DinoFlow: Is Flow Cytometry Getting a Foundation Model?

A foundation-model approach could shift flow analysis from task-specific models toward reusable representations learned across large collections of cytometry data.

By Juan Manuel Ojeda · CFCM

PRIMARY EVIDENCE · PEER-REVIEWED

Primary evidence

O'Fallon B, Morrison M, Grespan MM, Spies NC, Ng DP. DinoFlow: Self-supervised pretraining in flow cytometry enables accurate detection of common hematopathological disorders. Cytometry Part B. First published 15 June 2026. DOI: 10.1002/cyto.b.70043.

What it actually demonstrates

The dataset contained 52,625 routine samples from ARUP Laboratories (2019–2024), using three 10-marker tubes on five NaviosEX instruments. The latest 4,993 samples were held out for temporal testing and excluded from pretraining. This is a dataset total, not the number used for training.

Evidence boundary

Evidence boundary: this supports self-supervised pretraining and reusable tube-level representations in the reported dataset; it does not establish universal transfer across laboratories, instruments or panel designs.

Primary paper →

Most automated cytometry systems are built for a particular panel, disease or classification task. Foundation models propose a different architecture: learn a general representation first, then adapt it to downstream tasks with less task-specific training.

For clinical cytometry, the attraction is obvious. A reusable representation could support classification, quality control, anomaly detection or cohort exploration. The difficult part is demonstrating that representations remain meaningful across instruments, panels, laboratories and pre-analytical variation.

The clinical question is therefore not simply whether a model transfers. It is where transfer fails, how that failure is detected and what validation is required for each intended use.

CFCM view

Clinical value depends on reproducibility, transparent limitations and evidence in the intended workflow. Technical novelty is the beginning of validation, not its endpoint.

Editorial note

CFCM is an independent publication. References to manufacturers, instruments, reagents, software or therapies do not constitute endorsement. Technical content should be interpreted in the context of local validation, applicable regulation and manufacturer instructions for use.

Author and related reading

Juan Manuel Ojeda — Founder & Editor, CFCM

Professional background: Juan Manuel Ojeda has professional experience with Sysmex España in clinical flow cytometry and now works as a freelance consultant. CFCM is his independent editorial project; its views do not represent Sysmex or imply company endorsement. Employer or client confidential information and intellectual property are excluded.

Original CFCM report: the August date shown above. Expanded web version and source-check pass: 19 September 2026. This was not independent scientific review. Source publication dates are separate; this is not retrospective web publication.

Resources: validation and analysis →Clinical Notes: practical lessons →Related: Clinical Flow AI Under the EU AI Act

References

O'Fallon B, Morrison M, Grespan MM, Spies NC, Ng DP. DinoFlow: Self-supervised pretraining in flow cytometry enables accurate detection of common hematopathological disorders. Cytometry Part B. First published 15 June 2026. DOI: 10.1002/cyto.b.70043.

Primary paper →