About
I’m Gerard Martí Juan. I work on machine learning and AI for medical images, and I have been kicking at the problem for a while now — first in Alzheimer’s disease, then multiple sclerosis, and fetuses, (which are the least cooperative imaging subjects I have ever worked with). I work as a Senior AI Researcher at Eden, building healthcare AI for the real world.
What I actually find interesting
Most of the interesting problems in medical imaging are outside the walls of academia. A model that reaches a good number on a curated benchmark and a model that survives contact with an owerworked radiology that wants systems that help their work are rarely the same model, and the gap between them is where the actual work is: data harmonisation, quality control, and actually understanding the final user’s needs, and being able to translate that into the model’s capabilities. This is a very hard problem, but it is the one that gives better returns on performance and impact than any other.
Background
I did my Ph.D. in Information and Communication Technologies at Universitat Pompeu Fabra, finishing in 2021 Cum Laude, on data-driven methods for characterising heterogeneity in Alzheimer’s disease — essentially, on the fact that “Alzheimer’s disease” describes a lot of rather different-looking brains. Before that, an MSc in Computer Vision at Universitat Autònoma de Barcelona and a BSc in Informatics Engineering at Universitat Politècnica de Catalunya.
Personal
Outside my work, I like to spend time with my wife, read and play interesting books and games, and follow the (mis)fortunes of FC Barcelona, going to matches whenever I can.
Elsewhere
The best way to reach me is email: gerardmartijuan(at)gmail.com.
Colophon
This site is a Jekyll build on GitHub Pages. It started life as Jekyll Now years ago; by now every layout, include and stylesheet has been rewritten, but the original attribution stays in the page source where it belongs.
A few deliberate choices:
- No trackers, no analytics, no cookies. I have no idea who visits, and I like it that way. Two scripts do still come from a CDN: the search index loads from unpkg the first time you focus the search box, and the pages with charts or equations pull those libraries from jsDelivr. Everything else is served from here.
- No JavaScript frameworks. Dark mode is a set of CSS custom properties plus a handful of lines of vanilla JS.
- Type is Newsreader, an open-licence variable serif, self-hosted rather than pulled from Google Fonts.
- Icons are inline SVG. The CV page used to load Font Awesome, Google Fonts and a 284 KB icon webfont to draw five contact links. Now it loads nothing.
- Publications come from one YAML file that renders both the list and the timeline chart above it, so the two cannot drift apart.
- CI does the nagging. Every push runs a small Python validator that checks internal links, missing image alt text, duplicate element IDs and malformed chart data, so the site fails loudly rather than quietly.
- AI-native infrastructure. Code is supported by my good pals Claude and GPT Sol, who helps debug and write the site’s code. However, writing is all me.
The whole thing is on GitHub if you want to take anything from it.