Hi Dorian! Thanks for this interview. Let’s start with your professional journey. How did you end up managing the Challenge project?

It’s a long story! I’ve always been curious about many things, which made choosing a path difficult. I wasn’t particularly gifted in one subject, but rather reasonably good at everything—whether it was math, life sciences, physics, or even literature and ancient Greek. With no clear distinction to guide me, I felt torn between scientific and literary studies.

So, I chose a compromise: an Economic and Commercial Scientific Track (ECS) preparatory class, which allowed me to study various fields. 

Then, as I developed a real interest in mathematics and its applications, I moved towards a third-year degree in mathematics and computer science at Paris 1 Panthéon-Sorbonne University.

I continued with an interdisciplinary Master’s degree, focused on applied mathematics, economics, and finance. But here’s the thing: I hated finance. However, I didn’t quit—I kept going, because even in what seemed like a mistake, I knew there was something to learn.

During that time, I discovered data science and deep learning (AI), which fascinated me. In my second year, I specialized in optimization and data science, choosing courses in machine learning and AI.

And then came the turning point: an internship in histology image processing. Histology is the study of biological tissues (animal and plant) at the microscopic level, and my role was to develop an algorithm to help detect cancerous cells in histology slides, assisting anatomopathologists (doctors who analyze patients’ biological samples: cells, tissues and organs) in their diagnosis.

That’s when it all clicked.

For the first time, I found a field that combined everything I loved: mathematics and data science, but applied to bio-imaging, which brings together biology (the subject of the images) and physics (the imaging and acquisition process); and a literary dimension, through scientific writing and discussions with professionals from different backgrounds.

After graduating, I looked for a way to stay in this field, which led me to my current position at France-BioImaging.

To the younger generation: it’s okay if your path isn’t straightforward! You may make choices that don’t turn out as expected, but everything can be connected in unexpected ways. If you feel lost, just keep moving forward and stay true to yourself—you’ll find your way!

What is your role at France-BioImaging?

As a research software engineer, I am the referent and project coordinator for the FBI Challenges.

How did the idea for the first Challenge come about?

Thanks to discussions within the FBI community, we identified a critical question:

“How can we find four common organelles (i.e. cell components as nucleus or mitochondria) in living cells by fluorescence microscopy without the constraints of the acquisition process?”

After several exchanges and exploratory imaging sessions, we built a robust dataset and defined a specific challenge topic—the two essential pillars of any bioimaging competition.

This led to “Light My Cells”, a challenge focused on using deep learning (AI) to reconstruct sharp fluorescence images from transmitted light images.

Why are competitions like Challenge valuable for bioimaging?

Advancing data science applications in bioimaging requires a deep understanding of both the imaging process (biology and microscopy) and data science techniques tailored to specific tasks.

For example, in our context, it is crucial to understand the advantages and limitations of fluorescence microscopy vs. transmitted light microscopy:

• Fluorescence microscopy, widely used in cell imaging, relies on biochemical fluorescent labeling to highlight specific cellular structures. While effective, this technique is time-consuming, costly, and potentially harmful to the cells, as it can cause damage ranging from minor alterations to complete cell death. To preserve biological samples, it is essential to minimize the number of fluorescent labels used in live-cell experiments.
• Transmitted light microscopy—including bright field, phase contrast, and DIC (Differential Interference Contrast) microscopy—offers a non-invasive alternative. These label-free techniques reduce phototoxicity and cell alteration, while still providing valuable structural information.

This raises an important question: Can we computationally generate fluorescence microscopy images using only transmitted light microscopy?

This was the goal of the “Light My Cells” Challenge. With “Light My Cells”, we sought to:

• Develop deep learning methods capable of generating multi-channel fluorescence images from a single transmitted light image.
• Encourage innovation in AI architectures and loss functions to handle challenges such as missing organelles and microscope variability (e.g. magnification, depth of field, numerical aperture).
• Promote open science by providing a publicly available training dataset and integrating winning AI models into bioimaging software.

Beyond this competition, Challenge encourages participants to push the boundaries of AI, innovate, and contribute to bioimaging research. The ultimate vision? Achieving fluorescence imaging in silico—generating fluorescence-like images directly from transmitted light data, without the need for physical labeling.

What has been the most challenging part for you as the organizer?

The entire challenge was challenging!

What are your hopes for the future of the Challenge?

If I can dream a little, I’d like to achieve at least 2 things:

1. Develop a large-scale, open-source bioimaging dataset in open format (ome.zarr), covering diverse biological models (fauna and flora) under a standardized acquisition framework with comprehensive metadata—like a smaller ImageNet for biology (ImageNet is a large heterogeneous dataset used to train AI for image recognition).
   
2. Launch a ‘Microscope Metrology’ challenge to develop methods for accurately recovering the Point Spread Function (PSF) directly from microscopy images. In fluorescence microscopy, images are inherently blurred, and the PSF defines how a single point appears in an image. However, current methods rely on approximations between: Theoretical PSF (ideal, physics-based model), Experimental PSF (measured from calibration samples), Real PSF (true system response), Effective PSF (influenced by optics, noise, and reconstruction). By improving PSF estimation, this challenge may enable technicians to assess microscope quality faster and standardize calibration while helping biologists optimize imaging parameters and enhance quantitative analysis.

Thank you, Dorian! Good luck with the evaluation phase of the ‘Fuse My Cells’ Challenge and the selection of the three winners. See you in April for the results announcement and ISBI 2025 participation!