RedDino

Red blood cell morphology is a routine but demanding part of hematology. The shape of an erythrocyte — sickled, spherocytic, target-shaped, elliptical, fragmented — carries diagnostic weight for anemias, hemoglobinopathies, and membrane disorders, and reading those shapes at scale is exactly the kind of task computer vision should be good at. In practice, models trained on one lab's smears tend to collapse on another's, because staining protocol, microscope, and camera introduce batch effects far larger than the morphological differences of interest.

RedDino is a self-supervised foundation model built specifically for this problem by Luca Zedda, Andrea Loddo, and Cecilia Di Ruberto at the University of Cagliari with Carsten Marr at Helmholtz Munich. It adapts the DINOv2 self-supervised framework to red blood cell imagery and is pretrained on a curated corpus of 1.25 million single-cell crops drawn from many acquisition modalities and sources. The result is a frozen feature extractor: downstream classifiers are fitted on its embeddings rather than by fine-tuning the backbone.

The design deliberately targets erythrocytes rather than leukocytes, which distinguishes it from broader digital hematology systems such as Uni-Hema that span detection, segmentation, and language tasks across many cell types. RedDino does one thing — produce robust RBC representations — and the paper's contribution is as much the ablation of which DINOv2 design choices actually help for cells as the model itself.

#Key Features

• Erythrocyte-specific DINOv2 adaptation: The KoLeo regularizer is removed because it suppresses exactly the rare pathological morphologies that matter clinically; Sinkhorn-Knopp centering replaces moving-average centering; and DINOv2's natural-image augmentations are swapped for a cell-appropriate pipeline.
• Three released model sizes: Small (22M parameters, 384-dimensional features), base (86M, 768-d), and large (304M, 1024-d) ViT backbones with patch size 14, all downloadable and loadable through timm.
• Frozen-feature evaluation protocol: Performance is measured by linear probing and nearest-neighbour classification on frozen embeddings, so the reported gains reflect representation quality rather than downstream fine-tuning capacity.
• Cross-source generalization: Leave-one-source-out evaluation tests transfer to smears acquired under protocols and equipment absent from training, the failure mode that defeats conventionally trained RBC classifiers.
• Open release with documented weights: Code is published under Apache-2.0 and the checkpoints carry a model card describing architecture, intended use, training data, and benchmark results. The small and base checkpoints are licensed CC-BY-4.0; the large checkpoint is CC-BY-NC-4.0.

#Technical Details

The pretraining corpus was assembled from 18 public datasets covering more than 420 individuals. Two extraction routes were used: a fine-tuned CellPose model that yielded over 3 million segmented cells, and non-overlapping 224x224 patch extraction that produced the 1,250,781 images used for training. White blood cell images were retained to counter class imbalance. On the Elsafty benchmark under leave-one-source-out evaluation, RedDino-large reaches 88.5% weighted F1 with linear probing against 86.0% for a DINOv2-large baseline, and RedDino-base reaches 78.8% weighted F1 under 1-nearest-neighbour classification against 75.5% for DINOv2-base. RedDino-small reaches 84.3% weighted F1 on Chula against 81.5% for DINOv2-large, and RedDino-large reaches 86.6% on DSE. Reported baselines include ResNet-50, DinoBloom, and unmodified DINOv2. The authors note that classes with overlapping morphology remain hard to separate, reflecting the absence of sharp clinical thresholds between shape categories.

#Applications

RedDino is intended as a drop-in backbone for computational hematology: RBC shape classification, morphology-based screening of peripheral blood smears, batch-effect-robust feature extraction for cross-site studies, and retrieval or clustering over cell populations. Because strong results are obtained with linear probes on frozen features, a lab with a few hundred labeled cells from its own scanner can build a usable classifier without GPU-intensive fine-tuning. The three model sizes let deployments trade accuracy against compute, with the small variant practical for high-throughput or edge settings.

#Impact

The work is a concrete demonstration that self-supervised recipes tuned on natural images need domain-specific surgery to work on cells — the KoLeo finding in particular is a transferable insight for anyone adapting DINOv2 to microscopy, where the biologically interesting samples are the rare ones a uniformity regularizer will flatten. Releasing three checkpoint sizes with a documented model card lowers the barrier for hematology groups that lack pretraining budgets. Two limitations bound its use: the paper is a preprint under a no-derivatives license and has not yet completed peer review, and the large checkpoint's non-commercial license restricts incorporation into commercial diagnostic products, though the small and base checkpoints carry no such restriction.