List of Papers Browse by Subject Areas Author List
Abstract
Clinical text can narrow down what to segment, but recent text-guided designs emphasize spatial alignment while overlooking frequency content that governs texture and boundaries. We propose Dual- Domain Cross-Modal Decoding (DD-CMD) for clinical text-guided pulmonary infection segmentation, integrating two complementary forms of language guidance during decoding. In the spatial domain, Text-Guided Spatial Cross-Attention (TGSA) aligns multi-scale visual tokens with text semantics and updates features through gated residual fusion. In the frequency domain, Spectral-Text Adaptive Modulation (STAM) applies a 2D DCT to compute learnable band-energy statistics and predicts text-conditioned FiLM parameters to recalibrate decoder channels for frequency-aware decoding. DD-CMD embeds TGSA and STAM into a coarse-to-fine decoder (7×7 → 56×56) and restores full-resolution masks using a lightweight two-stage refinement module. Experiments on QaTa- COV19 and MosMedData+ show that DD-CMD achieves 91.46% Dice / 84.26% mIoU and 81.95% Dice / 69.42% mIoU, respectively, with average gains of +1.96 Dice and +2.67 mIoU over the strongest prior baselines. Code: https://github.com/maklachur/DD-CMD.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4478_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/maklachur/DD-CMD
Link to the Dataset(s)
https://www.kaggle.com/datasets/aysendegerli/qatacov19-dataset
https://www.medseg.ai/covid-19
https://github.com/HUANGLIZI/LViT
BibTex
@InProceedings{RahMd_DualDomain_MICCAI2026,
author = { Rahman, Md Maklachur AND Hammond, Tracy},
title = { { Dual-Domain Cross-Modal Decoding for Clinical Text-Guided Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a novel framework named DD-CMD (Dual-Domain Cross-Modal Decoding) for the task of medical image segmentation guided by clinical text. The core idea is to recognize that existing methods overly focus on spatial domain alignment while neglecting the crucial role of frequency domain information in texture and boundary representation. Therefore, DD-CMD introduces two complementary text-guided mechanisms in the decoding stage: (1) Text-Guided Spatial Attention (TGSA) for aligning visual features with text semantics in the spatial domain; (2) Spectral-Text Adaptive Modulation (STAM), which computes learnable frequency band energy statistics via 2D Discrete Cosine Transform (DCT) and predicts text-conditioned FiLM parameters to recalibrate the decoder channels in the frequency domain. The framework adopts a coarse-to-fine decoding strategy (7×7 → 56×56) and recovers full-resolution masks through a lightweight two-stage refinement module.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The motivation is novel and profound: The core insight of the paper, that language guidance should not only align “where” in the spatial domain but also regulate “how to respond” in the frequency domain to preserve texture and boundaries, is highly enlightening. This goes beyond the current majority of works that only focus on spatial alignment, providing a new and valuable perspective for multimodal medical image analysis. 2.The technical solution is ingeniously and coherently designed: The design logic of the TGSA and STAM modules is clear, corresponding to the guidance requirements in the spatial and frequency domains respectively. Particularly, the STAM module, which combines DCT, learnable gating, and FiLM modulation, is an elegant and effective way of frequency domain modeling. The entire architecture (coarse-to-fine decoding + lightweight high-resolution refinement) also demonstrates a good engineering trade-off. 3.The experimental results are outstanding and well-validated: On the two challenging datasets, QaTa-COV19 and MosMedData+, DD-CMD achieved significantly better performance than the current state-of-the-art (SOTA) baseline models (for example, a Dice improvement of +3.33% on MosMedData+). The ablation study (Table 2) systematically proved the effectiveness of the TGSA, STAM, and refinement modules, and directly verified the model’s advantage in boundary accuracy through the HD95 metric. The visualization results (Figures 2 and 3) are also very convincing, clearly demonstrating the model’s ability in detail recovery. 4.The positioning of related work is accurate: The paper clearly points out the limitations of existing text-guided methods (overemphasis on spatial alignment) and proposes its own solution based on this, with a very precise problem positioning.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The distinction from recent frequency-domain multi-modal works needs to be strengthened: Although the motivation of the paper is very novel, it must be noted that at MICCAI 2025, Yu et al. published a work titled “Frequency-domain multi-modal fusion for language-guided medical image segmentation” (Reference [20]). This work also explored multi-modal fusion in the frequency domain. Although the implementation details of DD-CMD (using DCT and FiLM) are different, there is a certain similarity in the core idea (utilizing frequency-domain information for multi-modal fusion). The paper needs to discuss in more depth the differences with this work, for example, emphasizing that DD-CMD performs conditional channel recalibration in the decoder, not just feature fusion, and clarifying why this design is more suitable for segmentation tasks. 2.The universality of the STAM module needs to be verified: The STAM module relies on DCT, a global orthogonal transform. Although it performs well in tasks with specific texture patterns such as lung infection, its universality in other types of medical image segmentation tasks (such as organ segmentation, where boundaries may be more important than texture) is not clear. The paper only validated on the lung infection dataset, which is somewhat limited. 3.The analysis of computational complexity can be more detailed: Although the paper reported the number of parameters (47.77M) and FLOPs (20.31G) and compared them with some baselines, it did not detail the specific overhead introduced by DCT and the additional MLP (used to predict FiLM parameters). For a model aimed at practical use, a more fine-grained efficiency analysis (such as the proportion of running time for each module) would be more valuable. 4.The freezing strategy of the text encoder: The paper chose to freeze PubMedBERT to retain biomedical semantics and prevent overfitting. This is a reasonable strategy, but it may also limit the model’s ability to fine-tune for specific task text descriptions. A brief discussion on the impact of fine-tuning the text encoder (even based on preliminary experiments) would be beneficial.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
DD-CMD is a high-quality paper that makes a profound and novel contribution by introducing frequency-domain modeling into text-guided medical image segmentation. This is not merely a simple combination of existing techniques but an innovation driven by a deep understanding of the problem’s essence. The technical solution is ingeniously designed, and the experimental validation is comprehensive with outstanding results, especially in the improvement of boundary accuracy, which directly supports the core argument of the paper.
Although there is potential overlap with the work of Yu et al. (MICCAI 2025), DD-CMD demonstrates sufficient distinction and incremental contribution through its unique STAM module design and specific application in the decoder. Its consistently superior performance on two independent datasets strongly validates the method’s effectiveness and robustness.
Overall, this work exhibits strong originality and practicality, meeting MICCAI’s expectations for high-quality research. Therefore, I give it a high score of 5 and recommend its acceptance. I suggest that the authors conduct a more detailed comparison and discussion with reference [20] in the final version and consider verifying the method’s universality in a broader range of tasks in future work.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper proposes the DD-CMD framework, whose main contributions are: integrating spatial-domain text-guided cross-attention (TGSA) and frequency-domain text-conditioned FiLM modulation (STAM) during decoding to achieve complementary dual-domain language guidance, together with a coarse-to-fine cross-modal decoding structure and a two-stage refinement module for efficient full-resolution mask recovery. Experimental results verify the effectiveness of the proposed method.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Different from existing text-guided medical image segmentation methods that only focus on spatial-domain semantic alignment, this work innovatively introduces both spatial-domain text-guided cross-attention and frequency-domain text-conditioned modulation, respectively addressing the core problems of accurate lesion localization and boundary/texture fidelity. 2.A coarse-to-fine decoding architecture is adopted, which concentrates computationally intensive cross-modal interactions at low-resolution stages while recovering pixel-level details through a lightweight two-stage high-resolution refinement module, balancing model performance and clinical deployment feasibility. 3.Extensive experimental results on public clinical datasets demonstrate that the proposed model significantly outperforms state-of-the-art baseline methods in both segmentation accuracy and boundary fidelity.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.In the context of Equation (1), the visual features V (s)and text features T (s) are clearly defined, but what V~(s) represents is never defined anywhere. 2.The paper does not provide the specific mathematical expression of the loss function nor the weight parameters, which reduces the transparency of experimental details. 3.Equation (2) introduces a learnable parameter α (s) to control the influence of cross-modal attention, but there is no experimental evidence demonstrating whether its introduction is effective or what its benefits are compared to using a fixed value. 4.The model has only been validated in a single scenario of pulmonary infection. Whether it can be extended to other clinical segmentation tasks to demonstrate the generalizability of the proposed method’s design remains to be shown. 5.Although the paper achieves promising metrics on the MosMedData+ dataset, the dataset itself consists fundamentally of 3D CT volumetric data. The approach of processing 2D slices sequentially, as adopted in this work, inherently disregards the three-dimensional spatial continuity of lesions within the physical volume. It is recommended that the authors supplement their study with comparative experiments against established 3D segmentation networks, such as V-Net or 3D U-Net. 6.The current prompt texts are likely the default annotations provided with the dataset. In scenarios where textual descriptions are missing or extremely sparse, is it feasible to employ large language models (LLMs) to generate the necessary descriptions? Furthermore, do you believe the current architecture is robust enough to withstand the “hallucination noise” introduced by LLMs? Or what specific mechanisms should be improved or incorporated into the architecture to mitigate the introduction of such noise?
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The overall structure of the paper follows a certain logical flow; however, the presentation of some sections is rather disjointed, and the exposition of the methodology remains somewhat vague. While a number of experiments have been conducted, several shortcomings persist in the current work.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have responded to my core comments. While the paper still lacks sufficient novelty, I recommend its acceptance after comprehensive consideration.
Review #3
- Please describe the contribution of the paper
This paper proposes a new decoder architecture called DD-CMD for medical CT image segmentation using a text-guided approach. The proposed decoder considers both spatial and frequency domains, driving improved segmentation performance. In spatial domain, text and image features are aligned to guide the model on where to focus, while in frequency domain, the model adjusts which textures and boundaries to emphasize based on the text.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
One of the main strengths of this paper is that, unlike existing medical image segmentation approaches, it incorporates frequency domain information in addition to spatial information, which helps improve performance. The paper also points out that segmentation of lung infections is difficult due to subtle textures, and introduces frequency content that is related to texture and boundary, going beyond existing text guided methods that mainly focus on spatial alignment. The architecture of the proposed model, including TGSA and STAM, is clearly described in both the figure and the methodology section. In addition, the use of a coarse-to-fine decoding strategy helps reduce computational cost. The use of PubMedBERT for text guidance can also be considered as a domain-specific design tailored for medical applications.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The paper lacks sufficient justification for choosing DCT as the frequency representation. Although frequency is emphasized in the introduction, the actual frequency modeling appears relatively simple. Additionally, the use of FiLM based modulation seems to be an application of an existing method within the decoder, which raises some concerns about the level of novelty. Moreover, despite sufficient training, the performance improvement is relatively limited.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
This paper clearly defines the problem of lung infection segmentation, and the attempt to incorporate frequency content is a meaningful direction. The experimental results and ablation studies support the proposed methodology well. However, it would be helpful to include comparisons with other frequency representations, which could further justify the use of the FiLM based approach.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper clearly identifies the limitations of existing lung segmentation approaches and proposes a meaningful method that combines spatial and frequency domain representations. The ablation studies effectively demonstrate the contribution of the proposed method, and reproducibility is supported through code release. However, the proposed method appears to be closer to a combination of existing techniques, such as FiLM based modulation, which limits its novelty. In addition, the performance improvements are relatively modest considering the training cost. Overall, the paper is well structured and technically sound, but somewhat limited in terms of novelty and impact, making a score of 4(weak accept) appropriate.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have sufficiently addressed my concern.
Author Feedback
We thank all reviewers (R1–R3) and the meta-reviewer (MR) for their valuable feedback.
(MR/R1-Q1/R3) Difference from prior frequency-domain work, Ref. [20]: Both use frequency info, but differ in purpose. FMISeg [20] employs DWT to split the input into LF/HF images, processed via dual visual branches with language-frequency interaction. DD-CMD proposes a novel unified visual stream without decomposition and uses DCT for spectral calibration. TGSA aligns decoder features with text; STAM applies DCT to attended features, computes gated spectral energy, and predicts FiLM parameters for channel recalibration. DD-CMD performs decoder-side spectral modulation, not LF/HF feature fusion.
(MR/R3) DCT choice, FiLM design, & gains: We agree that FiLM exists, but our novelty is integrating it for decoder-level spectral calibration conditioned on text after TGSA alignment. DCT is chosen for being real-valued, orthogonal, and parameter-free, unlike DFT, which needs complex phase handling, providing stable energy compaction on low-res decoder maps without learned spectral encoders. STAM is not just a transform; it gates DCT coefficients, gathers channel-wise spectral energy, fuses with text, and predicts FiLM parameters for boundary/texture channels. Outcome: +3.33 Dice/+4.35 mIoU on MosMedData+, consistent results on QaTa-COV19.Removing STAM or FiLM reduces Dice/mIoU and worsens HD95 (Table 2).
(MR/R2-Q1,Q2,Q3) Method details & α: In Eq. (1), V~^(s) is the LocalConvMix output before cross-attention. It provides TGSA visual queries and, with text features, estimates the agreement gate g^(s). We will clarify this notation. As stated in Sec. 3, we use equally weighted Dice and cross-entropy loss: L=L_Dice+L_CE. Eq. (2) defines α^(s) as a learnable residual scale for stage-specific text influence. A fixed α applies the same text influence across decoder stages, although stages differ in semantic localization and structural refinement, making the decoder less adaptive. Thus, we use learnable α^(s) to calibrate text influence during training. Removing the adaptive gate, TGSA, or STAM degrades performance (Table 2).
(MR/R1-Q2/R2-Q4,Q5) Validation scope & 2D CT: DD-CMD follows the standard LViT [13] protocol used by recent baselines (TGCAM [9], MMI-UNet [4], RecLMIS [11], ViTexNet [3]) for fair comparison. Each sample is an image-text-mask triplet. Although MosMedData+ derives from CT volumes, the benchmark is slice-wise. A 3D comparison would require volumetric inputs, 3D labels, and volume-level text, which defines a different task that does not isolate our text-guided decoder contribution.
(R1-Q4/R2-Q6) Text encoder & prompt reliability: We freeze PubMedBERT to retain biomedical priors and avoid overfitting with limited text diversity. Task adaptation occurs through TGSA, STAM, gates, FiLM, and refinement modules, not by updating the full encoder. For sparse text, Table 3 shows very short descriptions hurt performance on benchmarks, while L=24 remains stable with no clear benefit from longer text. DD-CMD accepts PubMedBERT-encoded text, but we did not use LLM prompts. To reduce hallucination, prompts should be clinically verified or constrained. DD-CMD limits over-conditioning via gated residual TGSA, where g^(s) scales text influence by image-text agreement, while the image pathway preserves visual evidence. The no-text variant in Table 2 shows reliable visual-only segmentation, supporting text as controlled guidance, not a dependency.
(R1-Q3/R3) Complexity & gain-cost trade-off: DD-CMD uses only TGSA/STAM at 7×7–56×56, with lightweight 112×112/224×224 refinement. STAM overhead is minimal: a fixed, parameter-free DCT, a C-dimensional descriptor, and 2C FiLM outputs per stage. DD-CMD uses 47.77M/20.31G FLOPs, comparable to MMI-UNet (56.2M/22.1G) and much less than LViT (54.1G) and LGA (381.1G). On MosMedData+, with prior mIoU 63–65%, DD-CMD reaches 69.42% (+4.35), showing effective dual-domain decoding beyond spatial-only alignment.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper proposes DD-CMD, a text-guided pulmonary infection segmentation framework combining spatial cross-attention and frequency-domain text-conditioned modulation, with strong reported results and useful ablations on QaTa-COV19 and MosMedData+. Reviewers’ scores are mixed due to concerns about novelty and clarity. In rebuttal, the authors should specifically clarify the distinction from prior frequency-domain multimodal segmentation work, justify the DCT/FiLM design choices, address the limited validation scope and 2D treatment of CT data, and provide missing methodological details.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Overall, the reviewers’ consensus is that this is a sound and relevant contribution with good empirical support. My final recommendation is acceptance. The authors should use the camera-ready version to sharpen the discussion of the relationship to prior frequency-domain multimodal segmentation work, clarify the mathematical notation and loss formulation, and more explicitly state the limits of the current validation setting.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper initially received mixed reviews (2 positive and 1 negative R2). After rebuttal, R2 raised his score, although the level of novelty is moderate (R2, R3). Hence, I recommend the acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors have responded to core comments sufficiently.