GloBIAS blog series

We aim to publish here:

• reflections and outcomes from the different workgroups on a bi-monthly schedule on this page

• reviews on the GloBIAS seminar series, this will be summaries based on transcripts from the YouTube videos (coming soon: including translations into several languages)

April 2025

Navigating the AI Frontier: Large Language Models in BioImage Analysis

by the GloBIAS Communication work group

French version

Spanish version

Simplified Chinese version

Introduction

The field of bioimage analysis is rapidly evolving, and large language models (LLMs) are emerging as powerful tools for researchers and analysts. This blog post summarizes key insights from a recent GloBIAS seminar on LLMs, offering an introduction tailored for life scientists. The presentation addressed the motivation, applications, and challenges of using LLMs in bioimage analysis, highlighting their potential to revolutionize how we extract meaningful data from biological images.

What are Large Language Models?

LLMs are a type of neural network that can be used for various tasks, including translation and code generation. While not originally designed for knowledge or information retrieval, LLMs are increasingly used to answer questions and provide insights. However, it’s important to be aware of potential pitfalls, such as "hallucinations," where the model generates incorrect or misleading information.

Key Applications in Bioimage Analysis

• Code Generation: LLMs can translate English text into executable code, such as Python, making it easier to automate image analysis tasks.

• Image Modification: LLMs can modify microscopy images based on textual instructions, such as blurring or enhancing image quality.

• Image Description: LLMs can describe the contents of an image, which is useful for automated image annotation.

Addressing Reproducibility

One common critique of LLMs is their limited reproducibility, as they may produce different outputs for the same input. However, using LLMs for code generation can mitigate this issue. The generated code, once executed, will consistently produce the same result.

How LLMs Work: A Simplified Overview

Traditional image processing architectures typically involve encoder-decoder networks, which transform an input image into an output image. LLMs, based on transformer architecture, differ by incorporating three elements: input, output, and a shifted output. This allows LLMs to translate between different forms of data, such as images and text.

Practical Techniques for Optimizing LLMs

• Prompt Engineering: Crafting specific prompts to guide the LLM towards the desired output.

• Retrieval Augmented Generation (RAG): Enhancing the accuracy of LLMs by integrating them with a domain-specific knowledge base.

• Function Calling: Using LLMs to identify and parameterize functions that can perform specific tasks.

• Fine-Tuning: Customizing a pre-trained LLM with domain-specific data to improve its performance.

Responsible Use of LLMs in Research

• Understand the Code: Always understand the code generated by LLMs and verify its correctness.

• Question Established Methods: Use manual measurements to check results.

• Disclose LLM Usage: Be transparent about using LLMs in your research methods.

• Share Prompts: Share effective prompts with the community to facilitate collective learning.

Conclusion

Large language models hold immense promise for bioimage analysis, offering tools for automation, annotation, and data extraction. By understanding their capabilities and limitations, and by adopting best practices for their use, bioimage analysts can harness the power of LLMs to drive new discoveries and insights.

Actionable Takeaways

• Explore LLMs for automating routine image analysis tasks.

• Experiment with prompt engineering to optimize LLM performance.

• Contribute to open-source benchmarks to evaluate and improve LLMs.

• Engage with the GloBIAS community to share knowledge and experiences.

This blog post aims to equip you with the knowledge to start exploring and integrating LLMs into your bioimage analysis workflows. The potential of these models is vast, and their responsible application promises to unlock new frontiers in life science research.

(Note: The blog post has been written to be formal yet accessible, suitable for professionals in the field. It incorporates key points, examples, and recommendations from the presentation, and encourages further exploration and community engagement.)

March 2025

Shaping the BioImage Analyst Community in East Asia: EABIAS

by Shao-Chun (Peggy) Hsu

As an imaging scientist in Taiwan for over 14 years, I had been working in isolation without the support of communities, relying primarily on self-learning to develop my skills. However, this began to change in 2024. Bioimage analysts from Taiwan and Japan came together to establish EABIAS, thanks to the efforts of Cheng-Yu, an image analyst studying in the UK. Born in Japan and raised in Taiwan, Cheng-Yu played a key role in connecting us.

With an introduction from “Mr. Japan” Kota Miura and financial support from GloBIAS, I had the honour of attending the first GloBIAS in-person workshop held in Gothenburg. 

Before the event, I was assigned to two working groups—one focused on studying the image analysis workflow from an assigned paper, and the other on hosting the career path session. To prepare, I attended two online meetings per week, making this the most intensive pre-workshop preparation I had ever experienced. Yet, it was absolutely worth it! Through these regular online meetings, I connected with bioimage analysts from various countries and also experienced how networking can be fostered through dedicated activities.

For the first time, I felt part of a like-minded community. I also realized that collaboration makes everything more enjoyable. Having spent years promoting image analysis at National Taiwan University, I began to ask myself: How could I reach more people through networking? One thing became clear—we need more bioimage analysts working together in Taiwan. 

On my long journey home, I started reaching out to friends working in core facilities across Taiwan, sharing the education materials generously provided by the community and discussing the possibility of organizing a local course series led by us. In our first year, we will start with ImageJ, the most widely used image analysis tool, laying the foundation for transitioning to Python in the second year. 

The first local workshop organized by EABIAS—ImageJ Micro-Image Analysis and Programming—will take place every Monday afternoon for seven weeks, starting on March 3rd. The course will be conducted in Chinese, and we still have available slots for online participants. If you prefer learning in your native language, we warmly invite you to join us!

More information:

• ImageJ Micro-Image Analysis and Programming (registration now closed)

• EABIAS

January 2025

GloBIAS Seminar Series

by Caterina Fuster-Barceló 

• Over the last few months, the Education and Training group has launched an exciting new initiative: the GloBIAS Seminar Series. This series is designed to support bioimage analysts and the broader community with informative talks that highlight the latest tools, software, and innovations in bioimage analysis. It's also a space for learning, sharing, and connecting. Inspired by initiatives like the Euro-BioImaging Virtual Pub, we recognized a strong interest in having a dedicated platform for presenting and discussing topics in bioimage analysis.

• The GloBIAS Seminar Series will feature one seminar during the last week of each month, with presentations scheduled across different time zones to accommodate diverse audiences. Each quarter will focus on a specific theme, allowing us to delve deeper into key topics over three consecutive months. This approach ensures both continuity and a range of perspectives on each topic before moving on to the next.

• In the first quarter, we focused on “LLMs for Bioimage Analysis,” featuring presentations by Robert Haase, Caterina Fuster-Barceló, and Loïc Alain Royer. They each showcased unique tools and approaches for integrating large language models (LLMs) into bioimage analysis. Recordings of these sessions are available on our YouTube channel.

• Our current theme is Infrastructure for Deploying Image Analysis Workflows. So far, we've had Ankur Kumar present on “Designing and Implementing Systems to Process Petabyte-Scale Imaging Datasets,” and Christian Tischer discussing “NextFlow for Bioimage Analysis.” Coming up, we’re excited to have Beatriz Serrano-Solano join us on Wednesday, February 26, 2025, from 2:00 pm to 3:30 pm CET to discuss “Image Analysis Using Galaxy.” This session promises to be an excellent addition to our line-up!

• We encourage presenters and participants alike to continue any unanswered questions or discussions in the Image.sc forum, a platform open to the entire community. Ongoing conversations led by Robert Haase and Caterina Fuster-Barceló can be found here and here.

• There are many more topics on the horizon, and we invite you to join the workgroup, suggest ideas, or even propose talks. Thank you all of you for being part of the journey!

• Don’t forget to register!