Hematology foundation model family of DINOv2 vision transformers for white blood cell embeddings from blood and bone marrow smears.
No providers recorded yet. Browse all providers
A differential blood count still depends on a trained cytologist examining at least 200 white blood cells per patient, and the automated classifiers built to relieve that work have a familiar weakness: a network trained on one hospital's smears degrades on another's. Stain chemistry, microscope optics and camera settings shift the images more than the morphological differences between a myelocyte and a metamyelocyte, and most public hematology datasets are too small, and too differently labeled, to train through that variation.
DinoBloom is a family of self-supervised vision transformers built to supply a reusable single-cell feature extractor for hematology. It was developed by Valentin Koch, Sophia J. Wagner and colleagues in Carsten Marr's group at Helmholtz Munich, with the Technical University of Munich, the Dr. von Haunersches Kinderspital of LMU Munich, and King's College London, and was published at MICCAI 2024. Rather than labeling more cells, the authors pooled 13 public and research datasets of peripheral blood and bone marrow smears, more than 380,000 white blood cell images in all, and let a DINOv2 teacher-student objective learn what stays constant across them.
The released encoders are used frozen: downstream users fit a linear probe, a k-nearest neighbor classifier, or a multiple instance learning head on top of the embeddings. Later Marr Lab models such as CytoSAE and GenBloom continue this line of work.
Training used the DINOv2 framework with an image-level DINO loss on the class token and a patch-level iBOT masked-token loss, images resized to 224×224, on eight NVIDIA A100 40 GB GPUs. Runs were short, 45 minutes for DinoBloom-B and four hours for DinoBloom-G, and performance peaked between 4,000 and 8,000 iterations. The largest source is the Bone Marrow Cytomorphology (BMC) set of 171,373 expert-annotated cells from 945 patients; others include the AML Hehr and AML Matek leukemia cohorts, Raabin-WBC, NuClick, LISC, Chula, BCCD and a private LabAnonymous set of over 40,000 peripheral blood images. BMC and AML Hehr were split 80/20, with only the training portion used for pretraining.
On Acevedo (11 classes), DinoBloom-G reached a weighted F1 of 89.1 with 1-nearest neighbor classification against 77.8 for DINOv2 ViT-G and 83.3 for Phikon; in linear probing it reached 91.8 against 88.0 for CTransPath. For AML subtyping on AML Hehr with attention-based MIL across four genetic subtypes plus healthy controls, DinoBloom-G reached a weighted F1 of 93.1, while the best non-DinoBloom encoders reached 81.9 (ResNet-50) and 81.8 (Phikon). On the 21-class, heavily imbalanced BMC bone marrow task, DinoBloom-B reached 65.8 balanced accuracy in 1-NN evaluation against 42.0 for the next best baseline. Larger variants helped on peripheral blood but not consistently on bone marrow, where DinoBloom-S gave the best linear-probe result.
DinoBloom serves as a drop-in feature extractor for computational hematology: white blood cell classification in peripheral blood and bone marrow, patient-level prediction of genetically defined AML subtypes through multiple instance learning, and checking a new dataset for batch effects before training on it. The authors also show UMAP embeddings of a patient's cells as a visual summary of their cell-type distribution, where myeloblast clusters that are nearly empty in healthy controls stand out in AML patients, suggesting morphology-based gating analogous to flow cytometry. Because linear probes work well on its features, a laboratory with modest labeled data can build classifiers without GPU-heavy fine-tuning.
DinoBloom established a shared, openly downloadable backbone for single-cell hematology imaging, and follow-up work uses it as a frozen encoder: CytoSAE learns its concept dictionary on DinoBloom-B features, and RedDino, a red blood cell counterpart, benchmarks against it. It also showed that continued self-supervised training on pooled, heterogeneous public cohorts outperforms both natural-image and histopathology encoders on blood cells. Limitations are clear from the design: the corpus is dominated by white blood cells, one of the 13 training sources is not public, and all evaluations are retrospective in-silico benchmarks rather than prospective clinical validation.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.