Deciphex has taken the wraps off CipherX, the pathology AI engine that has been running quietly inside its clinical and pharmaceutical businesses since January. Foundation models turn a tissue image into numbers, the Irish company argues, yet they cannot say which structure on the slide produced a prediction.
Two layers make up CipherX. Sitting underneath is a foundation-model layer that pulls together a set of encoders (including DCX-3, Deciphex’s own fusion model) and chooses which one handles a given task according to the usage rights attached to each dataset.
Above that sits a semantic layer. Its job is to take those representations apart into what Deciphex calls glyphs: units of tissue structure that recur, each one confirmed and named by a subspecialist pathologist. Glyphs then combine into signatures, the named histological structures a pathologist recognizes, so nothing in an output is opaque and every finding can be traced back to the slide.
The company argues this matters commercially as much as scientifically. Swap out the foundation model underneath and everything stacked on top of it is wasted work, whereas a layer that interprets can outlive the swap.
Regulators want the same thing: auditable infrastructure that does not hinge on one model version. Deciphex says tile-level performance has converged across pathology foundation models, so the next gains will come from data and interpretation rather than bigger training sets.
CipherX runs inside Diagnexia, the clinical diagnostic service, and Patholytix, the pharma research arm.
Why it matters. Pathology AI is moving from model announcements to harder deployment questions about audits, traceability and recertification.
