomously. https://doi.org/10.5281/zenodo.17669681 LLMs Bio-Image Analysis, NFDI4DS Conference, Robert Haase, November 25th 2025 29 Bio-Image Analysis Communities •Image/data analysis + data management support •Networking opportunities •Trainings https://nfdi4bioimage.de/help-desk/ https://www.globias.org/aboutglobias/globias-association https://image.sc/ Image Science Community Forum National Infrastructure for Research Data Management in BioImaging LLMs Bio-Image Analysis, NFDI4DS Conference, Robert Haase, November 25th 2025 30 Collaborators & contributors Christian Tischer (EMBL Heidelberg), Jean-Karim Hériché (EMBL Heidelberg), Nico Scherf (MPI CBS Leipzig), Loïc A. Royer (CZ Biohub), Elena Nicolay (UFZ Leipzig), Kevin Yamauchi (ETH Zurich / Basel), Seth Hinz, Teun A.P.M. Huijben, Jordão Bragantini (CZ Biohub San Francisco), llan Theodoro (CZ Biohub and the State University of Campinas, Brazil), Ian C. (University of British Columbia), Pradeep Rajasekhar (WEHI Australia) Acknowledgements Communities & platforms Lea Kabjesz, Lea Gihlein, Mara Lampert RH acknowledges the financial support by the Federal Ministry of Education and Research of Germany and by Sächsische Staatsministerium für Wissenschaft, Kultur und Tourismus in the programme Center of Excellence for AI-research „Center for Scalable Data Analytics and Artificial Intelligence Dresden/Leipzig“, project identification number: ScaDS.AI BIDS Team @ ScaDS.AI / UL Funding https://doi.org/10.5281/zenodo.17669681