Blog
In this episode we interview Dr. Martin Jones on his journey from quantum physics to biomedical imaging, what CLEM (Correlative Light and Electron Microscopy) is and why it matters, the real promise and real limitations of AI in microscopy, and the power of combining data modalities. Martin is Deputy Head of Microscopy Prototyping in the Electron Microscopy STP at the Francis Crick Institute. He holds a PhD in physics and specialises in developing hardware and software solutions for imaging and image analysis. This interview is led by Dr. Bruno Sanguinetti, Dotphoton's CTO.
Martin traces a path that will be familiar to many in the field: a background in adaptive systems and quantum physics in the early 2000s, followed by a serendipitous move into biology when exactly the right person mentioned exactly the right opening at exactly the right moment. He talks about the difference between AI as a tool for understanding intelligence and AI as pure engineering optimisation, why deep learning has been genuinely transformative for bio-image analysis, and where multimodal imaging is headed next. As always, we close by asking our guest what they see as the defining theme for society in the years ahead.
A bit of both. Academia means short contracts, so there's always an element of luck in what's available. Martin was coming to the end of a postdoc in Leeds, keeping an eye on options, and sent a speculative message to an old MSc colleague — Katie Bentley, who he'd studied evolutionary and adaptive systems with at the University of Sussex around 2001. It just happened that she and her boss had recently visited Germany to see one of the new light-sheet microscopes, and they were looking for a physicist who could build things and write custom image analysis code. Biology outside of standard bioinformatics pipelines is a rare skill set. That message led to a role at Cancer Research UK, which eventually led to the Francis Crick Institute - and a whole career direction.
Directly, in multiple ways. The hardware side came first: knowing how to get light into tight spaces, understanding diffraction limits, appreciating the quantum efficiency of different cameras — these are things experimental physicists deal with constantly, but that aren't always obvious to biologists. Martin's entry into the electron microscopy team at CRUK came because he attended a talk on Correlative Light and Electron Microscopy, immediately recognised that the hardware challenge described — piping light into a vacuum chamber through millimetre-scale gaps — was structurally similar to work from his PhD, and emailed the organiser on the spot. She replied: come downstairs and let's talk.
More broadly, physics training means being comfortable with multi-dimensional problems, with theory and experiment side by side, and with not being pigeonholed. That last point has become especially relevant as AI and image analysis have converged.
It's worth separating two things. One is the original AI ambition — understanding how intelligence works, what a brain actually does. That remains an open and very hard question. The other is the engineering use: given data and a target, can we optimise our way to a good answer? That second use is genuinely transforming bio-image analysis right now.
In particular, convolutional neural networks and deep learning methods have made a step change in what's possible. Martin's background, shown literally in the visualisation behind him during the interview, is a prime example: a huge EM dataset — probably hundreds of gigabytes, imaged over days — with thousands of mitochondria in a piece of brain tissue, each one individually segmented and colour-coded by a deep learning model. Attempting this with traditional methods doesn't really work. With deep learning it does, and increasingly it even becomes routine. One of the most powerful properties is composability: someone builds a model for one dataset, the next researcher adapts it with a small amount of additional annotation rather than starting from scratch. The hand-tuning that used to be a dark art is replaced by optimisation.
There's a slight of hand worth acknowledging. For supervised methods — which is most of what's in production — you still need significant quantities of manually annotated ground truth data to train the network. That annotation bottleneck is real. Martin's team has been exploring citizen science as one route around it: inspired by the Galaxy Zoo project, which crowdsourced the annotation of millions of astronomical images, they set up their own citizen science projects on the Zooniverse platform to gather training data at scale. It's worked well.
The deeper challenge is interpretability and robustness. A model trained to predict certain things can predict them inappropriately very easily — especially when the inference-time data doesn't match the statistical distribution of the training data. Understanding that input space, knowing when a model is operating outside its competence, is still a hard problem. Martin notes a parallel with quantum physics: in both cases you're working with high-dimensional spaces where the meaningful dimensions aren't always orthogonal and aren't always obvious, and physics training helps build intuition for that even when the intuition can't be fully articulated.
Certain hardware limits are being approached. Cameras are close to 100% quantum efficiency. Certain optical limits have been pushed by super-resolution techniques. So where does progress come from? Martin's view is that the next wave is multimodal imaging — combining different imaging modalities in ways that are more than the sum of their parts.
CLEM is the example he knows best. Fluorescence microscopy gives you functional specificity (you label exactly what you care about, and it glows) but limited resolution. Electron microscopy gives you nanometre-scale resolution of almost everything, but no inherent specificity — everything shows up, and if you're not an expert, it's hard to know what's what. Combine them and you get: I know this structure is functionally relevant because the fluorescence told me so, and I know exactly what it looks like at the ultrastructure level because the EM told me so. You can also use the fluorescence to guide where the EM acquires — saving enormous amounts of imaging time on a modality where a single cell at full resolution can be a terabyte and take days.
Add X-ray microscopy (micro-CT, centimetre-scale at micron resolution), histopathology, and other modalities, and the combinations multiply. AI is particularly well suited to extracting patterns across these heterogeneous datasets — things a human can't see because there's too much data, or because the modalities are too different to hold in mind simultaneously. Generative approaches — inferring what one modality would show from another — are an active area, though rigorous validation is essential to make sure the predictions are real rather than learned artefacts.
Integration or rather, open integration across silos.
In science and beyond, development tends to happen behind closed doors. Different fields, different institutions, different companies are often solving the same problems in parallel without knowing it. The analogy from Martin's own work is clear: CLEM gets you something more than the sum of its parts precisely because it crosses the boundary between two previously separate imaging traditions. The same principle applies at the level of knowledge, of file formats, of research communities. If the goal is to make things better for more people, then keeping things hidden — for profit margins, for competitive advantage, for institutional inertia — works against that goal. The next few years could bring a lot of progress if those silos open up.
Dotphoton provides innovative image compression solutions for big image data. Dotphoton’s unique set of algorithms and cutting edge approach to camera calibration enables file size reduction by a factor of 6—10, while preserving the raw quality of images. Dotphoton's deep understanding of latest insights from the quantum information field ensures it stays ahead as the highly reliable partner, trusted by the European Space Agency, Bosch, and the leading biomedical centres across the world.
We use cookies to enhance your browsing experience and analyse site traffic. By clicking "Accept all" you consent to our use of cookies. Manage preferences