pith. sign in

arxiv: 2606.27794 · v1 · pith:ZLYDB6C3new · submitted 2026-06-26 · 💻 cs.CV

Text as Illumination: Spatial Contrastive Retinex Learning for Language-guided Medical Image Segmentation

Jian Shi , Cheng Zhen , Pingping Zhang , Rui Xu , Yanan Lv , Yili Ma , Huan Bi , Haojie Li
show 1 more author
Huchuan Lu
This is my paper

Pith reviewed 2026-06-29 04:57 UTC · model grok-4.3

classification 💻 cs.CV
keywords language-guided medical image segmentationRetinex modeltext modulationcontrastive learningsemantic consistencyfeature modulationmulti-scale supervision
0
0 comments X

The pith

Treating text embeddings as semantic illumination in a Retinex-inspired network improves semantic consistency in language-guided medical image segmentation.

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

The paper establishes that explicit fine-grained constraints from text can be provided by modeling text embeddings as illumination maps to modulate visual features. This Retinex-inspired approach uses positive and negative maps to enhance relevant regions and suppress others, combined with a multi-scale supervision loss that enforces cross-modal similarity in foreground areas. Existing methods suffer from mismatch due to implicit interactions, so this aims to produce segmentation outputs that better align with clinical text descriptions. Sympathetic readers would care because it could lead to more reliable automated delineation of lesions and structures using both image and text data.

Core claim

The central claim is that by treating text embeddings as semantic illumination for feature modulation via the Retinex-inspired Text Modulation Block and Consistent Detail Compensation Block, along with the Multi-Scale Illumination Supervision Loss, the TIRNet framework ensures precise cross-modal alignment and achieves state-of-the-art performance in language-guided medical image segmentation on the MosMedData+ and QaTa-COV19 datasets.

What carries the argument

The Retinex-inspired Text Modulation Block (RTMB), which employs positive and negative illumination maps to enhance text-relevant foreground features and suppress background interference.

If this is right

  • Semantic consistency is enforced at each decoder stage through illumination maps.
  • Region-Grounded Contrastive Loss concentrates cross-modal similarity in text-relevant regions.
  • Background Suppression Loss provides pixel-level supervision for negative maps.
  • High-frequency details are recovered via consistency-gated mechanism.
  • State-of-the-art results are demonstrated on two medical segmentation datasets.

Where Pith is reading between the lines

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

  • The modulation blocks could be tested for transfer to non-medical vision-language segmentation tasks.
  • Varying the number of decoder stages where the blocks are applied might reveal optimal placement for different image resolutions.
  • The approach may require additional checks when clinical text contains ambiguous or conflicting descriptions.

Load-bearing premise

The Retinex model can be directly adapted to text-visual feature interaction in medical images to ensure semantic consistency without introducing artifacts or requiring extensive hyperparameter tuning.

What would settle it

Performance that drops or introduces visible artifacts when the illumination maps are removed or when tested on datasets with mismatched text-image pairs would falsify the claim.

Figures

Figures reproduced from arXiv: 2606.27794 by Cheng Zhen, Haojie Li, Huan Bi, Huchuan Lu, Jian Shi, Pingping Zhang, Rui Xu, Yanan Lv, Yili Ma.

Figure 1
Figure 1. Figure 1: Limitations of existing methods. Both implicit interaction (LViT [12]) and coarse supervision (TeViA [26]) methods fail to effectively focus on foregrounds, while our TIRNet achieves accurate localization. To address the above issues, we present Text-as-Illumination Retinex Net￾work (TIRNet), a novel framework for LMIS. Specifically, TIRNet integrates two key blocks at each decoder stage. The Retinex-inspi… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the TIRNet. It integrates RTMB and CDCB into each decoder stage. The RGC-Loss maximizes cross-modal similarity in text-relevant foregrounds and suppresses background activations to enlarge the foreground-background margin. target regions that are semantically aligned with the text embedding while sup￾pressing irrelevant background responses, producing the modulated decoder fea￾ture F ′ dec with… view at source ↗
Figure 3
Figure 3. Figure 3: Visual comparison of segmentation results of different methods (Green: True Positive, Red: False Negative, Blue: False Positive). Following the data split protocol of LViT [12], we conduct extensive experiments on both datasets. To quantitatively assess performance, we report both sample￾level averages (m-Dice, m-IoU), and global metrics (g-Dice, g-IoU) aggregated over the entire dataset. Implementation De… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative visualization of the component ablation study. Panels (a)–(f) cor￾respond to the methods in the 1st–6th rows of [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
read the original abstract

Language-guided Medical Image Segmentation (LMIS) has shown great potential to improve the delineation of anatomical structures and lesions by integrating clinical textual information. Existing methods generally rely on either implicit interaction between textual and visual features or auxiliary coarse-grained supervision for cross-modal alignment. However, these methods lack explicit and fine-grained constraints to ensure semantic consistency, causing a mismatch between language and the segmentation outputs. To address this issue, we propose Text-as-Illumination Retinex Network (TIRNet), a novel Retinex-inspired framework that treats text embeddings as semantic illumination for feature modulation, thereby improving semantic consistency in LMIS. TIRNet introduces two key blocks integrated at each decoder stage: (1) the Retinex-inspired Text Modulation Block (RTMB), which employs positive and negative illumination maps to enhance text-relevant foreground features and suppress background interference; and (2) the Consistent Detail Compensation Block (CDCB), which selectively recovers high-frequency details via a consistency-gated mechanism conditioned on illumination reliability. Furthermore, we propose a Multi-Scale Illumination Supervision Loss (MSIS-Loss), comprising a Region-Grounded Contrastive Loss (RGC-Loss) that enforces cross-modal similarity to be concentrated in text-relevant foreground regions and suppressed in background regions, and a Background Suppression Loss (BS-Loss) that provides pixel-level supervision for negative illumination maps, jointly ensuring a precise cross-modal alignment at each decoder stage. Extensive experiments on the MosMedData+ and QaTa-COV19 datasets demonstrate that TIRNet achieves state-of-the-art performance in LMIS. The code is available at: https://github.com/anaanaa/TIRNet.

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

0 major / 3 minor

Summary. The manuscript proposes TIRNet, a Retinex-inspired framework for language-guided medical image segmentation (LMIS). Text embeddings are treated as semantic illumination to modulate features via the Retinex-inspired Text Modulation Block (RTMB) and Consistent Detail Compensation Block (CDCB) integrated at each decoder stage. A Multi-Scale Illumination Supervision Loss (MSIS-Loss) is introduced, consisting of Region-Grounded Contrastive Loss (RGC-Loss) and Background Suppression Loss (BS-Loss) to enforce cross-modal alignment. Experiments on MosMedData+ and QaTa-COV19 datasets claim state-of-the-art performance, with code released at https://github.com/anaanaa/TIRNet.

Significance. If the empirical SOTA results hold under rigorous validation, the work offers a concrete mechanism for explicit fine-grained text-visual alignment in LMIS via positive/negative illumination maps and region-grounded contrastive supervision. Credit is due for the public code release, which enables direct reproducibility of the reported performance on the two cited datasets.

minor comments (3)
  1. Abstract: the SOTA claim would be strengthened by explicit mention of the evaluation metrics (e.g., Dice, IoU), number of runs, and whether error bars or statistical tests accompany the reported gains over prior LMIS methods.
  2. The Retinex analogy is used to motivate the positive/negative illumination maps in RTMB and the consistency-gated recovery in CDCB; a brief discussion of how the medical-image domain differs from classical Retinex assumptions (e.g., illumination smoothness) would improve clarity without altering the central claim.
  3. Section describing MSIS-Loss: the interaction between RGC-Loss and BS-Loss at multiple decoder scales is central to the alignment argument; a single consolidated equation or pseudocode block would make the joint supervision easier to follow.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of our work on TIRNet and the recommendation for minor revision. The referee's summary correctly reflects the core contributions, including the Retinex-inspired design with RTMB, CDCB, and MSIS-Loss for explicit cross-modal alignment in LMIS. No major comments were provided in the report.

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained empirical construction

full rationale

The paper introduces TIRNet as a Retinex-inspired architecture with explicitly defined components (RTMB, CDCB, MSIS-Loss with RGC-Loss and BS-Loss) whose purpose is to enforce cross-modal alignment via positive/negative illumination maps and region-grounded contrastive supervision. These are presented as design choices motivated by the Retinex analogy rather than derived from prior results by the same authors. No equations reduce a prediction to a fitted input by construction, no uniqueness theorem is imported via self-citation, and the central claim (SOTA on MosMedData+ and QaTa-COV19) rests on experimental outcomes with released code. The method is therefore externally falsifiable and does not collapse to its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Since only the abstract is available, specific free parameters, axioms, and invented entities cannot be identified from the text. The method introduces new blocks and losses, but details are not provided.

pith-pipeline@v0.9.1-grok · 5853 in / 1126 out tokens · 34310 ms · 2026-06-29T04:57:07.602169+00:00 · methodology

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

30 extracted references

  1. [1]

    In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention

    Bozorgpour, A., Kolahi, S.G., Azad, R., Hacihaliloglu, I., Merhof, D.: CENet: Con- text enhancement network for medical image segmentation. In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention. pp. 120–129 (2025)

    2025

  2. [2]

    In: IEEE/CVF International Conference on Computer Vision

    Cai, Y., Bian, H., Lin, J., Wang, H., Timofte, R., Zhang, Y.: Retinexformer: One- stage Retinex-based Transformer for low-light image enhancement. In: IEEE/CVF International Conference on Computer Vision. pp. 12504–12513 (2023)

    2023

  3. [3]

    In: European Conference on Computer Vision Workshops

    Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M.: Swin- Unet: Unet-like pure Transformer for medical image segmentation. In: European Conference on Computer Vision Workshops. pp. 205–218 (2022)

    2022

  4. [4]

    In: International Conference on Machine Learning

    Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for con- trastive learning of visual representations. In: International Conference on Machine Learning. pp. 1597–1607 (2020)

    2020

  5. [5]

    In: IEEE International Conference on Image Processing

    Degerli, A., Kiranyaz, S., Chowdhury, M.E., Gabbouj, M.: Osegnet: Operational segmentation network for Covid-19 detection using chest X-Ray images. In: IEEE International Conference on Image Processing. pp. 2306–2310 (2022)

    2022

  6. [6]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Guo, Y., Zeng, X., Zeng, P., Fei, Y., Wen, L., Zhou, J., Wang, Y.: Common vision- language attention for text-guided medical image segmentation of pneumonia. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 192–201 (2024)

    2024

  7. [7]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Hu, J., Li, Y., Sun, H., Song, Y., Zhang, C., Lin, L., Chen, Y.W.: LGA: A lan- guage guide adapter for advancing the SAM model’s capabilities in medical im- age segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 610–620 (2024)

    2024

  8. [8]

    In: IEEE/CVF International Conference on Computer Vision

    Huang, S.C., Shen, L., Lungren, M.P., Yeung, S.: GLoRIA: A multimodal global- local representation learning framework for label-efficient medical image recogni- tion. In: IEEE/CVF International Conference on Computer Vision. pp. 3942–3951 (2021) 10 Jian Shi et al

    2021

  9. [9]

    IEEE Transactions on Medical Imaging44(4), 1821–1835 (2024)

    Huang, X., Li, H., Cao, M., Chen, L., You, C., An, D.: Cross-modal conditioned re- construction for language-guided medical image segmentation. IEEE Transactions on Medical Imaging44(4), 1821–1835 (2024)

    2024

  10. [10]

    Nature Methods18(2), 203–211 (2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: A self-configuring method for deep learning-based biomedical image segmentation. Nature Methods18(2), 203–211 (2021)

    2021

  11. [11]

    Scientific American237(6), 108– 129 (1977)

    Land, E.H.: The Retinex theory of color vision. Scientific American237(6), 108– 129 (1977)

    1977

  12. [12]

    IEEE Transactions on Medical Imaging43(1), 96–107 (2023)

    Li, Z., Li, Y., Li, Q., Wang, P., Guo, D., Lu, L., Jin, D., Zhang, Y., Hong, Q.: LViT: Language meets Vision Transformer in medical image segmentation. IEEE Transactions on Medical Imaging43(1), 96–107 (2023)

    2023

  13. [13]

    IEEE Transactions on Industrial Informatics 21(12), 9619–9630 (2025)

    Liu, Z., Geng, K., Cheng, X., Wang, Z., Yin, G., Sun, Y., Ma, T.: RetinexDet: En- hancing multispectral object detection via retinex state space duality and wavelet- based frequency adaptive fusion. IEEE Transactions on Industrial Informatics 21(12), 9619–9630 (2025)

    2025

  14. [14]

    In: International Conference on Learning Representations (2019)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019)

    2019

  15. [15]

    Digital Diagnostics1(1), 49–59 (2020)

    Morozov, S.P., Andreychenko, A.E., Blokhin, I.A., Gelezhe, P.B., Gonchar, A.P., Nikolaev, A.E., Pavlov, N.A., Chernina, V.Y., Gombolevskiy, V.A.: MosMedData: data set of 1110 chest CT scans performed during the COVID-19 epidemic. Digital Diagnostics1(1), 49–59 (2020)

    2020

  16. [16]

    In: International Conference on Machine Learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748–8763 (2021)

    2021

  17. [17]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedi- cal image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 234–241 (2015)

    2015

  18. [18]

    In: IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    Rui, S., Chen, L., Tang, Z., Wang, L., Liu, M., Zhang, S., Wang, X.: Multi-modal vision pre-training for medical image analysis. In: IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 5164–5174 (2025)

    2025

  19. [19]

    IEEE Transactions on Medical Imaging42(4), 935–946 (2022)

    Shi, J., Sun, B., Ye, X., Wang, Z., Luo, X., Liu, J., Gao, H., Li, H.: Semantic de- composition network with contrastive and structural constraints for dental plaque segmentation. IEEE Transactions on Medical Imaging42(4), 935–946 (2022)

    2022

  20. [20]

    In: International Conference on Medical Image Com- puting and Computer-Assisted Intervention

    Tomar, N.K., Jha, D., Bagci, U., Ali, S.: TGANet: Text-guided attention for im- proved polyp segmentation. In: International Conference on Medical Image Com- puting and Computer-Assisted Intervention. pp. 151–160 (2022)

    2022

  21. [21]

    In: IEEE/CVF International Conference on Computer Vision

    Wang, W., Zhou, T., Yu, F., Dai, J., Konukoglu, E., Van Gool, L.: Exploring cross-image pixel contrast for semantic segmentation. In: IEEE/CVF International Conference on Computer Vision. pp. 7303–7313 (2021)

    2021

  22. [22]

    In: European Conference on Computer Vi- sion

    Wu, Y., He, K.: Group normalization. In: European Conference on Computer Vi- sion. pp. 3–19 (2018)

    2018

  23. [23]

    IEEE Transactions on Circuits and Systems for Video Technology35(4), 3234–3249 (2024)

    Xu, M., Xiao, T., Liu, Y., Tang, H., Hu, Y., Nie, L.: CMIRNet: Cross-modal interactive reasoning network for referring image segmentation. IEEE Transactions on Circuits and Systems for Video Technology35(4), 3234–3249 (2024)

    2024

  24. [24]

    IEEE Transactions on Geoscience and Remote Sensing61, 1–12 (2023)

    Yan, T., Wan, Z., Zhang, P., Cheng, G., Lu, H.: TransY-Net: Learning fully trans- former networks for change detection of remote sensing images. IEEE Transactions on Geoscience and Remote Sensing61, 1–12 (2023)

    2023

  25. [25]

    In: IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: LAVT: Language- aware Vision Transformer for referring image segmentation. In: IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 18155–18165 (2022) Title Suppressed Due to Excessive Length 11

    2022

  26. [26]

    IEEE Transactions on Medical Imaging45(2), 477–489 (2026)

    Zeng, Q., Luo, H., Lu, Z., Xie, Y., Wang, Z., Zhang, Y., Xia, Y.: Harnessing text insights with visual alignment for medical image segmentation. IEEE Transactions on Medical Imaging45(2), 477–489 (2026)

    2026

  27. [27]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Zhong, Y., Xu, M., Liang, K., Chen, K., Wu, M.: Ariadne’s thread: Using text prompts to improve segmentation of infected areas from chest X-ray images. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 724–733 (2023)

    2023

  28. [28]

    Engineering Applications of Artificial Intelligence144, 110073 (2025)

    Zhou, X., Song, Q., Nie, J., Feng, Y., Liu, H., Liang, F., Chen, L., Xie, J.: Hybrid cross-modality fusion network for medical image segmentation with contrastive learning. Engineering Applications of Artificial Intelligence144, 110073 (2025)

    2025

  29. [29]

    In: International Workshop on Deep Learning in Medical Image Analysis

    Zhou, Z., Rahman Siddiquee, M.M., Tajbakhsh, N., Liang, J.: UNet++: A nested U-Net architecture for medical image segmentation. In: International Workshop on Deep Learning in Medical Image Analysis. pp. 3–11 (2018)

    2018

  30. [30]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Zhu, W., Chen, X., Qiu, P., Farazi, M., Sotiras, A., Razi, A., Wang, Y.: SelfReg- UNet: Self-regularized UNet for medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 601–611 (2024)

    2024