List of Papers Browse by Subject Areas Author List
Abstract
In latent diffusion-based segmentation, the model is trained to recover anatomically consistent regions from noisy latent mask representations conditioned on the input image. Recent diffusion-based approaches provide strong generative priors, but their conditioning is often limited to image or latent features, without explicit structural or semantic guidance. This may lead to unstable predictions and imprecise boundary delineation. In this work, we propose a text-guided multi-frequency latent diffusion framework for medical image segmentation, termed TMF-Seg. The model introduces an Image–Text Conditional Fusion module that encodes image and text conditions and integrates them through cross-attention to obtain a semantically informed representation. We further design a Text-Guided Multi-Frequency Enhancement module that decomposes the fused features into high-, mid-, and low-frequency components and adaptively reweights them according to the textual context, providing complementary semantic and structural priors for the diffusion process. The enhanced conditional features are then fed into a latent diffusion U-Net for efficient single-step segmentation. Experiments on three public datasets, Kvasir-SEG, MosMedData+, and QaTa-COV19, show that TMF-Seg consistently improves segmentation accuracy over several strong baselines. The code is publicly available at https://github.com/qczggaoqiang/TMF-Seg.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2061_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/qczggaoqiang/TMF-Seg
Link to the Dataset(s)
Kvasir-SEG: https://datasets.simula.no/kvasir-seg/
MosMedData+: https://github.com/HUANGLIZI/LViT
QaTa-COV19: https://github.com/HUANGLIZI/LViT
BibTex
@InProceedings{GaoQia_TextGuided_MICCAI2026,
author = { Gao, Qiang AND Wang, Yi AND Zhang, Yong AND Du, Lan AND Li, Yong AND Chen, Cunjian},
title = { { Text-Guided Multi-frequency Latent Diffusion for 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 TMF-Seg, a latent diffusion framework for medical image segmentation that strengthens the conditioning pathway rather than modifying the diffusion backbone itself. The method adds two main components:
an Image–Text Conditional Fusion (ITCF) module that uses cross-attention to fuse image latents with text embeddings from Bio-ClinicalBERT, and a Text-Guided Multi-Frequency Enhancement (TGME) module that decomposes fused features into high-, mid-, and low-frequency components and adaptively reweights them according to text-derived semantic context.
The authors keep the autoencoder and diffusion model frozen and train only the proposed condition encoder. They evaluate the approach on three public datasets—Kvasir-SEG, MosMedData+, and QaTa-COV19—and report consistent gains over CNN, CNN-transformer, and latent-diffusion baselines, with additional ablation evidence supporting the contribution of the ITCF and TGME modules.
- 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.
Clear methodological idea. The paper has a focused contribution: improving latent diffusion segmentation by enriching the conditioning signal with both semantic guidance from text and structural guidance from frequency-aware enhancement. This is more coherent than papers that combine many loosely connected modules. Well-motivated design. The paper argues that latent diffusion segmentation quality depends strongly on conditioning, and that image-only conditioning may not provide enough semantic or structural guidance in hard regions. The proposed ITCF and TGME directly target this motivation. Architectural modularity. A practical strength is that the method leaves the autoencoder and diffusion backbone frozen and only redesigns the condition encoder. This makes the contribution relatively modular and easier to integrate into existing latent diffusion segmentation pipelines. Reasonable empirical coverage. The method is tested on three public datasets spanning colonoscopy, CT, and chest X-ray segmentation, which is better than validating on only a single dataset or modality. Consistent quantitative improvements. The paper reports that TMF-Seg outperforms all listed baselines across the three datasets. Compared with SDSeg, the gains are modest but consistent: Kvasir-SEG Dice: 91.03 → 92.63 MosMedData+ Dice: 80.97 → 81.34 QaTa-COV19 Dice: 89.92 → 91.42. Ablation study supports both modules. On QaTa-COV19, adding ITCF or TGME individually improves over SDSeg, and the full model performs best, supporting complementarity between the two components. Paper organization is straightforward. The framework diagram and module descriptions are reasonably easy to follow, and the method-to-experiment flow is logical. The figures on pages 3–4 help clarify the overall pipeline and the TGME design.
- 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.
Novelty appears incremental rather than strong. The paper combines known ingredients—latent diffusion segmentation, image-text fusion via cross-attention, and frequency-based feature decomposition/reweighting. The contribution is a sensible integration of these ideas, but based on the paper alone it does not yet read as a fundamentally new segmentation paradigm. The closest prior art cited by the paper already includes text-guided segmentation/diffusion and texture/semantic enhancement, so the novelty margin over existing latent diffusion conditioning approaches seems moderate. The paper should position itself more sharply against recent text-guided and semantics-enhanced segmentation works, especially SDSeg, TSLDeg, LViT, STPNet, and the cited text-guided diffusion literature. Text supervision is insufficiently specified. A central claim is that the method benefits from textual descriptions, but the paper does not clearly explain where these text descriptions come from, how they are constructed for each dataset, whether they are standardized templates, whether they are manually curated, and what information they contain relative to the mask labels. This is important because performance may depend substantially on prompt design and text quality. Without this detail, both fairness and reproducibility are weakened. The method section states that the input consists of an image and its textual description, but the dataset section does not explain the source of the text. Limited analysis of the text contribution. Although the ablation shows ITCF helps, there is no deeper study of text robustness: e. g. , different prompt templates, noisy/incorrect text, empty text, generic text, or prompt sensitivity across datasets. Because the paper’s main claim is text-guided segmentation, this missing analysis is noticeable. Ablation scope is narrow. The ablation is conducted only on QaTa-COV19.Stronger support would come from at least one additional dataset, and ideally from analysis of frequency-band weights, DCT design choices, or alternative fusion baselines. No statistical analysis or variance reporting. The reported gains on some datasets are fairly small, especially on MosMedData+. The paper does not report standard deviations, confidence intervals, or multi-seed results, so it is hard to judge whether the improvements are robust. Limited efficiency discussion. The paper states training hardware and basic settings, but does not quantify added computational overhead, parameter increase, or inference-time cost relative to SDSeg. Since the method adds text encoding and multi-frequency branches, such analysis would strengthen the practical case. Clinical impact remains indirect. The datasets are public and diverse, but there is no validation on a clinical workflow, no expert reader study, and no downstream task analysis. So the work is best viewed as a methodological paper rather than a clinically translated contribution
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
-
- 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
The paper addresses an interesting direction for diffusion-based segmentation by enriching the conditioning pathway with language and frequency-aware structure modeling. The overall framework is coherent and the empirical results are promising. To strengthen the paper, I encourage the authors to clarify the provenance and construction of the text descriptions for all datasets, add a more comprehensive comparison to closely related text-guided/semantic-enhanced segmentation methods, and include stronger robustness analysis around prompt quality and statistical significance. A runtime/parameter comparison against SDSeg would also improve the practical evaluation
- 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?
I lean Weak Accept because the paper presents a coherent and technically sound improvement to latent diffusion segmentation, supported by consistent gains on three public datasets and a reasonable ablation showing the benefit of both proposed modules. The idea of strengthening diffusion conditioning using image-text fusion and text-guided multi-frequency enhancement is sensible and appears effective in practice.
At the same time, I do not view the novelty as especially high. The work appears more like a careful combination of existing ideas than a major conceptual advance, and the paper does not yet analyze several important issues deeply enough—especially the origin and role of the text descriptions, robustness to prompt design, statistical significance of the gains, and computational overhead. The ablation is also limited to one dataset. These limitations prevent me from scoring it as a clear accept. Still, the method is well motivated, the empirical results are positive, and the contribution seems above the rejection threshold for a methodological MICCAI paper
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 TMF-Seg, a text-guided multi-frequency latent diffusion framework for medical image segmentation. The core contributions are: 1.Propose an Image-Text Conditional Fusion (ITCF) module that fuses image latent features and medical text embeddings via cross-attention to inject semantic guidance into diffusion-based segmentation. 2.Design a Text-Guided Multi-Frequency Enhancement (TGME) module that decomposes fused features into high/mid/low frequency components via DCT and adaptively weights them using text semantics to strengthen structural and boundary priors. 3.Enhance the conditioning of a frozen latent diffusion model to achieve more accurate and stable segmentation, validated on three public datasets (Kvasir-SEG, MosMedData+, QaTa-COV19).
- 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.Compared to prior works, this work innovatively integrates textual semantics and multi-frequency structural modeling into the diffusion conditioning, which is well justified for medical images with ambiguous boundaries and low contrast. 2.The TGME module uses DCT-based frequency decomposition and text-guided adaptive weighting. It allows the model to dynamically emphasize high-frequency details (boundaries), mid-frequency structures, or low-frequency global context based on textual prompts. This is biologically plausible and practically beneficial for medical segmentation. 3.The experiments are complete, and the evaluated datasets are diverse. 4.The paper is well organized, with intuitive figures and reasonable explanations. The motivation, method, and logic are easy to follow.
- 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 training loss was not descripted in the whole paper. Please add this content. 2.Can you ensure the alignment between image and text features? How to verify if text and images are aligned? Please provide corresponding attention map. 3.The novelty in core technical components is limited, for example, the ITCF module uses standard cross-attention for image-text fusion. 4.The paper does not provide examples or details of the text prompts used for each dataset. Please provide the details of the text prompts. 5.No robustness testing on noisy, low-quality, or pathological images. 6.Please provide detailed training strategy, such as loss functions.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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
1.Please regraph the method figure, i.e., place the model input on the left and the output on the right. 2.Please highlight the improvement of your method in Fig. 3.
- 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?
The novelty is relatively limited, and its clinical application will be evaluated. The paper is well organized, with intuitive figures and reasonable explanations. The motivation, method, and logic are easy to follow.
- Reviewer confidence
Very confident (4)
- [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 clearly to my question. After careful consideration, I have decided to accept this article.
Review #3
- Please describe the contribution of the paper
1.The main novelty is the DCT-based multi-frequency design. The model splits the fused feature into high-, mid-, and low-frequency parts. This lets the model handle fine details, boundaries, and global structure in a more explicit way. 2.The text feature is used to predict the weight of each frequency branch. This means the model can adaptively decide which frequency information is more important for the current image.
- 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.
A major strength of the paper is the method design. The model decomposes the fused feature into different frequency components, which gives a more fine-grained way to model details, boundaries, and global structure. It then uses a weighting module to adaptively combine these frequency branches, so the text signal can provide more effective guidance during segmentation. This module provides a new angle for fusing text features and image features.
- 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) A major weakness is that the ablation study is limited to module-level comparisons. It only shows that the proposed module is useful as a whole, but it does not verify whether the high-, mid-, and low-frequency branches actually capture the different types of information claimed by the authors, such as fine details, boundaries, and global structure. Therefore, the current experiments do not fully support the interpretation of the proposed frequency decomposition design. 2) In addition, the paper does not provide enough analysis of the adaptive weighting process, for example how the weights are distributed across different frequency branches in practice. More experimental evidence would be helpful to clarify the contribution of the frequency decomposition design to the fusion of text and image features. 3) The related work discussion is not sufficient, author/s should discuss existing works that adopt Multi-Frequency in greater domain and differentiate the contribution against them.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 paper has a clear idea, and the method is interesting. In particular, the frequency decomposition design gives a new way to fuse text and image features, and the results are better than several baselines. However, the current evidence is not strong enough to fully support the main claim of the paper. The ablation study is limited to module-level comparisons, so it does not show whether the high-, mid-, and low-frequency branches really capture different types of information, as claimed by the authors. The paper also lacks an analysis of the adaptive weighting process, for example, how the weights are distributed across the three frequency branches.
- Reviewer confidence
Very confident (4)
- [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.
Although the paper does not introduce a fundamentally new methodological idea, it presents a meaningful application that demonstrates the feasibility and practical value of applying the multi-frequency concept in this setting. The rebuttal has clarified several of my concerns. Given the limited space of the MICCAI format, the paper is acceptable.
Author Feedback
We thank the reviewers for their positive feedback on the novel method design (R1, R2, R4), effective experimental results (R1, R2), clear organization (R1, R2), and ablation study (R1), and thank the AC for acknowledging these strengths. Following the rebuttal guidelines, we address remaining concerns below without new experiments.
Q1: Frequency branch validation (R4) We clarify that the three frequency branches are defined by pre-defined, non-overlapping DCT coefficient index sets registered as fixed buffers (not learned): high-freq uses top-16 DCT indices (capturing fine-grained details and boundaries), mid-freq uses low-8 indices (structural transitions), and low-freq uses bottom-8 indices (global anatomical context). These index sets are mathematically disjoint by construction, guaranteeing spectral separation. The DCT filters are fixed (register_buffer), so each branch provably receives distinct frequency content.
Q2: Adaptive weight distribution analysis (R4) The text-to-frequency mapping is a two-layer MLP (768→128→3) followed by Softmax, producing per-sample weights α_h+α_m+α_l=1 conditioned on each sample’s textual description via mean-pooled Bio-ClinicalBERT embeddings. For example, “Bilateral pulmonary infection, two infected areas, all left lung and all right lung.” and “one medium white round polyp, located in center of the image” encode very different semantic signals, leading to different frequency weight distributions. The Grad-CAM results in Fig. 4 further confirm that TGME shifts attention toward structure-aware regions, consistent with text-driven frequency modulation.
Q3: Text prompt details (R1, R2) Text descriptions are publicly available clinical annotations. For QaTa-COV19 and MosMedData+, they are released by LViT [10] authors as structured annotations (.xlsx format) and described with specific lesion characteristics per image (e.g., for one QaTa-COV19 sample: “Bilateral pulmonary infection, three infected areas, upper middle lower left lung and upper middle lower right lung”). For Kvasir-SEG, descriptions were obtained from corresponding paper authors upon request and will be released with our code upon acceptance. All prompts are per-image clinical descriptions providing inherent semantic diversity rather than generic templates. Text robustness analysis is feasible and will be explored.
Q4: Training loss and variance reporting (R2,R1) Following SDSeg [12], we adopt L = L_noise + λL_latent, where L_noise predicts Gaussian noise and L_latent minimizes latent estimation error (λ=1, both MAE). Only the condition encoder (ITCF+TGME) is trained; Diffusion backbone and autoencoder remain frozen. Variance reporting and statistical analysis will be included in the revised paper.
Q5: Related work on multi-frequency methods (R4) Our TGME differs from prior multi-frequency works (e.g., [15]) in two key aspects: (1) frequency decomposition is applied in the latent conditioning space rather than image space, which is novel for diffusion-based segmentation; (2) frequency reweighting is dynamically text-guided per sample, rather than statically learned. We will expand the related work discussion to highlight these distinctions.
Q6: Cross-modal fusion design and alignment (R2) Cross-attention in ITCF computes attention scores between each spatial image position (query) and text tokens (key/value), enabling image features to selectively attend to semantically relevant textual cues. Grad-CAM improvements in Fig. 4 support that meaningful semantic alignment is learned. We will include explicit cross-attention map visualizations upon acceptance.
Q7: Novelty (R1, R2) The contribution lies in the principled integration of text-guided semantic conditioning and frequency-domain structural priors within latent diffusion conditioning, unexplored for medical image segmentation. Consistent gains over SDSeg, TSLDeg, and LViT (Table 1) support its practical value.
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.
This paper introduces TMF-Seg, a text-guided multi-frequency latent diffusion framework that enhances medical image segmentation via image-text fusion and adaptive frequency feature reweighting. Reviewers commend the modular design, consistent performance gains across three diverse public datasets, and intuitive integration of clinical text semantics and structural frequency information. There are some limitations which include incremental technical novelty, incomplete text prompt documentation, and narrow ablation studies that fail to fully validate frequency branch functionality and cross-modal alignment.
- 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.
Most of concerns have been addressed or clearified adequately.
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.
Overall, the work appears above the acceptance threshold. The contribution is not a major conceptual breakthrough, and several analyses would still strengthen the paper, but the method is well motivated, the reported improvements are consistent, and the rebuttal resolves enough of the key methodological and reproducibility concerns to support 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.
three accept, i agree.
