Proprietary histopathology foundation model encoding tissue tiles into embeddings for the CipherX pathology AI engine in clinical and pharma work.
No providers recorded yet. Browse all providers
DCX-3 is a pathology foundation model developed by Deciphex, the Dublin-based computational pathology company behind the Diagnexia clinical diagnostic service and the Patholytix pharmaceutical research business. Deciphex describes DCX-3 as a "fusion model": an encoder that converts histopathology image tiles into numerical representations of their morphology, which every downstream tool then works from. It was disclosed publicly on 22 September 2026 as part of the launch of CipherX, the company's pathology AI engine, which has run in Deciphex's production workflows since January 2026.
DCX-3 and CipherX are different things, and the distinction matters for anyone reading the announcement. CipherX is an engine with two layers. Its foundation-model layer holds several encoders, DCX-3 among them alongside third-party models, and picks one per task and according to the usage rights attached to each dataset. Above that sits a proprietary semantic layer that breaks an encoder's embeddings into "glyphs", recurring units of representation that a subspecialist pathologist has confirmed correspond to a structural tissue element, and composes glyphs into named histological "signatures". DCX-3 is the Deciphex-built encoder in the first layer; the semantic layer and the other encoders are not part of it.
The motivation Deciphex gives for this split is a limitation of raw pathology encoders: a prediction made from an embedding cannot easily be traced back to anything visible on the slide, and swapping the encoder forces every tool built on it to be redone. By anchoring outputs to pathologist-named structures, CipherX aims to keep the downstream vocabulary stable even as the encoder beneath it, DCX-3 or another, is replaced.
Deciphex has not published a technical report, preprint or model card for DCX-3, and its architecture, parameter count, pretraining objective and training corpus are undisclosed. The company has separately described its foundation-model program as drawing on a databank of about 3.5 million slides assembled through data-sharing agreements with Novartis and Charles River Laboratories, digitized clinical archives and consented patient contributions, without attributing a specific corpus to DCX-3.
The only performance claim made for the encoder itself is that it places among the leading pathology foundation models on the THUNDER and EVA suites, alongside Virchow2, UNI2-h, H-optimus-1, RudolfV 2 and Midnight. Deciphex has not released per-task scores, and DCX-3 has not been added to the public THUNDER leaderboard maintained by the benchmark's authors, so the comparison rests on the company's own report.
The headline figures in the launch, 99.85% negative predictive value for adenocarcinoma and 98.76% for melanoma in routine Diagnexia casework, belong to downstream CipherX detection tools built on its signatures. They measure those deployed tools, not the DCX-3 encoder.
DCX-3 reaches users only through CipherX. In Diagnexia, CipherX runs three Assist tools outside the diagnostic pathway, covering image quality assessment, complexity-based case triage and post-authorization review, plus adenocarcinoma and melanoma detection tools, with every case signed out by a pathologist. Through Patholytix Research Services, the same engine supports translational research and clinical-trial histopathology, digital tissue biomarker development and digital companion diagnostics for pharmaceutical sponsors.
DCX-3 represents a commercial pathology lab building its own tile encoder while explicitly treating encoders as interchangeable parts beneath a stable, pathologist-auditable semantic layer. That framing reflects a view that tile-level performance across leading pathology foundation models has converged and that interpretability and model-invariance now matter more for regulated clinical and pharma use. Its reach outside Deciphex is limited: CipherX is not offered as a licensed product, no weights, code or API for DCX-3 have been released, and its benchmark standing has not been independently reproduced.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.