Step-Attention Refinement of DINOv3 Features for Efficient Anterior Eye Segmentation

Cl\'ement Playout; Jean-Mathieu Gagnon; Lama S\'eoud; Mathieu Duchesneau; Philippe Baumstimler; S\'ebastien Gagn\'e

REVIEW 4 major objections 5 minor 23 references

A DINOv3-based model with step-attention refinement reaches 85.55% mIoU on clinical anterior eye segmentation and transfers best to unseen datasets.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 01:19 UTC pith:O6WJQEAA

load-bearing objection A plausible decoder idea and a solid in-domain benchmark, but the cross-domain headline rests on margins that are within noise and the paper needs OOD error bars before the robustness claim can stand. the 4 major comments →

arxiv 2607.27087 v2 pith:O6WJQEAA submitted 2026-07-29 cs.CV

Step-Attention Refinement of DINOv3 Features for Efficient Anterior Eye Segmentation

Philippe Baumstimler , Jean-Mathieu Gagnon , S\'ebastien Gagn\'e , Mathieu Duchesneau , Cl\'ement Playout , Lama S\'eoud This is my paper

classification cs.CV

keywords anterior eye segmentsemantic segmentationDINOv3foundation model adaptationstep-attention feature refinementdomain shiftclinical imagingtransformer decoder

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Anterior eye segment segmentation, used in ocular biometrics and clinical analysis, must cope with varied acquisition conditions and scarce annotations. This paper proposes a compact architecture built on a DINOv3 ViT-Small backbone, with a step-attention feature refinement module that progressively fuses features from multiple transformer layers before a light convolutional decoder. The central claim is that for the large, low-frequency anatomical structures of the eye, semantically rich deep representations matter more than high-resolution decoding. Fully fine-tuned, the model reaches 85.55% mIoU on a private clinical dataset of 333 images spanning eight acquisition protocols, outperforming convolutional and transformer baselines including heavier DINOv3-based decoders, and shows the strongest robustness to domain shift on four unseen public segmentation datasets. If right, this makes a strong parameter-efficient baseline for clinical eye segmentation and points to decoder design as the key to adapting foundation models for medical dense prediction.

Core claim

The central claim is that a step-attention feature refinement module—K+1 transformer-style blocks (multi-head self-attention with RoPE and MLP) that fuse DINOv3 ViT-Small features from layers 3, 6, 9, and 12 deep-to-shallow in low resolution, followed by a four-layer convolutional decoder—yields state-of-the-art anterior eye segment segmentation. Fully fine-tuned at 32.9M parameters, it reaches 85.55% mIoU on a private clinical dataset, beating UNet, SegFormer, DPT, and SegDINO; on four unseen public datasets it attains the highest mIoU in every reported case. The authors attribute this to adaptation of pretrained representations without expensive high-resolution decoding.

What carries the argument

The step-attention feature refinement (SAFR) module: a sequence of K+1 refinement stages, each combining multi-head self-attention with rotary position embeddings and a MLP, that progressively fuse feature maps extracted from layers 3, 6, 9, and 12 of a DINOv3 ViT-Small, going from deepest to shallowest. Only the final refined representation is passed to a lightweight convolutional decoder (bilinear upsampling, 5x5 convolutions, SiLU, layer norm, 1x1 projection). This machinery adapts the foundation-model features to the segmentation task while keeping all refinement at low resolution, and is the main claimed source of the accuracy and cross-domain robustness.

Load-bearing premise

The central premise is that anterior eye segmentation is dominated by large, low-frequency structures, so accurate segmentation can be achieved by refining deep semantic representations at low resolution without high-resolution skip connections; if fine structures such as eyelash occlusions, thin vessels, or pathological boundaries turn out to be decisive, the approach's advantage could vanish or reverse.

What would settle it

An experiment that isolates performance on fine-grained regions—for example, adding a test set of images with prominent eyelash occlusions or measuring mIoU on scleral vessels—and comparing against a high-resolution decoder with skip connections could falsify the core hypothesis. If the high-resolution model outperforms the step-attention model on those fine-detail classes while the two are comparable on coarse classes, the low-frequency assumption is the limiting factor.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

A compact 32.9M-parameter model can outperform much larger and more complex decoders for coarse anatomical segmentation, implying that parameter-efficient clinical deployment is feasible without sacrificing accuracy.
Decoder design, not just backbone choice, is decisive when adapting DINOv3 representations to dense prediction: the step-attention refinement beats an MLP decoder (SegDINO) and a heavier convolutional decoder (DPT) using the same backbone.
The approach generalizes to unseen public datasets spanning biometric and clinical settings, with the highest mIoU on all four external test sets in the reported comparison, indicating robustness to domain shift in acquisition conditions.
Even with a frozen encoder, the method matches DPT-level performance with nearly five times fewer trainable parameters, so the module can be used as an efficient fine-tuning strategy.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The low-frequency assumption could extend to other medical segmentation targets dominated by large organs (e.g., lung, liver, kidney in CT), where a step-attention refinement might replace heavy decoders with little loss; this is a testable hypothesis beyond the eye domain.
The ablation's finding that removing the class token improves performance suggests that global tokens can be detrimental for dense prediction; testing this in other DINOv3-based segmenters could lead to simpler, better adaptation protocols.
The significant residual domain gap on MOBIUS (mIoU 59.61 vs 85.55 in-domain) implies that the module does not eliminate acquisition-specific variation; combining SAFR with explicit domain adaptation or augmentation might close the gap further, an extension the paper does not explore.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

A plausible decoder idea and a solid in-domain benchmark, but the cross-domain headline rests on margins that are within noise and the paper needs OOD error bars before the robustness claim can stand.

read the letter

The step-attention refinement idea is worth a look. The specific combination — progressive deep-to-shallow fusion of DINOv3 layer features with RoPE and a light conv upsampler — isn't in SegDINO or DPT, and the paper makes a genuinely useful empirical point: the decoder still matters a lot when adapting a strong self-supervised backbone to dense prediction. The in-domain experiment is one of the better parts. Patient-level split, five seeds, per-class IoU with standard deviations, and an honest discussion of why the rare classes (caruncle, artifacts) behave badly. The parameter efficiency is real: frozen-encoder version beats SegDINO by a wide margin with only 11.3M trainable parameters, and the fully fine-tuned version is competitive with much larger decoders.

The soft spot is the headline robustness claim, and the stress-test note gets this right. Table 1 does show this method as the best mIoU on each of the four OOD datasets — the reader's phrase "contradicted by its own Table 1" is incorrect. But on SLID the lead over DPT is 0.03 mIoU (85.53 vs 85.50), which is well within any reasonable noise estimate. No OOD error bars or significance tests are reported, even though the in-domain table has per-class standard deviations. That asymmetry is conspicuous. The OOD numbers are also macro-averages over different class subsets (Eyes-Defy-Anemia has only palpebral conjunctiva, SLID/MOBIUS have pupil/iris/sclera, SBVPI has scattered structures), so a single comparison table overstates consistency. The abstract's "strongest robustness to domain shift across four unseen public datasets" is too strong if one of the four leads is a coin flip.

Other limitations are more minor but still worth noting: no code or data release, so the main private dataset is unavailable, and the ablation study is a bit sparse — it would help to see the effect of varying the extracted layer set, since that's a free parameter the paper doesn't explore.

This paper is for readers working on anterior eye segmentation or on efficient dense-prediction decoders for DINO-family backbones. They will get a reasonable architectural baseline and a clear experimental setup they can build on. It deserves a serious referee, not a desk rejection. In review, I'd ask for OOD error bars, a significance test or at least a bootstrap confidence interval on the cross-domain numbers, and a rephrasing of the robustness claim to match what the evidence actually supports. With those changes, the paper would be solid for its niche.

Referee Report

4 major / 5 minor

Summary. The paper proposes a lightweight DINOv3 ViT-Small-based architecture with a step-attention feature refinement module for anterior eye segment (AES) segmentation. It introduces a private clinical dataset (LightX, 333 images, 7 classes) and evaluates the method on this in-domain set and on four public out-of-distribution datasets. The authors report that the fully fine-tuned model achieves 85.55% mIoU on LightX, outperforming convolutional and transformer baselines, and claim the strongest robustness to domain shift across the four unseen datasets.

Significance. If the claims hold, the work provides a parameter-efficient decoder design for adapting a foundation model to medical segmentation, plus a new clinical AES dataset and extensive external validation. The in-domain experimental design is solid in several respects: patient-level split, five seeds, per-class mIoU with standard deviations, and a reasonable set of baselines with a consistent DINOv3 backbone for the closest comparisons. The main weakness is that the headline cross-domain robustness claim is not supported by the paper's own numbers on SLID, and the out-of-distribution results are reported without any variance or significance testing, making the ranking statistically fragile.

major comments (4)

[Abstract & §3 Comparison study, Table 1] The abstract claims 'strongest robustness to domain shift across four unseen public AES segmentation datasets.' Table 1 contradicts this: on SLID, UNet-EfficientNet-b5 reaches 86.09 mIoU while the proposed method reaches 85.53 mIoU. Thus the method is best on three of the four OOD datasets, not all four. The text in §3 ('with our model consistently achieving the best cross-dataset performance') is also inaccurate. Please correct the claim or justify it by a different aggregation (e.g., macro-average across datasets) and state that explicitly.
[§3 Comparison study, Table 1] No per-seed standard deviations or statistical significance tests are reported for any out-of-distribution result, although Table 2 gives per-class stds for the in-domain set. The OOD margins are often small: on SLID the proposed method (85.53) is only 0.03 mIoU above DPT (85.50), and on Eyes-Defy-Anemia the lead over SegFormer-b3 is 1.23 points. Without error bars or paired tests, the claimed cross-domain superiority may be noise. Report std/confidence intervals for all OOD metrics and run paired significance tests between the proposed method and each baseline.
[§3 Ablation study, Table 3] The ablation table and its discussion are internally inconsistent. The text states 'removing the class token greatly improves performance,' but the first row is identified as the proposed approach and has the highest mIoU (79.09±1.65); rows lacking components show lower mIoU. If the first row includes the class token, the sentence is contradicted by the table. If the row labels/checkmarks are mis-rendered, they must be corrected. This ambiguity undermines the ablation conclusion about the class token and about SAFR's contribution.
[§2.2] The core design hypothesis—that AES segmentation primarily involves large, low-frequency structures and therefore does not need high-resolution skip connections—is asserted without direct evidence. The paper's own failure cases (eyelash occlusions, fine-grained details in LightX; MOBIUS acquisition differences) suggest that high-resolution information may matter precisely when the method struggles. A comparison against a variant with high-resolution skip connections, or at least a quantitative analysis of boundary/small-structure performance, would strengthen the claim that low-resolution refinement is sufficient.

minor comments (5)

[Table 3 caption] 'pirvate' should be 'private'.
[References] Reference [5] is cited for 'EyePACS AES' but [5] describes the EyePACS telemedicine system, not an AES dataset. Provide a dedicated dataset reference or clarify the provenance.
[§3 Data] MOBIUS is cited through the SSBC 2020 paper [17], but the dataset itself is not clearly specified (which subset, annotation protocol). Please clarify the exact MOBIUS version and evaluation protocol used.
[Table 2] The Artifacts class result for the proposed fully fine-tuned model (95.5±1.3 mIoU) is much higher than all baselines (63-82) and seems implausible given the class has only 1.0% pixel ratio. Verify this number and explain the large gap, or check for label leakage/artifact definition issues.
[Abstract / §3 Implementation details] The paper emphasizes efficiency and lightweight design but reports only parameter counts. Add inference time, FLOPs, or memory usage to substantiate the efficiency claim.

Circularity Check

0 steps flagged

No significant circularity; empirical architecture benchmark with independent held-out and external evaluation.

full rationale

The paper is an empirical architecture study, not a derivation. Performance is measured on a held-out test split (LightX) and four unseen public datasets (Eyes-Defy-Anemia, SLID, MOBIUS, SBVPI), with the proposed method compared against external baselines. No predicted quantity is defined in terms of a fitted constant or an input parameter; the mIoU values are measured outcomes. The step-attention refinement module is a novel architectural component whose contribution is tested by ablation (Table 3), and the design hypothesis in Section 2.2 is a qualitative motivation, not a fitted input. DINOv3 and SegDINO are external prior works and are not used to justify the paper's own empirical claims. The abstract's cross-domain claim is consistent with Table 1: Ours achieves the highest mIoU on each of the four OOD datasets. The absence of significance tests on OOD comparisons is a statistical robustness concern, not circularity. There are no self-citations, no uniqueness theorems, and no renamed known results. The result is therefore self-contained evidence for the stated conclusions.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The paper's central contribution is an empirical benchmark on a private dataset; it performs no theoretical derivation. The main free choices are training/architecture hyperparameters and the unproven low-frequency hypothesis. The proposed SAFR module is not a hidden physical entity; it is directly evaluated on held-out test data, so it carries an empirical handle.

free parameters (4)

Extracted transformer layers L = {3,6,9,12}
Chosen following SegDINO; all multi-scale fusion and the K+1 refinement blocks depend on this hand-picked set.
Number of refinement blocks = K+1 = 5
Four fused feature levels plus one final block; selected ad hoc to keep the decoder lightweight.
Per-architecture learning rate = in [1e-5, 1e-3], cosine to 1e-6
Selected per architecture on the validation set; this can favor one decoder design over another.
Class loss weights = inverse squared frequency
Generalized Dice reweighting mitigates class imbalance; it directly shapes per-class mIoU, notably for caruncle, palpebral conjunctiva, and artifacts.

axioms (4)

domain assumption DINOv3 ViT-Small pretrained on LVD-1689M provides transferable representations for clinical AES images.
Used as backbone in section 2.1; no domain-specific pretraining and no proof that the pretraining distribution covers slit-lamp/smartphone eye images.
ad hoc to paper AES structures are large, low-frequency, and need little high-resolution decoding detail.
Stated as a hypothesis in section 2.2; the whole low-resolution refinement design and the claim that it suffices rely on this premise.
domain assumption Summing refined deep features with shallower features in the latent space preserves the spatial information needed for boundaries.
The fusion rule in Eq. (4) uses element-wise addition without learned alignment; position info is only via RoPE inside MHSA, not a proven guarantee.
domain assumption Macro-averaged mIoU over each dataset's available classes is a fair way to compare cross-domain robustness when datasets have different class label sets.
Table 1 footnote states macro average across available classes; datasets cover different class subsets, so a single 'best cross-dataset' claim is not uniquely defined.

invented entities (1)

Step-Attention Feature Refinement (SAFR) module independent evidence
purpose: Progressively fuses multi-level DINOv3 features in the low-resolution latent space before convolutional decoding.
Architectural construction, not a physical entity; evaluated on the held-out LightX test split and public OOD datasets, so it has an independent empirical handle.

pith-pipeline@v1.3.0-alltime-deepseek · 3542 in / 3806 out tokens · 191040 ms · 2026-08-03T01:19:21.989078+00:00 · methodology

0 comments

read the original abstract

Anterior eye segment (AES) segmentation is a key component of both ocular biometrics and emerging clinical image analysis applications. However, heterogeneous acquisition conditions and limited annotations in medical settings hinder the robustness and generalization of existing methods. Foundation models (FMs) such as DINOv3 offer strong transfer capabilities, but efficiently adapting their representations to dense prediction tasks remains challenging. In this study, we investigate robust AES segmentation in clinical settings, and propose a lightweight architecture built upon a distilled DINOv3 ViT-Small backbone. We introduce a step-attention feature refinement module that progressively adapts multi-level transformer representations before convolutional decoding, enabling efficient exploitation of pretrained features with few parameters. We evaluate the proposed approach on a private dataset of 333 clinically acquired AES images spanning eight ophthalmic acquisition protocols and annotated for seven anatomical classes. Compared with convolutional and transformer-based baselines, including DINOv3-based methods, our approach achieves the best overall performance, reaching 85.55\% mIoU when fully fine-tuned. It also demonstrates the strongest robustness to domain shift across four unseen public AES segmentation datasets. These results establish a strong baseline for robust AES segmentation in clinical settings and highlight the importance of decoder design for effectively adapting FMs representations to medical segmentation tasks.

Figures

Figures reproduced from arXiv: 2607.27087 by Cl\'ement Playout, Jean-Mathieu Gagnon, Lama S\'eoud, Mathieu Duchesneau, Philippe Baumstimler, S\'ebastien Gagn\'e.

**Figure 2.** Figure 2: Qualitative segmentation results for the proposed method and competing [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

23 extracted references · 5 linked inside Pith

[1]

Com- puters in Biology and Medicine197, 111026 (2025)

Camporeale, M., Clemente, F., Dimauro, G., Lomonte, N., Maglietta, R., Pas- ciolla, C., Sacco, D., Zaccaria, G.M.: Highly reliable personalized noninvasive hemoglobin estimation by using vision transformers and dual fine-tuning. Com- puters in Biology and Medicine197, 111026 (2025). https://doi.org/https://doi. org/10.1016/j.compbiomed.2025.111026, https:...

arXiv 2025
[2]

In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)

Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)

2021
[3]

In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW)

Chaudhary, A.K., Kothari, R., Acharya, M., Dangi, S., Nair, N., Bailey, R., Kanan, C., Diaz, G., Pelz, J.B.: Ritnet: real-time semantic segmentation of the eye for gaze tracking. In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW). pp. 3698–3702. IEEE (2019)

2019
[4]

Chinese Academy of Sciences’ Institute of Automation (CASIA): Casia iris image database version 4.0 (2018), https://hycasia.github.io/dataset/casia-irisv4/

2018
[5]

Journal of Diabetes Science and Technology3(3), 509–516 (May 2009)

Cuadros, J., Bresnick, G.: Eyepacs: An adaptable telemedicine system for diabetic retinopathy screening. Journal of Diabetes Science and Technology3(3), 509–516 (May 2009). https://doi.org/10.1177/193229680900300315, http://dx.doi.org/10. 1177/193229680900300315

doi:10.1177/193229680900300315 2009
[6]

https://doi.org/10.21227/t5s2-4j73, https://dx.doi.org/10.21227/ t5s2-4j73

Dimauro, G., Maglietta, R., Bai, T., Kasiviswanathan, S.: Eyes-defy- anemia (2022). https://doi.org/10.21227/t5s2-4j73, https://dx.doi.org/10.21227/ t5s2-4j73

doi:10.21227/t5s2-4j73 2022
[7]

Nature Biomedical Engineering4(8), 767–777 (Jun 2020)

Li, W., Yang, Y., Zhang, K., Long, E., He, L., Zhang, L., Zhu, Y., Chen, C., Liu, Z., Wu, X., Yun, D., Lv, J., Liu, Y., Liu, X., Lin, H.: Dense anatomi- cal annotation of slit-lamp images improves the performance of deep learning for the diagnosis of ophthalmic disorders. Nature Biomedical Engineering4(8), 767–777 (Jun 2020). https://doi.org/10.1038/s4155...

doi:10.1038/s41551-020-0577-y 2020
[8]

Loshchilov, I., Hutter, F.: Decoupled weight decay regularization (2019), https: //arxiv.org/abs/1711.05101

Pith/arXiv arXiv 2019
[9]

Ophthalmology Science5(4), 100757 (2025)

Nahass, G.R., Koehler, E., Tomaras, N., Lopez, D., Cheung, M., Palacios, A., Peterson, J.C., Hubschman, S., Green, K., Purnell, C.A., Setabutr, P., Tran, A.Q., Yi, D.: Open-source periorbital segmentation dataset for ophthalmic applications. Ophthalmology Science5(4), 100757 (2025). https://doi.org/10.1016/j.xops.2025. 100757, http://dx.doi.org/10.1016/j....

arXiv 2025
[10]

ArXiv preprint (2021)

Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. ArXiv preprint (2021)

2021
[11]

In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) Medical Image Computing and Computer-Assisted Intervention – MICCAI
[12]

In: Uhl, A., Busch, C., Marcel, S., Veldhuis, R.N.J

Rot, P., Vitek, M., Grm, K., Emeršič, v., Peer, P., Štruc, V.: Deep sclera seg- mentation and recognition. In: Uhl, A., Busch, C., Marcel, S., Veldhuis, R.N.J. (eds.) Handbook of Vascular Biometrics (HVB), pp. 395–432. Springer (2020). https://doi.org/10.1007/978-3-030-27731-4_13

doi:10.1007/978-3-030-27731-4_13 2020
[13]

Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khalidov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., Massa, F., Haziza, D., Wehrstedt, L., Wang, J., Darcet, T., Moutakanni, T., Sentana, L., Roberts, C., Vedaldi, A., Tolan, J., Brandt, J., Couprie, C., Mairal, J., Jégou, H., Labatut, P., Bojanowski, P.: DINOv3 (2025), https://a...

Pith/arXiv arXiv 2025
[14]

Su, J., Lu, Y., Pan, S., Murtadha, A., Wen, B., Liu, Y.: Roformer: Enhanced transformer with rotary position embedding (2023), https://arxiv.org/abs/2104. 09864

2023
[15]

Sudre, C.H., Li, W., Vercauteren, T., Ourselin, S., Jorge Cardoso, M.: Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In: Cardoso, M.J., Arbel, T., Carneiro, G., Syeda-Mahmood, T., Tavares, J.M.R., Moradi, M., Bradley, A., Greenspan, H., Papa, J.P., Madabhushi, A., Nascimento, J.C., Cardoso, J.S., Belagianni...

2017
[16]

Tan, M., Le, Q.V.: Efficientnet: Rethinking model scaling for convolutional neural networks (2020), https://arxiv.org/abs/1905.11946

Pith/arXiv arXiv 2020
[17]

In: IEEE International Joint Conference on Biometrics (IJCB)

Vitek, M., Das, A., Pourcenoux, Y., Missler, A., Paumier, C., Das, S., De Ghosh, I., Lucio, D.R., Zanlorensi Jr., L.A., Menotti, D., Boutros, F., Damer, N., Grebe, J.H., Kuijper, A., Hu, J., He, Y., Wang, C., Liu, H., Wang, Y., Sun, Z., Osorio- Roig, D., Rathgeb, C., Busch, C., Tapia Farias, J., Valenzuela, A., Zampoukis, G., Tsochatzidis, L., Pratikakis,...

arXiv 2020
[18]

Vitek, M., Tomašević, D., Das, A., Nathan, S., Özbulak, G., Özbulak, G.A.T., Calbimonte, J.P., Anjos, A., Bhatt, H.H., Premani, D.D., Chaudhari, J., Wang, C., Jiang, J., Zhang, C., Zhang, Q., Ganapathi, I.I., Ali, S.S., Velayudan, D., Assefa, M., Werghi, N., Daniels, Z.A., John, L., Vyas, R., Khiarak, J.N., Saeed, T.A., Nasehi, M., Kianfar, A., Panahi, M....

Pith/arXiv arXiv 2025
[19]

Annals of the New York Academy of Sciences1551(1), 201–209 (Aug 2025)

Wong, D., Ng, Y., Eppenberger, L.S., Cherecheanu, A.P., Anghelache, A., Toma, E., Coroleuca, R., Garcia-Feijoo, J., Garhöfer, G., Schmetterer, L.: Toward au- tomated assessment of conjunctival hyperemia: A semisupervised artificial intelli- gence approach. Annals of the New York Academy of Sciences1551(1), 201–209 (Aug 2025). https://doi.org/10.1111/nyas....

doi:10.1111/nyas.70009 2025
[20]

In: Neural Information Processing Systems (NeurIPS) (2021)

Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. In: Neural Information Processing Systems (NeurIPS) (2021)

2021
[21]

Frontiers in Digital Health 7(Jan 2026)

Xu, M., Sun, Y., Cheng, H., Zhou, Y., Maimaiti, N., Chen, P., Miao, Q., Xu, P., Ye, J.: Slid: a slit-lamp image dataset for deep learning-based anterior eye anatomical segmentation and multi-lesion detection. Frontiers in Digital Health 7(Jan 2026). https://doi.org/10.3389/fdgth.2025.1716501, http://dx.doi.org/10. 3389/fdgth.2025.1716501

arXiv 2026
[22]

arXiv preprint arXiv:2509.00833 (2025), https://arxiv.org/abs/2509.00833

Yang, S., Wang, H., Xing, Z., Chen, S., Zhu, L.: Segdino: An efficient de- sign for medical and natural image segmentation with dino-v3. arXiv preprint arXiv:2509.00833 (2025), https://arxiv.org/abs/2509.00833

Pith/arXiv arXiv 2025
[2015]

pp. 234–241. Springer International Publishing, Cham (2015)

2015

[1] [1]

Com- puters in Biology and Medicine197, 111026 (2025)

Camporeale, M., Clemente, F., Dimauro, G., Lomonte, N., Maglietta, R., Pas- ciolla, C., Sacco, D., Zaccaria, G.M.: Highly reliable personalized noninvasive hemoglobin estimation by using vision transformers and dual fine-tuning. Com- puters in Biology and Medicine197, 111026 (2025). https://doi.org/https://doi. org/10.1016/j.compbiomed.2025.111026, https:...

arXiv 2025

[2] [2]

In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)

Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)

2021

[3] [3]

In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW)

Chaudhary, A.K., Kothari, R., Acharya, M., Dangi, S., Nair, N., Bailey, R., Kanan, C., Diaz, G., Pelz, J.B.: Ritnet: real-time semantic segmentation of the eye for gaze tracking. In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW). pp. 3698–3702. IEEE (2019)

2019

[4] [4]

Chinese Academy of Sciences’ Institute of Automation (CASIA): Casia iris image database version 4.0 (2018), https://hycasia.github.io/dataset/casia-irisv4/

2018

[5] [5]

Journal of Diabetes Science and Technology3(3), 509–516 (May 2009)

Cuadros, J., Bresnick, G.: Eyepacs: An adaptable telemedicine system for diabetic retinopathy screening. Journal of Diabetes Science and Technology3(3), 509–516 (May 2009). https://doi.org/10.1177/193229680900300315, http://dx.doi.org/10. 1177/193229680900300315

doi:10.1177/193229680900300315 2009

[6] [6]

https://doi.org/10.21227/t5s2-4j73, https://dx.doi.org/10.21227/ t5s2-4j73

Dimauro, G., Maglietta, R., Bai, T., Kasiviswanathan, S.: Eyes-defy- anemia (2022). https://doi.org/10.21227/t5s2-4j73, https://dx.doi.org/10.21227/ t5s2-4j73

doi:10.21227/t5s2-4j73 2022

[7] [7]

Nature Biomedical Engineering4(8), 767–777 (Jun 2020)

Li, W., Yang, Y., Zhang, K., Long, E., He, L., Zhang, L., Zhu, Y., Chen, C., Liu, Z., Wu, X., Yun, D., Lv, J., Liu, Y., Liu, X., Lin, H.: Dense anatomi- cal annotation of slit-lamp images improves the performance of deep learning for the diagnosis of ophthalmic disorders. Nature Biomedical Engineering4(8), 767–777 (Jun 2020). https://doi.org/10.1038/s4155...

doi:10.1038/s41551-020-0577-y 2020

[8] [8]

Loshchilov, I., Hutter, F.: Decoupled weight decay regularization (2019), https: //arxiv.org/abs/1711.05101

Pith/arXiv arXiv 2019

[9] [9]

Ophthalmology Science5(4), 100757 (2025)

Nahass, G.R., Koehler, E., Tomaras, N., Lopez, D., Cheung, M., Palacios, A., Peterson, J.C., Hubschman, S., Green, K., Purnell, C.A., Setabutr, P., Tran, A.Q., Yi, D.: Open-source periorbital segmentation dataset for ophthalmic applications. Ophthalmology Science5(4), 100757 (2025). https://doi.org/10.1016/j.xops.2025. 100757, http://dx.doi.org/10.1016/j....

arXiv 2025

[10] [10]

ArXiv preprint (2021)

Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. ArXiv preprint (2021)

2021

[11] [11]

In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) Medical Image Computing and Computer-Assisted Intervention – MICCAI

[12] [12]

In: Uhl, A., Busch, C., Marcel, S., Veldhuis, R.N.J

Rot, P., Vitek, M., Grm, K., Emeršič, v., Peer, P., Štruc, V.: Deep sclera seg- mentation and recognition. In: Uhl, A., Busch, C., Marcel, S., Veldhuis, R.N.J. (eds.) Handbook of Vascular Biometrics (HVB), pp. 395–432. Springer (2020). https://doi.org/10.1007/978-3-030-27731-4_13

doi:10.1007/978-3-030-27731-4_13 2020

[13] [13]

Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khalidov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., Massa, F., Haziza, D., Wehrstedt, L., Wang, J., Darcet, T., Moutakanni, T., Sentana, L., Roberts, C., Vedaldi, A., Tolan, J., Brandt, J., Couprie, C., Mairal, J., Jégou, H., Labatut, P., Bojanowski, P.: DINOv3 (2025), https://a...

Pith/arXiv arXiv 2025

[14] [14]

Su, J., Lu, Y., Pan, S., Murtadha, A., Wen, B., Liu, Y.: Roformer: Enhanced transformer with rotary position embedding (2023), https://arxiv.org/abs/2104. 09864

2023

[15] [15]

Sudre, C.H., Li, W., Vercauteren, T., Ourselin, S., Jorge Cardoso, M.: Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In: Cardoso, M.J., Arbel, T., Carneiro, G., Syeda-Mahmood, T., Tavares, J.M.R., Moradi, M., Bradley, A., Greenspan, H., Papa, J.P., Madabhushi, A., Nascimento, J.C., Cardoso, J.S., Belagianni...

2017

[16] [16]

Tan, M., Le, Q.V.: Efficientnet: Rethinking model scaling for convolutional neural networks (2020), https://arxiv.org/abs/1905.11946

Pith/arXiv arXiv 2020

[17] [17]

In: IEEE International Joint Conference on Biometrics (IJCB)

Vitek, M., Das, A., Pourcenoux, Y., Missler, A., Paumier, C., Das, S., De Ghosh, I., Lucio, D.R., Zanlorensi Jr., L.A., Menotti, D., Boutros, F., Damer, N., Grebe, J.H., Kuijper, A., Hu, J., He, Y., Wang, C., Liu, H., Wang, Y., Sun, Z., Osorio- Roig, D., Rathgeb, C., Busch, C., Tapia Farias, J., Valenzuela, A., Zampoukis, G., Tsochatzidis, L., Pratikakis,...

arXiv 2020

[18] [18]

Vitek, M., Tomašević, D., Das, A., Nathan, S., Özbulak, G., Özbulak, G.A.T., Calbimonte, J.P., Anjos, A., Bhatt, H.H., Premani, D.D., Chaudhari, J., Wang, C., Jiang, J., Zhang, C., Zhang, Q., Ganapathi, I.I., Ali, S.S., Velayudan, D., Assefa, M., Werghi, N., Daniels, Z.A., John, L., Vyas, R., Khiarak, J.N., Saeed, T.A., Nasehi, M., Kianfar, A., Panahi, M....

Pith/arXiv arXiv 2025

[19] [19]

Annals of the New York Academy of Sciences1551(1), 201–209 (Aug 2025)

Wong, D., Ng, Y., Eppenberger, L.S., Cherecheanu, A.P., Anghelache, A., Toma, E., Coroleuca, R., Garcia-Feijoo, J., Garhöfer, G., Schmetterer, L.: Toward au- tomated assessment of conjunctival hyperemia: A semisupervised artificial intelli- gence approach. Annals of the New York Academy of Sciences1551(1), 201–209 (Aug 2025). https://doi.org/10.1111/nyas....

doi:10.1111/nyas.70009 2025

[20] [20]

In: Neural Information Processing Systems (NeurIPS) (2021)

Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. In: Neural Information Processing Systems (NeurIPS) (2021)

2021

[21] [21]

Frontiers in Digital Health 7(Jan 2026)

Xu, M., Sun, Y., Cheng, H., Zhou, Y., Maimaiti, N., Chen, P., Miao, Q., Xu, P., Ye, J.: Slid: a slit-lamp image dataset for deep learning-based anterior eye anatomical segmentation and multi-lesion detection. Frontiers in Digital Health 7(Jan 2026). https://doi.org/10.3389/fdgth.2025.1716501, http://dx.doi.org/10. 3389/fdgth.2025.1716501

arXiv 2026

[22] [22]

arXiv preprint arXiv:2509.00833 (2025), https://arxiv.org/abs/2509.00833

Yang, S., Wang, H., Xing, Z., Chen, S., Zhu, L.: Segdino: An efficient de- sign for medical and natural image segmentation with dino-v3. arXiv preprint arXiv:2509.00833 (2025), https://arxiv.org/abs/2509.00833

Pith/arXiv arXiv 2025

[23] [2015]

pp. 234–241. Springer International Publishing, Cham (2015)

2015