Over the last decade, artificial intelligence has significantly transformed image analysis, with deep learning methods now achieving remarkable performance across a wide range of visual tasks. More recently, Large Language Models and promptable vision models like SAM (Segment Anything Model) have started to become part of our everyday scientific workflows. At the same time, many of us are still learning how these powerful models work, how they can help accelerate technical progress, and where their limitations lie.

This seminar offers an accessible overview of the evolution of AI for biomedical image analysis, from early neural networks to the transformer-based models that underpin many of today’s AI tools. Rather than focusing on technical details, I will present an intuitive understanding of the ideas behind these methods, and how the field has progressed from task-specific models to increasingly general and interactive systems. I will also discuss how technical progress can often depend just as much on domain knowledge as on the choice of AI model itself. In practice, domain knowledge must drive dataset design, annotation strategy, and evaluation criteria - key but often neglected aspects of AI-driven biomedical image analysis. To illustrate this, I will present three recent research projects in which a data-centric approach has been a key component.

This talk aims to offer a practical and realistic perspective on modern AI. We will see how, when combined with expertise in medicine and biology, these methods can help us address relevant problems more effectively and incorporate them more meaningfully into our scientific practice.