pith. sign in

arxiv: 2604.05388 · v1 · submitted 2026-04-07 · 💻 cs.CV

LUMOS: Universal Semi-Supervised OCT Retinal Layer Segmentation with Hierarchical Reliable Mutual Learning

Yizhou Fang , Jian Zhong , Li Lin , Xiaoying Tang This is my paper

Pith reviewed 2026-05-10 19:43 UTC · model grok-4.3

classification 💻 cs.CV
keywords semi-supervised learningOCT segmentationretinal layersmulti-granularitypseudo labelscross-domain generalizationhierarchical promptingprogressive training
0
0 comments X

The pith

LUMOS uses dual decoders and progressive learning to segment OCT retinal layers semi-supervisedly across varying label granularities.

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

The paper tackles annotation scarcity and inconsistent label details in OCT retinal layer segmentation by proposing a universal semi-supervised framework. It introduces a dual-decoder network with hierarchical prompting to limit noise from pseudo labels and a reliable progressive multi-granularity learning module that weights reliable regions and trains from easier to harder tasks. This combination allows stable alignment between different granularity levels of supervision from both labeled and unlabeled data. Experiments across six datasets show improved accuracy and better generalization to new imaging domains compared to prior techniques.

Core claim

LUMOS establishes that a Dual-Decoder Network with Hierarchical Prompting Strategy combined with Reliable Progressive Multi-granularity Learning achieves stable cross-granularity consistency in semi-supervised OCT segmentation, outperforming existing methods while generalizing well across domains and label granularities.

What carries the argument

The Dual-Decoder Network with Hierarchical Prompting Strategy (DDN-HPS) that suppresses pseudo-label noise propagation and the Reliable Progressive Multi-granularity Learning (RPML) that uses region-level reliability weighing to select consistency targets progressively from coarse to fine tasks.

If this is right

  • The approach makes better use of unlabeled OCT scans for layer segmentation despite limited annotations.
  • It enables a single model to handle datasets with both coarse and detailed layer labels without retraining.
  • Cross-domain application becomes more feasible for different OCT scanners and patient groups.
  • Annotation requirements for new datasets can be reduced by leveraging consistency across granularities.
  • Performance gains are realized through progressive training that minimizes error accumulation in pseudo labels.

Where Pith is reading between the lines

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

  • This strategy may extend to other segmentation problems in medical imaging where label detail varies between sources.
  • Public OCT datasets could be combined more easily for training without harmonizing their annotation styles first.
  • Clinical deployment might benefit from the method's robustness to domain shifts in real-world OCT acquisitions.
  • Further gains could come from integrating additional reliability metrics beyond region-level weighting.

Load-bearing premise

The assumption that region-level reliability weighing combined with progressive training will ensure stable selection of cross-granularity consistency targets and suppress pseudo-label noise without introducing systematic biases or requiring dataset-specific tuning.

What would settle it

An experiment on a seventh OCT dataset where the reliability-based target selection shows no higher agreement with ground truth than uniform selection, and overall segmentation accuracy does not exceed that of single-granularity semi-supervised baselines.

read the original abstract

Optical Coherence Tomography (OCT) layer segmentation faces challenges due to annotation scarcity and heterogeneous label granularities across datasets. While semi-supervised learning helps alleviate label scarcity, existing methods typically assume a fixed granularity, failing to fully exploit cross-granularity supervision. This paper presents LUMOS, a semi-supervised universal OCT retinal layer segmentation framework based on a Dual-Decoder Network with a Hierarchical Prompting Strategy (DDN-HPS) and Reliable Progressive Multi-granularity Learning (RPML). DDN-HPS combines a dual-branch architecture with a multi-granularity prompting strategy to effectively suppress pseudo-label noise propagation. Meanwhile, RPML introduces region-level reliability weighing and a progressive training approach that guides the model from easier to more difficult tasks, ensuring the reliable selection of cross-granularity consistency targets, thereby achieving stable cross-granularity alignment. Experiments on six OCT datasets demonstrate that LUMOS largely outperforms existing methods and exhibits exceptional cross-domain and cross-granularity generalization capability.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents LUMOS, a semi-supervised universal framework for OCT retinal layer segmentation. It introduces a Dual-Decoder Network with Hierarchical Prompting Strategy (DDN-HPS) to suppress pseudo-label noise propagation and Reliable Progressive Multi-granularity Learning (RPML) that combines region-level reliability weighing with progressive training to enable stable cross-granularity consistency targets. Experiments on six OCT datasets are reported to show large outperformance over existing methods along with strong cross-domain and cross-granularity generalization.

Significance. If the empirical results and robustness claims hold, the work would advance semi-supervised medical image segmentation by demonstrating how heterogeneous label granularities can be exploited without dataset-specific tuning. This addresses a practical bottleneck in OCT analysis where annotations vary in detail across studies, potentially lowering annotation costs while improving model deployment across domains.

major comments (2)
  1. [§3.2] §3.2 (RPML formulation): The region-level reliability weighing is presented as ensuring stable selection of cross-granularity targets, yet no derivation, sensitivity analysis, or ablation isolates its behavior when initial pseudo-labels contain domain-shift noise. This is load-bearing for the abstract's claim of exceptional cross-domain generalization, as early errors could be amplified rather than suppressed.
  2. [Table 2] Table 2 (cross-dataset results) and §4.3: The reported outperformance lacks an ablation that removes the reliability weighing component while keeping progressive training and DDN-HPS fixed. Without this, it is impossible to confirm that RPML, rather than the dual-decoder architecture alone, drives the cross-granularity gains.
minor comments (2)
  1. [§3.1] The hierarchical prompting strategy in DDN-HPS would benefit from an explicit equation defining the prompt embedding concatenation to improve reproducibility.
  2. [Figure 4] Figure 4 caption should explicitly state the number of runs and standard deviation for the reported Dice scores to clarify statistical significance.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback on our manuscript. We address each major comment below and have revised the manuscript to incorporate the requested analyses and ablations.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (RPML formulation): The region-level reliability weighing is presented as ensuring stable selection of cross-granularity targets, yet no derivation, sensitivity analysis, or ablation isolates its behavior when initial pseudo-labels contain domain-shift noise. This is load-bearing for the abstract's claim of exceptional cross-domain generalization, as early errors could be amplified rather than suppressed.

    Authors: We agree that the manuscript would benefit from an explicit derivation and sensitivity analysis of the region-level reliability weighing under domain-shift noise. The RPML formulation computes per-region reliability weights from cross-decoder prediction consistency to down-weight unreliable pseudo-label regions, combined with a progressive schedule that begins with coarser, more stable targets before introducing finer granularity. This design is intended to limit early error propagation, but we acknowledge the absence of an isolated study. In the revised version we will add a brief derivation of the reliability weight in §3.2 and include a sensitivity analysis (with controlled noise injection on initial pseudo-labels) in the supplementary material. revision: yes

  2. Referee: [Table 2] Table 2 (cross-dataset results) and §4.3: The reported outperformance lacks an ablation that removes the reliability weighing component while keeping progressive training and DDN-HPS fixed. Without this, it is impossible to confirm that RPML, rather than the dual-decoder architecture alone, drives the cross-granularity gains.

    Authors: We concur that an ablation isolating the reliability-weighing term—while retaining progressive training and DDN-HPS—is required to attribute the cross-granularity gains specifically to this component of RPML. The current experiments ablate the full RPML module and DDN-HPS separately, but do not perform the requested partial removal. We will run this ablation and report the results in the revised §4.3, updating the experimental discussion and adding the numbers to Table 2 or a new supplementary table. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical validation of proposed architecture and training strategy

full rationale

The paper introduces DDN-HPS and RPML as novel components for semi-supervised OCT segmentation and supports its claims solely through experimental results on six datasets. No equations, derivations, or self-citations are presented that reduce any 'prediction' or result to the method's own inputs by construction. The reliability weighing and progressive training are framed as design choices whose benefits are measured externally rather than defined into existence. This is a standard empirical ML contribution with independent content.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review prevents exhaustive identification; the method implicitly relies on standard semi-supervised assumptions such as consistency between model predictions at different granularities and the existence of reliable pseudo-label regions.

axioms (1)
  • domain assumption Pseudo-labels generated at one granularity level can provide useful supervision signals for another granularity level when filtered by region reliability
    This underpins the RPML component for cross-granularity alignment and is invoked to justify stable training.

pith-pipeline@v0.9.0 · 5475 in / 1164 out tokens · 61742 ms · 2026-05-10T19:43:45.917396+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages · 1 internal anchor

  1. [1]

    INTRODUCTION Optical Coherence Tomography (OCT) provides high-resolution cross-sectional images that are crucial for diagnosing oph- thalmic diseases like diabetic retinopathy and glaucoma [1]. Accurate retinal layer segmentation [2] is fundamental to clinical diagnosis but faces significant challenges: manual segmentation is time-consuming with high inte...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  2. [2]

    METHOD 2.1. LUMOS Paradigm Overview Conventional semi-supervised methods typically address a single annotation granularity, trainingNseparate models for Ntasks involving different annotation granularities. This hinders label reuse across granularities and inflates com- putation and storage, ultimately limiting performance and generalization. LUMOS unifies...

    work page

  3. [3]

    EXPERIMENTS 3.1. Dataset and Preprocessing We assemble multiple multi-granularity OCT layer segmenta- tion datasets acquired across different devices as the internal sets: HC-MS [9], GCN [10], and OCTA-500 [11], containing 3,430, 482, and 9,416 samples with 8, 8, and 5 granularity levels, respectively. HC-MS [9] and GCN [10] are split into training, valid...

    work page 2017

  4. [4]

    Experiments on six datasets demonstrate that LUMOS and its components significantly outperform existing methods in cross-domain and cross-granularity adaptation

    CONCLUSION This study proposes LUMOS, a semi-supervised universal framework for multi-granularity OCT retinal layer segmenta- tion. Experiments on six datasets demonstrate that LUMOS and its components significantly outperform existing methods in cross-domain and cross-granularity adaptation. However, limitations persist regarding handling severe patholog...

    work page

  5. [5]

    ACKNOWLEDGMENTS This study was supported by the National Key Research and Development Program of China (2023YFC2415400); the National Natural Science Foundation of China (T2422012); the Guangdong Basic and Applied Basic Research (2024B1515020088); the High Level of Special Funds (G030230001, G03034K003); the Guangdong Key Research and Development Program ...

    work page

  6. [6]

    Ethical approval was not required as confirmed by the open data usage terms gov- erning these datasets

    COMPLIANCE WITH ETHICAL STANDARDS This research study was conducted retrospectively using pub- licly available and anonymized datasets (OCTA-500, HC-MS, GCN, AMD, HEG, GOALS, OIMHS). Ethical approval was not required as confirmed by the open data usage terms gov- erning these datasets

    work page

  7. [7]

    Optical coherence tomography in coronary atherosclerosis assessment and intervention,

    Araki Makoto et al., “Optical coherence tomography in coronary atherosclerosis assessment and intervention,” Nature Reviews Cardiology, vol. 19, no. 10, pp. 684– 703, 2022

    work page 2022

  8. [8]

    Retinal layer segmentation in oct im- ages with boundary regression and feature polarization,

    Tan Yubo et al., “Retinal layer segmentation in oct im- ages with boundary regression and feature polarization,” IEEE Transactions on Medical Imaging, vol. 43, no. 2, pp. 686–700, 2023

    work page 2023

  9. [9]

    Exploring publicly accessi- ble optical coherence tomography datasets: a compre- hensive overview,

    Rozhyna Anastasiia et al., “Exploring publicly accessi- ble optical coherence tomography datasets: a compre- hensive overview,”Diagnostics, vol. 14, no. 15, pp. 1668, 2024

    work page 2024

  10. [10]

    Diffrect: Latent diffusion label rectification for semi-supervised medical image segmentation,

    Xinyu Liu, Wuyang Li, and Yixuan Yuan, “Diffrect: Latent diffusion label rectification for semi-supervised medical image segmentation,” inInternational Con- ference on Medical Image Computing and Computer- Assisted Intervention. Springer, 2024, pp. 56–66

    work page 2024

  11. [11]

    Balancing multi-target semi- supervised medical image segmentation with collabora- tive generalist and specialists,

    Wang You et al., “Balancing multi-target semi- supervised medical image segmentation with collabora- tive generalist and specialists,”IEEE Transactions on Medical Imaging, 2025

    work page 2025

  12. [12]

    Adaptive bidirectional displace- ment for semi-supervised medical image segmentation,

    Chi Hanyang et al., “Adaptive bidirectional displace- ment for semi-supervised medical image segmentation,” inProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2024, pp. 4070– 4080

    work page 2024

  13. [13]

    Masked-attention mask transformer for universal image segmentation,

    Cheng Bowe et al., “Masked-attention mask transformer for universal image segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1290–1299

    work page 2022

  14. [14]

    Unioctseg: Towards universal oct retinal layer segmentation via hierarchical prompt- ing and progressive consistency learning,

    Zhong Jian et al., “Unioctseg: Towards universal oct retinal layer segmentation via hierarchical prompt- ing and progressive consistency learning,” inInterna- tional Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025, pp. 629–639

    work page 2025

  15. [15]

    Retinal layer parcellation of optical coherence tomography images: Data resource for mul- tiple sclerosis and healthy controls,

    He Yufan et al., “Retinal layer parcellation of optical coherence tomography images: Data resource for mul- tiple sclerosis and healthy controls,”Data in Brief, vol. 22, pp. 601, 2018

    work page 2018

  16. [16]

    Multi-scale gcn-assisted two-stage network for joint segmentation of retinal layers and discs in peripapillary oct images,

    Li Jiaxuan et al., “Multi-scale gcn-assisted two-stage network for joint segmentation of retinal layers and discs in peripapillary oct images,”Biomedical Optics Express, vol. 12, no. 4, pp. 2204–2220, 2021

    work page 2021

  17. [17]

    Octa-500: a retinal dataset for opti- cal coherence tomography angiography study,

    Li Mingchao et al., “Octa-500: a retinal dataset for opti- cal coherence tomography angiography study,”Medical Image Analysis, vol. 93, pp. 103092, 2024

    work page 2024

  18. [18]

    Real-time automatic segmentation of optical coherence tomography volume data of the mac- ular region,

    Tian Jing et al., “Real-time automatic segmentation of optical coherence tomography volume data of the mac- ular region,”PLoS ONE, vol. 10, no. 8, pp. e0133908, 2015

    work page 2015

  19. [19]

    Dataset and evaluation algorithm design for goals challenge,

    Fang Huihui et al., “Dataset and evaluation algorithm design for goals challenge,” inInternational Workshop on Ophthalmic Medical Image Analysis. Springer, 2022, pp. 135–142

    work page 2022

  20. [20]

    Validated automatic segmenta- tion of amd pathology including drusen and geographic atrophy in sd-oct images,

    Chiu Stephanie J et al., “Validated automatic segmenta- tion of amd pathology including drusen and geographic atrophy in sd-oct images,”Investigative Ophthalmology & Visual Science, vol. 53, no. 1, pp. 53–61, 2012

    work page 2012

  21. [21]

    Oimhs: An optical coherence tomogra- phy image dataset based on macular hole manual seg- mentation,

    Ye Xin et al., “Oimhs: An optical coherence tomogra- phy image dataset based on macular hole manual seg- mentation,”Scientific Data, vol. 10, no. 1, pp. 769, 2023

    work page 2023

  22. [22]

    Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results,

    Antti Tarvainen and Harri Valpola, “Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results,”Ad- vances in Neural Information Processing Systems, vol. 30, 2017

    work page 2017

  23. [23]

    Semi-supervised semantic seg- mentation with cross-consistency training,

    Ouali Yassine et al., “Semi-supervised semantic seg- mentation with cross-consistency training,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 12674–12684

    work page 2020

  24. [24]

    Advent: Adversarial entropy min- imization for domain adaptation in semantic segmen- tation,

    Vu Tuan-Hung et al., “Advent: Adversarial entropy min- imization for domain adaptation in semantic segmen- tation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 2517–2526

    work page 2019