Cancer explained to artificial intelligence: Italy's model thinking oncologist

The project with Ieo was born within a precise idea: the big generalist models are not enough in medicine. Not because they are poor, but because they do not know where they stand. They don't know the protocols of that department, they haven't read those reports, they don't understand how that team thinks.

Reply's answer is called the Model Factory: an industrial process - from data qualification to training, from governance to deployment - designed to build models that really know the context in which they will operate. In healthcare, where an algorithmic hallucination can have real consequences on a real patient, the difference between a generic model and one trained on institution-specific knowledge is not a technical detail. It is the point.

"The real value will come from models built on the knowledge, data and expertise of each organisation," said Carlo Malgieri, Partner at Laife Reply, "The collaboration with Ieo was created precisely to systemise these elements.

A few days ago, Microsoft and Mayo Clinic announced a collaboration that, in its conceptual architecture, follows the same path. The model the two giants are developing will be trained on de-identified clinical data and longitudinal knowledge accumulated by one of the world's most respected medical institutions. It will be owned by Mayo Clinic - a non-trivial detail, signalling a clear stance on the issue of clinical data governance - and will be made accessible through Azure Foundry's Api.

Mustafa Suleyman, CEO of Microsoft AI, called Mayo Clinic 'the best collaboration imaginable' to accelerate towards what he called 'medical frontier intelligence'. The difference with generic models, as Mayo Clinic itself pointed out, is that health AI requires deep clinical context, longitudinal understanding, rigorous governance, and real-world validation.

This is exactly the same premise from which Reply starts with Ieo. The difference is one of scale and geographical ambition: Mayo Clinic and Microsoft aim to build a global model, accessible to anyone in the world via the cloud. Reply and Ieo are aiming at something more granular: a model that knows Ieo, that thinks like Ieo, that has been trained on the specific oncological complexity of one of Europe's most important Irccs.

The two approaches are not mutually exclusive. On the contrary, they will probably complement each other: large foundational models can become the basis on which to graft vertical specialisations, such as the one Reply is building with Ieo. The Reply Model Factory is designed exactly for this type of layered architecture.

What is emerging these days - between Milan and Rochester, Minnesota - is actually the confirmation of a paradigm shift that has been in the air for some time. Ia in healthcare is no longer about 'if', but about 'how' and 'who controls the data'. The question of model ownership, explicitly raised in the Mayo-Microsoft agreement, and that of control over the training process, central to the Reply Model Factory approach, are two sides of the same problem: how do we bring artificial intelligence into a high-risk context - medicine - without losing clinical, ethical and legal control?

The Ieo, with its patient-centred care model and data-driven strategy, is in this sense a natural partner for such a project. It is not a mere client, but a co-developer that brings with it a wealth of information built up over years of clinical and research activity. The Ia will not cure cancer on its own. But training a model that can read a breast report as an oncologist at the Ieo reads it is already a concrete step forward. And it is exactly the kind of step that, multiplied on a scale, can change medicine.