This last Saturday, 27 June, we hosted our last Open Day for the 2025-26 academic year. Throughout the whole year, visitors to most Open Days could see a pint of the work we do in our Optical Imaging lab. On a table, several devices containing light sources, optical fibres, and squiggly laptop traces were laid down and highlighted what looked light an engineering or physics demo. The question we answered a few times was, “what is the computer science problem?”
Our lab, based at the University of Birmingham, works on diffuse optics: shining near-infrared light through tissue and reading what comes back out to measure blood flow and oxygenation in the brain and muscle. We use several flavours of this, from continuous-wave fNIRS, the simplest and cheapest version, to time-domain and frequency-domain systems that extract more information by timing or modulating the light, to diffuse correlation spectroscopy (DCS), which tracks blood flow directly by watching how light speckles fluctuate as red blood cells move.
None of it involves ionising radiation, none of it requires the patient to lie still in a scanner, and all of it produces enormous amounts of data that someone has to make sense of. That’s where computer science comes in, at every stage.
Instrument design and automation
The light source, the detector, the fibres, and the electronics that stitch them together are increasingly run by embedded software, not knobs and switches. Real-time control loops, FPGA programming, sensor fusion, and robotics, think automated probe placement or wearable systems that adjust themselves as a person moves, all need researchers who think computationally. A better instrument is often a better piece of code, not just better hardware.The inverse problem
This is the part that should genuinely excite anyone who likes hard optimisation problems. Light doesn’t travel in a straight line through tissue, it scatters, bounces, and diffuses. Going from “light in, light out” to a map of oxygenation inside the head means solving an inverse problem that is typically ill-posed, often underdetermined, and always sensitive to noise and modelling assumptions. This is the same mathematical territory as image reconstruction in CT and MRI, and it rewards people who know numerical methods, regularisation, and Bayesian inference.Machine learning and data-driven modelling
Diffuse optics data is messy: motion artefacts, systemic physiology contaminating the brain signal, huge inter-subject variability. Classical signal processing only gets you so far. There’s real appetite for machine learning approaches that can denoise signals, separate systemic from neural components, or learn tissue models directly from data rather than from hand-derived physics.Medical AI and clinical translation
Ultimately this technology needs to answer real clinical questions: is this preterm infant’s brain getting enough oxygen, is this stroke patient recovering, is this patient in the ICU deteriorating. That means building models that are accurate, interpretable, and robust to the hospital and population they’re deployed in, and validating them against other neuroimaging methods like MRI or PET. This is applied AI with a direct line to patient outcomes.If you’re a computer scientist, or about to become one, and you assumed biomedical optics was someone else’s field, it isn’t. If any of this sounds like your kind of problem, come find us.“We’re short on people who can write good embedded code, solve inverse problems properly, or bring modern ML to a genuinely hard, physically-grounded dataset.”
— Rickson Mesquita, Principal Investigator