Pith. sign in

REVIEW 2 major objections 1 minor 14 references

A class-aware dual mixture-of-experts network segments scleral anomalies at 72.11 percent mean Dice on a new benchmark dataset.

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 · grok-4.3

2026-06-28 06:46 UTC pith:XWCINMK7

load-bearing objection This paper adds a new scleral anomaly dataset and a dual-MoE decoder on DINOv3 but reports no baselines, ablations, or numbers for its generalization claim. the 2 major comments →

arxiv 2606.04888 v1 pith:XWCINMK7 submitted 2026-06-03 cs.CV

HD-DinoMoE: A Class-Aware Hierarchical Dual Mixture-of-Experts Network for Scleral Anomaly Segmentation in Complex Acquisition Scenarios

Yinxiang Yu , Maoxiang Chu , Qi Niu , Guanghu Liu , Wei Xu , Haotian Wang , Zhi Chen , Yutian Zhu
show 2 more authors
Yuelong Fan Guanghao Liao
This is my paper
classification cs.CV
keywords scleral anomaly segmentationmixture of expertsDINOv3specular reflectionmedical image segmentationTraditional Chinese Medicinemulti-label datasetocular inspection
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.

This paper presents HD-DinoMoE to quantify scleral surface anomalies for Traditional Chinese Medicine ocular inspection in images from varied sources. The network fuses DINOv3 features using a class-aware dual-stream approach and decodes them with class-specific experts for three anomaly types. It employs a three-stage routing strategy, a Progressive Confidence Penalty Loss to control false positives in specular areas, and adaptive sample weighting. A new multi-label dataset ML-SASD is introduced with clinical, wild, and mixed settings. On the mix setting, it reaches 72.11% mean Dice and 58.44% mean IoU while generalizing competitively to a public dataset.

Core claim

HD-DinoMoE combines class-aware dual-stream DINOv3 feature fusion with class-specific multi-expert decoding, stabilized by three-stage backbone-frozen routing, to segment Vessels, Yellow and Black Spots, and Blood Spots. The Progressive Confidence Penalty Loss and Class-Aware Adaptive Sample Weighting further mitigate specular reflection issues and class imbalances. This yields 72.11% mean Dice and 58.44% mean IoU on the ML-SASD-Mix setting, with good boundary control and reduced false positives in specular regions, plus competitive results on SBVPI Vessels.

What carries the argument

The class-aware hierarchical dual mixture-of-experts network that performs dual-stream DINOv3 feature fusion and class-specific multi-expert decoding.

Load-bearing premise

The specific combination of dual-stream DINOv3 fusion, class-specific experts, three-stage routing, PCP loss, and CA-ASW sufficiently handles the challenges of distributional discrepancies, anomaly diversity, and specular reflections.

What would settle it

If a baseline model without these components matches or exceeds the 72.11% Dice and 58.44% IoU on ML-SASD-Mix while controlling specular false positives, the necessity of the proposed machinery would be questioned.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The manuscript introduces HD-DinoMoE, a class-aware hierarchical dual mixture-of-experts network that fuses DINOv3 features via dual streams and decodes with class-specific experts for segmenting Vessels, Yellow/Black Spots, and Blood Spots in scleral images. It proposes a three-stage backbone-frozen routing strategy, Progressive Confidence Penalty (PCP) Loss, and Class-Aware Adaptive Sample Weighting (CA-ASW), constructs the ML-SASD dataset (Clinical/Wild/Mix settings with pixel annotations), and reports mean Dice of 72.11% and mean IoU of 58.44% on ML-SASD-Mix while claiming competitive generalization on the Vessels subset of SBVPI.

Significance. If the performance numbers are shown to exceed strong baselines with ablations and statistical support, the work would offer a practical contribution to automated TCM ocular inspection by addressing multi-source shifts, anomaly variability, and specular reflections. The release of a new annotated multi-label scleral dataset is a clear positive; however, the current lack of comparative evidence makes it impossible to determine whether the architectural and loss innovations deliver meaningful gains.

major comments (2)
  1. [Abstract] Abstract: the assertion that the method 'shows competitive generalization on the Vessels subset of the public SBVPI dataset' supplies no numerical Dice/IoU values, standard deviations, comparison table, or definition of 'competitive'. This directly undermines the generalization component of the central claim.
  2. [Abstract] Abstract: the reported 72.11% Dice / 58.44% IoU on ML-SASD-Mix is presented without any baseline methods, ablation results on the dual-stream fusion / multi-expert decoder / PCP Loss / CA-ASW components, or statistical significance tests. These omissions are load-bearing because the soundness of the performance claim cannot be assessed without them.
minor comments (1)
  1. [Abstract] Abstract: phrases such as 'favorable boundary localization and specular-region false-positive control' are stated qualitatively; quantitative boundary metrics (e.g., Hausdorff distance) or region-specific false-positive rates would strengthen the description.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract. We agree that the abstract must be strengthened to stand alone and will revise it to include explicit numerical support and references to the detailed experiments. Point-by-point responses follow.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion that the method 'shows competitive generalization on the Vessels subset of the public SBVPI dataset' supplies no numerical Dice/IoU values, standard deviations, comparison table, or definition of 'competitive'. This directly undermines the generalization component of the central claim.

    Authors: We accept the observation. The full manuscript reports the SBVPI Vessels results (including Dice, IoU, standard deviations, and direct comparisons) in Section 4.3 and the corresponding table. The abstract will be revised to include these specific values and to define 'competitive' by reference to the quantitative margins shown in the comparison table. revision: yes

  2. Referee: [Abstract] Abstract: the reported 72.11% Dice / 58.44% IoU on ML-SASD-Mix is presented without any baseline methods, ablation results on the dual-stream fusion / multi-expert decoder / PCP Loss / CA-ASW components, or statistical significance tests. These omissions are load-bearing because the soundness of the performance claim cannot be assessed without them.

    Authors: The abstract is a concise summary; the manuscript contains baseline comparisons (Section 4), component-wise ablations on dual-stream fusion, multi-expert decoder, PCP Loss and CA-ASW (Section 5 and tables), and we can incorporate statistical significance where appropriate. We will revise the abstract to note the performance margins over baselines and the ablation-validated contributions of each proposed element. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical evaluation on new dataset with independent metrics

full rationale

The paper proposes HD-DinoMoE architecture and constructs the ML-SASD dataset, then reports Dice/IoU metrics on its Mix split and a qualitative generalization note on SBVPI. No equations, self-definitions, or self-citations reduce any claimed result to a fitted input by construction; performance numbers are direct outputs of evaluation on held-out annotations. The absence of SBVPI numbers is an evidence gap, not circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 2 invented entities

Central claim depends on the effectiveness of several newly introduced components whose hyperparameters and design choices are not detailed in the abstract; evaluation also rests on the quality of the newly created dataset annotations.

free parameters (2)
  • Routing and expert count in dual MoE decoder
    The class-specific multi-expert decoding and three-stage routing strategy require choices for number of experts and routing thresholds that are tuned to the training data.
  • Weighting coefficients in PCP Loss and CA-ASW
    The Progressive Confidence Penalty and Class-Aware Adaptive Sample Weighting introduce scalar weights that control the strength of the penalties and re-weighting.
axioms (2)
  • domain assumption Pre-trained DINOv3 features remain sufficiently invariant under the distributional shifts present in clinical and wild scleral images.
    The dual-stream fusion step assumes the backbone provides useful representations without further fine-tuning of the feature extractor.
  • domain assumption Pixel-wise ground-truth labels in ML-SASD accurately capture the three anomaly categories across all acquisition settings.
    All quantitative claims rest on the correctness of the newly collected annotations.
invented entities (2)
  • Progressive Confidence Penalty (PCP) Loss no independent evidence
    purpose: Penalize high-confidence false positives and reduce leakage in specular reflection regions during training.
    New loss term introduced to address SSR-specific failure modes.
  • Class-Aware Adaptive Sample Weighting (CA-ASW) no independent evidence
    purpose: Dynamically balance contributions across samples and anomaly classes during optimization.
    New weighting mechanism to handle class imbalance in the multi-label setting.

pith-pipeline@v0.9.1-grok · 5896 in / 1750 out tokens · 54125 ms · 2026-06-28T06:46:23.529446+00:00 · methodology

0 comments
read the original abstract

Traditional Chinese Medicine (TCM) ocular inspection provides empirical cues for assessing scleral surface anomalies, but its clinical use remains subjective and difficult to quantify. To support intelligent and quantifiable ocular inspection, this study presents the TCM-inspired Artificial Intelligence Ocular Auxiliary Diagnosis System (TAO) and focuses on pixel-level scleral surface anomaly segmentation. For clinical and user-acquired images affected by multi-source distributional discrepancies, diverse anomaly morphologies, and scleral specular reflection (SSR), we propose HD-DinoMoE, a class-aware hierarchical dual mixture-of-experts network. HD-DinoMoE combines class-aware dual-stream DINOv3 feature fusion with class-specific multi-expert decoding to segment Vessels, Yellow and Black Spots, and Blood Spots. A three-stage backbone-frozen routing strategy stabilizes dual-backbone adaptation; Progressive Confidence Penalty (PCP) Loss reduces high-confidence false positives and segmentation leakage in SSR regions; and Class-Aware Adaptive Sample Weighting (CA-ASW) balances sample- and class-level training contributions. We further construct the Multi-label Scleral Anomaly Segmentation Dataset (ML-SASD), a new benchmark with Clinical, Wild, and Mix settings and pixel-wise annotations for three anomaly categories. On ML-SASD-Mix, HD-DinoMoE achieves a mean Dice of 72.11% and a mean Intersection-over-Union of 58.44%, while maintaining favorable boundary localization and specular-region false-positive control. It also shows competitive generalization on the Vessels subset of the public SBVPI dataset. These results indicate that HD-DinoMoE provides a feasible segmentation solution for TAO under complex acquisition scenarios. The code and data access information are available at https://github.com/FX-CMX/HD-DinoMoE.

Figures

Figures reproduced from arXiv: 2606.04888 by Guanghao Liao, Guanghu Liu, Haotian Wang, Maoxiang Chu, Qi Niu, Wei Xu, Yinxiang Yu, Yuelong Fan, Yutian Zhu, Zhi Chen.

Figure 1
Figure 1. Figure 1: Functional overview of the TCM-inspired AI Ocular Auxiliary Diagnosis System However, translating the TAO system into real-world deployment faces multifaceted challenges spanning both data and algorithmic dimensions: 1) Extreme scarcity of pathology-oriented datasets: Although several ocular surface and sclera segmentation datasets have been publicly released(Vitek et al., 2020b; Proenca et al., 2010; De M… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the eye image acquisition protocol. (a) and (b) show the Five-eye Positioning Method for exposing different scleral regions, and (c) shows the schematic illustration of eye image acquisition. To improve the coverage of complex acquisition conditions in ML-SASD, we constructed a multi-source data acquisition framework consisting of medical atlases and images captured using commercial mobile … view at source ↗
Figure 3
Figure 3. Figure 3: Illustrations from Traditional Chinese Medicine Eye Diagnosis (3rd Edition) (Zheng and Zheng, 2016) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative image examples from the ML-SASD-Wild subset (1) ML-SASD-Clinical (controlled clinical subset): This subset was selected from the medical atlas Illustrated TCM Ocular Inspection and Syndrome Differentiation(Zheng and Zheng, 2016), which consists of ocular images with diverse morphologies captured by Dr. Zheng using a digital camera during years of clinical practice. The acquisition protocol f… view at source ↗
Figure 5
Figure 5. Figure 5: Representative annotated sclera examples from the ML-SASD dataset (First-stage Annotated) The annotation protocols of ML-SASD were designed in accordance with the cascaded semantic segmentation strategy in the TAO system, resulting in two annotation sets: First-stage Annotated Images and Second-stage Annotated Images. For First-stage Annotated Images, the task objective is to segment the sclera from ocular… view at source ↗
Figure 6
Figure 6. Figure 6: Representative examples of Ve and YBS from the Second-stage Annotated subset. Left: various examples of Ve. Right: various examples of YBS. The black dashed contours and arrows in the YBS examples are used only to highlight typical locations and morphological characteristics and do not represent precise annotation boundaries. (a) (b) (c) (d) (e) (f) (g) (h) Variou spots (a) (b) (c) (d) (e) (f) (g) (h) [PI… view at source ↗
Figure 7
Figure 7. Figure 7 [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Examples of regional overlap among scleral anomaly categories (Second-stage Annotated). (a) Regional overlap between BS and YBS, where blue and green highlighted regions indicate the BS and YBS annotations, respectively. (b) Regional overlap between Ve and BS, where red and blue highlighted regions indicate the Ve and BS annotations, respectively. In both panels, “All label” shows the combined visualizatio… view at source ↗
Figure 9
Figure 9. Figure 9: Comprehensive visualization of scleral anomaly annotations in ML-SASD-Mix (Second-stage Annotated): the “Original” panel shows the raw input image, while “Class 1”, “Class 2”, and “Class 3” correspond to the ground-truth masks of Ve, BS, and YBS, respectively. To improve annotation consistency, all masks in ML-SASD were produced following a unified annotation protocol. The annotations were independently pr… view at source ↗
Figure 10
Figure 10. Figure 10: Examples of SSR (Second-stage Annotated): The first column shows the original images, and the second column highlights the SSR regions on the original images. The black dashed contours and arrows are used only to highlight the typical locations and morphological characteristics of SSR and do not represent precise annotation boundaries. 3.4. Dataset Distribution Analysis To systematically characterize the … view at source ↗
Figure 11
Figure 11. Figure 11: Image-level sample distribution statistics [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Pixel-level class distribution across Clinical-Wild-Mix datasets:The left chart displays the absolute number of labeled pixels on a logarithmic scale for each segmentation class. The right chart illustrates the relative pixel ratio (%) of each class within the respective datasets. background. Without targeted intervention during training, the model may suffer from background-biased prediction collapse. Wi… view at source ↗
Figure 13
Figure 13. Figure 13: Image Resolution Distribution As shown in [PITH_FULL_IMAGE:figures/full_fig_p012_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Per-Class Lesion Area Distribution As shown in [PITH_FULL_IMAGE:figures/full_fig_p013_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Class co-occurrence matrix As summarized in [PITH_FULL_IMAGE:figures/full_fig_p013_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Maximum Mean Discrepancy matrix [PITH_FULL_IMAGE:figures/full_fig_p014_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Technical architecture of the TCM-inspired AI Ocular Auxiliary Diagnosis System segmentation is critical to the reliability of downstream regional scoring and auxiliary analysis. Therefore, this study focuses on the second-phase scleral anomaly segmentation task, and HD-DinoMoE is designed around this task. In the third component, TAO establishes a spatial mapping relationship based on the segmented scler… view at source ↗
Figure 18
Figure 18. Figure 18: HD-DinoMoE Overall architecture HD-DinoMoE is designed for the second-phase scleral surface anomaly segmentation task in the TAO system. Given a pre-extracted scleral region of interest (ROI), the model predicts three anomaly categories, namely Ve, YBS, and BS, in a multi-label manner. As shown in [PITH_FULL_IMAGE:figures/full_fig_p016_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Architecture diagram of CA-DSGF (GAP), as shown in [PITH_FULL_IMAGE:figures/full_fig_p018_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Design of the proposed TS-BFRA strategy: The snowflake icon denotes frozen parameters, while the flame icon denotes trainable parameters. Given that CA-DSGF contains two large-scale DINOv3-L encoders, conventional end-to-end joint optimization is highly susceptible to systematic representation collapse induced by cross-feature gradient interference, especially before the two foundation branches have adapt… view at source ↗
Figure 21
Figure 21. Figure 21: Architecture diagram of CS-MED Yinxiang Yu et al.: Preprint submitted to Elsevier Page 20 of 47 [PITH_FULL_IMAGE:figures/full_fig_p020_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: CA-ASW Loss Overall Process The input to the CA-ASW module is 𝑙 𝑖,𝑐 , which is derived from the loss functions Base and PCP in Section 4.6. When PCP Loss is not used, the input is defined as: 𝑙 𝑖,𝑐 =  (𝑖,𝑐) Base (20) When PCP Loss is used, the input is defined as: Yinxiang Yu et al.: Preprint submitted to Elsevier Page 25 of 47 [PITH_FULL_IMAGE:figures/full_fig_p025_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Radar chart comparison of the best-performing models from each group As shown in [PITH_FULL_IMAGE:figures/full_fig_p031_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Visual comparison of representative segmentation methods on ML-SASD-Mix underscores that, for multi-source heterogeneous scleral anomaly segmentation, large-scale visual pretrained features can provide more effective and transferable foundational representations. Among the DINOv3-related baselines, SegDINO-SAT and SegDINO-LVD achieve 70.44% and 70.06% mDice, respectively, clearly outperforming most non-fo… view at source ↗
Figure 26
Figure 26. Figure 26: Dice coefficient curves under different training strategies the validation Dice reaches a premature peak and subsequently plateaus, without improving synchronously with the training performance. This indicates that although end-to-end joint optimization can continue to fit the training set, its improvement on validation performance is limited, suggesting a certain tendency toward overfitting. One possible… view at source ↗
Figure 27
Figure 27. Figure 27: Visualization of expert feature activations and class-specific gating weights in CS-MED as a complementary expert. For the representative Ve structures in the figure, the sub-pixel reconstruction-based D2S expert shows strong high-frequency detail-capturing capability, with the heatmap focusing on sharp and slender Ve boundaries but appearing relatively fragmented. In contrast, the DPT expert provides nec… view at source ↗
Figure 28
Figure 28. Figure 28: Effect of PCP Loss parameters on mIoU and mGFPR As shown in [PITH_FULL_IMAGE:figures/full_fig_p038_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: Visual comparison of the effect of PCP Loss on suppressing false positives in reflective regions 5.5.5. Ablation Study of CA-ASW To validate the effectiveness of the adaptive sample-class weighting scheme in CA-ASW, this study compares five weighting strategies on ML-SASD-Mix, namely Easy, Hard, Focal, Curriculum, and Balanced, and further conducts ablation experiments on the related hyperparameters. The … view at source ↗
Figure 30
Figure 30. Figure 30: Radar chart comparison of the best-performing models from each group SegNeXt ConDSeg SegDINO nnU-Net U-Mamba HD-DinoMoE GT Original [PITH_FULL_IMAGE:figures/full_fig_p043_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: Comparison of prediction heatmaps on the SBVPI Vessels dataset For the boundary metric BF1, HD-DinoMoE shows a more evident advantage, reaching 53.09%, exceeding the second-best DINOUnet-LVD by 5.51 percentage points. This indicates that HD-DinoMoE provides better boundary localization capability for Vessels structures in cross-dataset testing. Considering the domain discrepancies between SBVPI and ML-SAS… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

14 extracted references · 3 canonical work pages · 2 internal anchors

  1. [1]

    Zhongguo Zhen Jiu = Chinese Acupuncture & Moxibustion 25, 287–291

    Formation, development and changes of the eye acupunture region-dividing acupoint location program. Zhongguo Zhen Jiu = Chinese Acupuncture & Moxibustion 25, 287–291. Chen, J., Mei, J., Li, X., Lu, Y., Yu, Q., Wei, Q., Luo, X., Xie, Y., Adeli, E., Wang, Y., Lungren, M.P., Zhang, S., Xing, L., Lu, L., Yuille, A., Zhou, Y.,2024. TransUNet:RethinkingtheU-net...

    2024

  2. [2]

    Das,A.,Pal,U.,FerrerBallester,M.A.,Blumenstein,M.,2014

    Towards multi-modal sclera and iris biometric recognition with adaptive liveness detection. Das,A.,Pal,U.,FerrerBallester,M.A.,Blumenstein,M.,2014. Multi-anglebasedlivelysclerabiometricsatadistance,in:2014IEEESymposium on Computational Intelligence in Biometrics and Identity Management (CIBIM), IEEE, Orlando, FL. pp. 22–29. De Marsico, M., Nappi, M., Ricc...

    2014

  3. [3]

    Dino U-Net: Exploiting High-Fidelity Dense Features from Foundation Models for Medical Image Segmentation

    Gao, Y., Li, H., Yuan, F., Wang, X., Gao, X., 2025b. Dino U-Net: Exploiting High-Fidelity Dense Features from Foundation Models for Medical Image Segmentation.arXiv:2508.20909. Yinxiang Yu et al.:Preprint submitted to ElsevierPage 45 of 47 Guan,H.,Liu,M.,2022. Domainadaptationformedicalimageanalysis:Asurvey. IEEETransactionsonBio-MedicalEngineering69,1173...

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  4. [4]

    nnU-Net:Aself-configuringmethodfordeeplearning-basedbiomedical image segmentation

    Isensee,F.,Jaeger,P.F.,Kohl,S.A.A.,Petersen,J.,Maier-Hein,K.H.,2021. nnU-Net:Aself-configuringmethodfordeeplearning-basedbiomedical image segmentation. Nature Methods 18, 203–211. Jacobs, R.A., Jordan, M.I., Nowlan, S.J., Hinton, G.E.,

    2021

  5. [5]

    Kirillov,A.,Mintun,E.,Ravi,N.,Mao,H.,Rolland,C.,Gustafson,L.,Xiao,T.,Whitehead,S.,Berg,A.C.,Lo,W.Y.,Dollar,P.,Girshick,R.,2023

    Mixtral of experts. Kirillov,A.,Mintun,E.,Ravi,N.,Mao,H.,Rolland,C.,Gustafson,L.,Xiao,T.,Whitehead,S.,Berg,A.C.,Lo,W.Y.,Dollar,P.,Girshick,R.,2023. Segment anything, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 4015–4026. Lee,H.J.,Kim,J.U.,Lee,S.,Kim,H.G.,Ro,Y.M.,2020. Structureboundarypreservingsegmentationformedicalim...

    2023

  6. [6]

    BritishJournalofOphthalmology , bjo–2024–325174

    Impact of artificial intelligence on the availability,accessibility,acceptabilityandqualityofophthalmicdiseasescreeningservices:Ascopingreview. BritishJournalofOphthalmology , bjo–2024–325174. Liu, C., Chen, Y., Shi, H., Lu, J., Jian, B., Pan, J., Cai, L., Wang, J., Yu, J., Gao, Z., Zhang, X., Bai, L., Zhang, Y., Li, J., Bercea, C.I., Ouyang, C., Chen,C.,...

    2024

  7. [7]

    Ma,X.,Xu,K.,Jiang,J.,Liu,R.,Yu,X.,2018.Layeredvasculaturesegmentationofcolorconjunctivalimagebasedonwavelettransform.Biomedical Signal Processing and Control 42, 9–17

    U-mamba: Enhancing long-range dependency for biomedical image segmentation. Ma,X.,Xu,K.,Jiang,J.,Liu,R.,Yu,X.,2018.Layeredvasculaturesegmentationofcolorconjunctivalimagebasedonwavelettransform.Biomedical Signal Processing and Control 42, 9–17. Maquiling,V.,Byrne,S.A.,Niehorster,D.C.,Nyström,M.,Kasneci,E.,2024. Zero-ShotSegmentationofEyeFeaturesUsingtheSeg...

    2018

  8. [8]

    Training language models to follow instructions withhumanfeedback,in:Koyejo,S.,Mohamed,S.,Agarwal,A.,Belgrave,D.,Cho,K.,Oh,A.(Eds.),AdvancesinNeuralInformationProcessing Systems, Curran Associates, Inc.. pp. 27730–27744. Perazzi,F.,Pont-Tuset,J.,McWilliams,B.,VanGool,L.,Gross,M.,Sorkine-Hornung,A.,2016. Abenchmarkdatasetandevaluationmethodology for video ...

    2016

  9. [9]

    Scientific Reports 15, 44142

    Non-invasive anemia detection from conjunctiva and sclera images using vision transformer with attention map explainability. Scientific Reports 15, 44142. Ronneberger,O.,Fischer,P.,Brox,T.,2015. U-net:Convolutionalnetworksforbiomedicalimagesegmentation,in:Navab,N.,Hornegger,J.,Wells, W.M., Frangi, A.F. (Eds.), Medical Image Computing and Computer-Assisted...

    2015

  10. [10]

    DINOv3.arXiv:2508.10104. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  11. [11]

    A new strategy for retinal vessel segmentation based on semi-dilation supervised learning and boundary reconstruction, in: 2024 7th International Conference on Pattern Recognition and Artificial Intelligence (PRAI), IEEE, Hangzhou, China. pp. 775–780. Wen,T.,Fu,X.,Evandrew,V.Y.,Yu,X.,Wang,M.,Chen,J.,Tang,L.,2026. Advancesinglaucomabiomechanicsfrom2000to20...

    2024

  12. [12]

    Exploiting DINOv3-based self-supervised features for robust few-shot medical image segmentation. Yan,W.,Wang,Y.,Gu,S.,Huang,L.,Yan,F.,Xia,L.,Tao,Q.,2019.Thedomainshiftproblemofmedicalimagesegmentationandvendor-adaptation by unet-GAN, in: Shen, D., Liu, T., Peters, T.M., Staib, L.H., Essert, C., Zhou, S., Yap, P.T., Khan, A. (Eds.), Medical Image Computing...

    work page arXiv 2019

  13. [13]

    Current Opinion in Ophthalmology 36, 477–486

    Artificial intelligence oculomics for systemic health and longevity medicine: 2025 and beyond. Current Opinion in Ophthalmology 36, 477–486. Yu, Z., Du, C., Liang, H., Zheng, X., Ma, Z., Wu, M., Ao, M., Lao, Q.,

    2025

  14. [14]

    (Eds.), Medical Image Computing and Computer Assisted Intervention – MICCAI 2025, Springer Nature Switzerland, Cham

    Endoscopic artifact inpainting for improved endoscopic image segmentation, in: Gee, J.C., Alexander, D.C., Hong, J., Iglesias, J.E., Sudre, C.H., Venkataraman, A., Golland, P., Kim, J.H., Park, J. (Eds.), Medical Image Computing and Computer Assisted Intervention – MICCAI 2025, Springer Nature Switzerland, Cham. pp. 191–201. Zheng, D., Zheng, Z.,

    2025