DinoFlow

A single tube from a leukemia/lymphoma workup holds roughly a quarter of a million cells, each measured on thirteen channels, and the diagnostic question — is this acute myeloid leukemia, is this T-cell acute lymphoblastic leukemia — is a property of the population, not of any one cell. Most machine learning built on clinical flow cytometry answers one such question at a time, fitting a supervised classifier per disorder — which works where positive cases are plentiful and breaks down for the rare entities a laboratory still has to catch.

DinoFlow, from the ARUP Institute for Research and Innovation at ARUP Laboratories with collaborators at the University of Utah, inverts the order. A transformer encoder is pretrained once, without labels, on 52,625 routine clinical samples, and compresses each tube into one fixed-size vector that lightweight heads then read for specimen viability or any of five common hematological disorders. The distinguishing choice is that the representation is tube-level from the start: earlier self-supervised cytometry work embedded individual events and had to pool them afterwards, but a diagnosis is rendered on the sample, not the cell.

Published in Cytometry Part B: Clinical Cytometry in June 2026, it is the earlier of the lab's two cytometry representation models: the successor EventHorizon generalizes across 17 antibody panels, while DinoFlow is trained on one fixed three-tube panel.

#Key Features

• Tube-level representation via a CLS token: Each event is a token in an unordered set, and the output at a prepended learnable class token is the sample representation — no pooling heuristic, and directly the unit a diagnosis is made on.
• DINO self-distillation adapted to event sets: The teacher network sees 4,096 events and the student 512, standing in for the original method's global and local image crops; augmentations are per-marker scaling, shifting, and Gaussian noise.
• A density loss to prevent collapse: Without a kernel-density ancillary loss, ramped from 0.005 to 0.01 over the first ten epochs, training collapsed to identical outputs for every input.
• Strongest where labels are scarcest: On T-ALL, with 101 training cases, it reaches an F1 of 0.926 against 0.858 for the next best model, while on common AML and CLL it sits level with supervised set transformers.
• Robust to instrument drift: Under in-silico shift, scale, and spillover perturbations at five times the drift measured across five cytometers over five years, F1 and auPRC stayed above 0.90 of the unperturbed baseline.

#Technical Details

Each event carries 10 antibody markers plus two forward- and one side-scatter measurement, and a single linear layer projects those 13 values into the model dimension; no positional encoding is applied. A standard transformer encoder stack with SwiGLU feedforward layers processes the sequence, and pretraining attaches a three-layer MLP projection head of dimension 4,096 for the DINO cross-entropy loss. An ablation over 0.8M, 6.3M, and 42M parameter variants settled on the 6.3M model, since the largest overfit without gaining accuracy. Training data are 52,625 samples collected at ARUP from January 2019 to August 2024, each three tubes (B-cell, T-cell, myeloid) run on Navios EX cytometers, compensated and arcsinh-transformed; the 4,993 most recent form a temporal test split. Fine-tuned, DinoFlow scored F1 0.878 on AML and 0.945 on CLL — narrowly behind set transformers at 0.884 and 0.952 — and led on B-NHL, B-ALL, and T-ALL, with viability regression at RMSE 4.51. Accuracy climbs with the number of events supplied at inference, most sharply for CLL cases below 1% disease burden, where F1 rises from about 0.75 at 4,096 events to over 0.80 at 64,000.

#Applications

The target is high-volume hematopathology triage, where one screening panel must surface many disorders. Frozen representations serve quality control as well as diagnosis: a UMAP of the concatenated three-tube embeddings separates bone marrow from peripheral blood and isolates hemodilute, clotted, and low-viability specimens, while a two-layer perceptron on those same features already beats self-organizing map representations on every diagnosis tested. Adding a target disorder becomes a matter of training a small head rather than a backbone — the practical argument for the approach where add-on tubes are ordered for rarer entities such as plasma cell neoplasms.

#Impact

DinoFlow demonstrates that DINO-style self-distillation carries over from vision to clinical flow cytometry at production scale, on a corpus an order of magnitude larger than the sub-1,000-sample cohorts typical of the field, and it is the prior work its laboratory's panel-agnostic successor is benchmarked against. The rare-disease result is the consequential one: unlabeled pretraining bought accuracy precisely where supervised training cannot. The implementation is public on GitHub but carries no license file, the pretrained weights are not released, and the training corpus is identified clinical data from one reference laboratory — so the reported accuracies are an internal retrospective validation, not an externally reproducible benchmark, and no prospective deployment is reported.