In our lab, we don't just push pixels or fine-tune search algorithms. We explore more delicate frontiers. The RADIOHED project was born from a simple yet dizzying ambition: can we, with our resources and our agility, build an intelligence capable of reading a chest X-ray and spotting pneumonia?
The challenge: medicine is the domain of nuance
Behind this question lies a real challenge. Medicine is the domain of extreme nuance. A blur on an image can be an artefact, a breath, or the sign of a serious condition. For a machine, learning this seems to demand an eternity of computation and millions of data points.
And yet, we got there. The key to it holds in one concept we want to share with you, because it defines the way we work: Transfer Learning.
The learning paradox
Imagine that to teach a child to recognise an apple tree, you first had to re-teach them the entire physics of light, the biology of plant cells and the geometry of fractals. They would never eat an apple.
Yet that's what many AIs do: they start from scratch, from the "dead pixel", to try to understand the world.
For RADIOHED, we chose a more human path, closer to the way we, as adults, learn. We leaned on a "giant" named DenseNet121. This neural network is no doctor. It spent its life looking at millions of everyday images: cats, cars, landscapes.
From that long training, it drew a universal skill: it knows how to see. It knows the grammar of shapes, the weight of shadows, the texture of surfaces.
Grafting specialised knowledge
The teaching behind Transfer Learning holds in this idea of reuse. Rather than forcing the machine to reinvent optics, we "froze" its generalist knowledge. We kept that expert photographer's eye it had forged for itself.
Then we performed a graft. We added a small final layer — we call it the "head" — and trained it specifically on 5,200 X-rays.
This is where the subtlety happens: the machine didn't learn to see X-rays; it learned to apply its universal gaze to a new domain. Because it already knew what a texture was, it quickly grasped the difference between healthy lung tissue and the opacity of an infection.
Why this efficiency moves us
The result surprised even us with its accuracy. With a score of 0.96 out of 1, our model holds its own against systems trained on colossal infrastructure. And yet it was built in a few minutes, on a simple laptop.
This is the deeper message of RADIOHED and of our lab: intelligence isn't a matter of brute force. It's a matter of connection.
Transfer Learning is a technological equaliser. It lets agile players lean on the collective heritage of AI to build cutting-edge solutions, quickly and with humility. We didn't build the reservoir of knowledge; we learned to plug into it with precision.
"We don't relearn how to see the world. We simply offer a speciality to a gaze already shaped by the experience of giants."
Key takeaways
Transfer Learning reuses the knowledge of an already-trained network rather than starting from the "dead pixel".
DenseNet121, trained on millions of everyday images, already knows how to see: shapes, shadows, textures.
We freeze that generalist knowledge and graft on a "head" trained on 5,200 chest X-rays.
The result: 0.96 out of 1, achieved in a few minutes on a simple laptop.
Intelligence isn't a matter of brute force, but of connection to the collective heritage of AI.