RedDino: Foundation Model for RBC Analysis

• RedDino is a self-supervised foundation model specifically designed for red blood cell image analysis using over one million curated RBC patches.
• It employs RBC-specific adaptations such as Sinkhorn–Knopp centering and contextual patch extraction to enhance feature quality and diagnostic relevance.
• The model demonstrates improved performance metrics in classification tasks by leveraging tailored augmentations and eliminating local-crop configurations.

Searching arXiv for papers relevant to “RedDino” and closely related name variants. RedDino denotes, in the current arXiv literature, two distinct referents. The canonical use is “RedDino: A foundation model for red blood cell analysis,” a self-supervised foundation model for red blood cell (RBC) image analysis that adapts the DINOv2 framework to hematology and is trained on a curated corpus of 1.25 million RBC image patches from 18 public datasets (Zedda et al., 11 Aug 2025). The same string also appears as a colloquial or misspelled reference to RAD-DINO, the unimodal chest X-ray encoder introduced in “Exploring scalable medical image encoders beyond text supervision” (Pérez-García et al., 2024). Both are DINO-derived biomedical encoders, but they address different imaging modalities, supervision regimes, and downstream tasks.

1. Nomenclature and referential scope

The term “RedDino” is not univocal. In one usage, it is the exact model name of a red blood cell foundation model released with code and pretrained checkpoints for RBC morphology analysis (Zedda et al., 11 Aug 2025). In another, it is a colloquial or misspelled reference to RAD-DINO, whose canonical name is RAD-DINO, sometimes stylized “raddino” in the manuscript (Pérez-García et al., 2024).

Term	Canonical referent	Domain
RedDino	Self-supervised RBC foundation model	RBC image analysis
RedDino (colloquial usage)	RAD-DINO	Chest X-ray encoding

This distinction matters because the two systems differ at every operational level. RedDino is purpose-built for RBC morphology, batch-effect robustness, and cross-source generalization in hematology. RAD-DINO is a scalable, general-purpose biomedical image encoder trained without language supervision on chest X-rays and evaluated on classification, semantic segmentation, report generation, and correlation with other medical records. A plausible implication is that the shared DINO lineage has encouraged informal name conflation, but the models are not interchangeable.

2. Problem setting, corpus construction, and data regime for RBC analysis

RedDino is motivated by the fact that RBC morphology is diagnostically informative yet technically difficult to model. RBCs exhibit a broad range of shape phenotypes, including normal discocytes, echinocytes, ovalocytes or elliptocytes, spherocytes, stomatocytes, sickle cells, clumps or agglomerates, and infection-induced changes such as malaria. In practice, stained blood smears are acquired under varying imaging conditions, and technical variation from staining protocols, illumination, slide preparation, and pressure-induced deformation introduces batch effects that can obscure biological signal (Zedda et al., 11 Aug 2025).

The data regime is unusually broad for this subfield. The training corpus comprises 18 datasets, 56,712 original smear images from more than 420 individuals, 3,076,269 segmented cells, and 1,250,781 non-overlapping 224×224 patch crops with preserved aspect ratio. The paper states that the abstract’s figure of approximately 1.25 million RBC images is consistent with the patch count, while the text also notes “over 3 million single RBC images,” reflecting the segmentation output used during curation. CellPose, fine-tuned and iteratively improved through manual corrections, was used to segment individual cells. In parallel, non-overlapping smear patches were extracted to preserve contextual content such as cells, background, membranes, and parasites.

The corpus spans “various imaging modalities, resolutions, and staining techniques,” with examples including DIAMOND, MIMM, SN-AM, IML, MP-IDB, DSE/RedTell, ALL-IDB, BCCD, MADHLOOM, AVI, and ThalassemiaPBS. The paper highlights brightfield imagery and stained smears, but it does not enumerate exact modality counts or per-dataset imaging settings. This omission is material because batch-effect robustness is a central claim, yet the precise modality composition is not fully specified.

For downstream evaluation, RedDino uses labeled out-of-distribution datasets rather than training directly on a single taxonomy. Elsafty contains 240,000 images across nine RBC shape classes from four sources; Chula contains 20,875 images across 12 classes; and DSE/RedTell contains 5,659 images across eight classes. The full class lists are not enumerated in the paper. Examples explicitly mentioned include echinocytes, ovalocytes, “borderline ovalocytes,” clumps, and Plasmodium falciparum–infected RBCs.

3. Architecture, objective, and RBC-specific adaptations

RedDino follows the DINOv2 teacher–student joint-embedding and distillation paradigm, but it modifies several design choices for RBC imagery (Zedda et al., 11 Aug 2025). The model family consists of three ViT-based variants: RedDino-small with 22M parameters and a 384-dimensional feature, RedDino-base with 86M parameters and a 768-dimensional feature, and RedDino-large with 304M parameters and a 1024-dimensional feature. Training inputs are 224×224 patches.

The core distributions and loss are reported as

$$p_s = \operatorname{softmax}(z_s/\tau_s), \qquad p_t = \operatorname{softmax}((z_t - c)/\tau_t),$$

and

$$L = \sum_{v \in \mathcal{V}_t} \sum_{v' \in \mathcal{V}_s} H\big(p_t^{(v)},\, p_s^{(v')}\big),$$

with teacher parameters updated by EMA,

$$\theta_t \leftarrow m\, \theta_t + (1-m)\, \theta_s.$$

The baseline DINO moving-average center update,

$$c \leftarrow (1-\alpha)\, c + \alpha\, \overline{z_t},$$

is not retained in the final RedDino recipe. Instead, the moving-average centering used in DINO and DINOv2 is replaced with Sinkhorn–Knopp centering, which the paper reports as improving feature quality.

Several RBC-specific ablations are central to the method. First, local crops were found to degrade performance in RBC analysis, and the final models eliminate the local-crop configuration that caused this decline. The paper’s interpretation is that small crops can miss diagnostic context such as neighboring cells, smear artifacts, or parasites. Second, patched smear images perform substantially better than isolated single-cell crops, indicating that contextual smear information is useful even when the target phenomenon is cell morphology. Third, the KoLeo uniformity regularizer used in DINOv2 is removed. The stated reason is that RBC populations contain many naturally similar cells, so enforcing strong uniformity can suppress informative rare or pathological phenotypes in the embedding space. Fourth, DINOv2’s default pixel-level augmentations are replaced with an RBC-tailored Albumentations pipeline described as containing 32 pixel-level augmentations, although the exact operations and hyperparameters are not enumerated.

Training uses the “same hyperparameters as original DINOv2” except for batch sizes chosen to fit two NVIDIA A100 80GB GPUs. Reported batch sizes are 512 for RedDino-small, 384 for RedDino-base, and 256 for RedDino-large. Each model is trained for 2,000 iterations, after which performance reportedly decreases. Estimated emissions are 4.15 kg CO2eq. The paper does not explicitly report temperatures, EMA momentum, optimizer, learning-rate schedule, multi-crop counts, projection-head sizes, or feature-pooling details, so exact reproducibility depends on the released code.

4. Evaluation protocols and empirical performance

RedDino is evaluated by linear probing, nearest-neighbor classification, and feature-space visualization, with accuracy, balanced accuracy, and weighted F1 as primary metrics (Zedda et al., 11 Aug 2025). Elsafty uses cross-source leave-one-source-out evaluation: a shallow classifier is trained on one source and tested on the other three. Chula and DSE use five-fold cross-validation for linear probing, while nearest-neighbor results are not reported there because of rare-class sample insufficiency.

On Elsafty, RedDino-large attains a linear-probe weighted F1 of $88.5 \pm 5.5$, compared with $86.0 \pm 5.6$ for the best baseline, DINOv2 large; the reported improvement of the best RedDino over the best baseline is +2.5 weighted F1, +2.1 balanced accuracy, and +2.6 accuracy. Under 1-NN classification, RedDino-base reaches weighted F1 $78.8 \pm 3.6$, balanced accuracy $81.8 \pm 2.7$, and accuracy $78.6 \pm 3.7$, corresponding to improvements of +3.2 weighted F1, +5.0 balanced accuracy, and +3.1 accuracy over the best baselines. Under 20-NN classification, RedDino-base reaches weighted F1 $82.6 \pm 2.8$, balanced accuracy $$L = \sum_{v \in \mathcal{V}_t} \sum_{v' \in \mathcal{V}_s} H\big(p_t^{(v)},\, p_s^{(v')}\big),$$0, and accuracy $$L = \sum_{v \in \mathcal{V}_t} \sum_{v' \in \mathcal{V}_s} H\big(p_t^{(v)},\, p_s^{(v')}\big),$$1, with reported gains of +3.4 weighted F1, +5.2 balanced accuracy, and +3.0 accuracy.

On Chula, the gains are similarly consistent. RedDino-small achieves weighted F1 $$L = \sum_{v \in \mathcal{V}_t} \sum_{v' \in \mathcal{V}_s} H\big(p_t^{(v)},\, p_s^{(v')}\big),$$2, balanced accuracy $$L = \sum_{v \in \mathcal{V}_t} \sum_{v' \in \mathcal{V}_s} H\big(p_t^{(v)},\, p_s^{(v')}\big),$$3, and accuracy $$L = \sum_{v \in \mathcal{V}_t} \sum_{v' \in \mathcal{V}_s} H\big(p_t^{(v)},\, p_s^{(v')}\big),$$4; RedDino-base achieves weighted F1 $$L = \sum_{v \in \mathcal{V}_t} \sum_{v' \in \mathcal{V}_s} H\big(p_t^{(v)},\, p_s^{(v')}\big),$$5, balanced accuracy $$L = \sum_{v \in \mathcal{V}_t} \sum_{v' \in \mathcal{V}_s} H\big(p_t^{(v)},\, p_s^{(v')}\big),$$6, and accuracy $$L = \sum_{v \in \mathcal{V}_t} \sum_{v' \in \mathcal{V}_s} H\big(p_t^{(v)},\, p_s^{(v')}\big),$$7; and RedDino-large achieves weighted F1 $$L = \sum_{v \in \mathcal{V}_t} \sum_{v' \in \mathcal{V}_s} H\big(p_t^{(v)},\, p_s^{(v')}\big),$$8, balanced accuracy $$L = \sum_{v \in \mathcal{V}_t} \sum_{v' \in \mathcal{V}_s} H\big(p_t^{(v)},\, p_s^{(v')}\big),$$9, and accuracy $$\theta_t \leftarrow m\, \theta_t + (1-m)\, \theta_s.$$0. The reported improvement of the best RedDino over the best baseline is +2.9 weighted F1, +3.0 balanced accuracy, and +3.6 accuracy.

On DSE/RedTell, the advantage is narrower. RedDino-large reaches weighted F1 $$\theta_t \leftarrow m\, \theta_t + (1-m)\, \theta_s.$$1, balanced accuracy $$\theta_t \leftarrow m\, \theta_t + (1-m)\, \theta_s.$$2, and accuracy $$\theta_t \leftarrow m\, \theta_t + (1-m)\, \theta_s.$$3. This is a slight improvement in weighted F1 and accuracy over the best baseline value of $$\theta_t \leftarrow m\, \theta_t + (1-m)\, \theta_s.$$4, but balanced accuracy is slightly lower than the best baseline, $$\theta_t \leftarrow m\, \theta_t + (1-m)\, \theta_s.$$5, by 0.6 points.

The feature analyses are used to support claims about morphological sensitivity. PCA and UMAP visualizations show class-separable structure. In patches containing Plasmodium falciparum infections, embeddings reportedly highlight parasites, membranes, and cell regions distinctly. Echinocytes are differentiated in brightfield images without supervision. On Elsafty source 1, UMAP clusters align with classes, but “Rounded RBCs,” “Ovalocytes,” and “Borderline Ovalocytes” overlap, whereas “Clumps” form distinct clusters. This suggests that RedDino captures meaningful morphology while retaining ambiguity where the clinical boundaries themselves are overlapping.

5. Relation to RAD-DINO and the broader DINO-based medical imaging literature

The other major referent associated with the string “RedDino” is RAD-DINO, a biomedical image encoder trained solely on unimodal biomedical imaging data, specifically chest X-rays, rather than on RBC smears (Pérez-García et al., 2024). RAD-DINO uses a ViT-B/14 backbone of approximately 87M parameters, is continually pretrained from a DINOv2 ViT-B initialization, and combines image-level self-distillation with patch-level masked image modeling. Its pretraining corpus, Multi-CXR, contains 838,336 images from 439,177 subjects and includes BRAX, CheXpert, MIMIC-CXR, NIH-CXR, PadChest, and a private outpatient cohort.

RAD-DINO is positioned against language-supervised biomedical encoders such as BioViL, BiomedCLIP, MRM, and CLIP variants. The paper argues that language supervision can be limiting in radiology because reports omit many clinically relevant visual details, create undesired invariances through catch-all phrases, and are scarce at scale because of protected health information constraints. On this basis, RAD-DINO challenges the assumption that text supervision is required for strong biomedical visual representations.

Empirically, RAD-DINO matches or exceeds state-of-the-art language-supervised models on several chest X-ray benchmarks. On VinDr-CXR linear probing with frozen encoders, RAD-DINO reports macro AUPRC 66.63, compared with 60.86 for MRM, 45.42 for BiomedCLIP, 34.19 for BioViL-T, 34.54 for CLIP@336, and 33.03 for CLIP@224. On report generation with a frozen encoder and a Vicuna-7B v1.5 decoder in an LLaVA-style pipeline, RAD-DINO reaches ROUGE-L 24.6, BLEU-4 9.3, RG_ER 22.8, and CheXbert Macro-F1-14 31.9, exceeding the listed biomedical baselines. It also correlates more strongly with sex, age, weight, and BMI than language-supervised chest X-ray encoders.

The relation between the two models is therefore architectural rather than task-specific. Both are unimodal, self-supervised, and DINO-derived. Both emphasize that image-only supervision can yield transferable medical representations. Yet RedDino is an RBC morphology foundation model with RBC-specific changes such as removal of KoLeo and elimination of local crops, whereas RAD-DINO is a chest X-ray encoder that retains the DINOv2 hybrid of global contrastive distillation and patch-level masked image modeling.

6. Practical usage, limitations, and common confusions

RedDino’s code is available at https://github.com/Snarci/RedDino, and pretrained models are distributed via a Hugging Face collection. The available variants are small, base, and large, corresponding to 22M, 86M, and 304M parameters, respectively (Zedda et al., 11 Aug 2025). The base model is described as offering a favorable performance/efficiency trade-off. For inference, the paper recommends extracting embeddings from 224×224 inputs and using cosine-similarity retrieval or kNN, with $$\theta_t \leftarrow m\, \theta_t + (1-m)\, \theta_s.$$6 reported as robust in their experiments, or linear probing with logistic regression or a small MLP.

Several limitations are explicit. The paper does not provide the full label taxonomy for the evaluation datasets; exact augmentation operations and hyperparameters are not listed; key DINOv2 training details such as temperatures, momentum, optimizer, learning-rate schedule, multi-crop counts, and projection dimensions are omitted; and the feature extraction method at inference is not fully specified. Some morphology classes remain difficult to separate, particularly overlapping ovalocyte-related categories. These limitations mean that RedDino is best understood as a strong empirical demonstration of RBC-specialized self-supervision rather than as a fully specified closed-form recipe.

A further source of confusion is purely nominal. RedDino should not be conflated with REDI, an unrelated DINOv3 method titled “REDI: Corpus Aware Patch Ranking for DINOv3 Token Reduction,” which studies token reduction on ImageNet-1K with a frozen DINOv3 ViT-B/16 backbone and a fixed keep–merge–compress operator (Im et al., 30 Jun 2026). REDI is a controlled reference scoring mechanism for patch ranking; it is not an RBC foundation model, and it is not RAD-DINO.

Taken together, the literature supports two stable conclusions. First, RedDino, in the strict sense, is a self-supervised RBC foundation model whose performance gains derive from RBC-specific modifications to DINOv2 and from training on contextual smear patches at scale. Second, the informal use of “RedDino” for RAD-DINO reflects name proximity rather than model identity. The former addresses computational hematology; the latter addresses chest X-ray representation learning beyond language supervision.