List of Papers Browse by Subject Areas Author List
Abstract
Brain tumor segmentation remains challenging because the three standard sub-regions, i.e., whole tumor (WT), tumor core (TC), and enhancing tumor (ET), often exhibit ambiguous visual boundaries. Integrating radiological description texts with imaging is promising; however, most multimodal approaches typically compress a report into a single global text embedding shared across all sub-regions, overlooking their distinct clinical characteristics. We propose TextSCP (text-modulated soft cascade architecture with prompts), a hierarchical text-guided framework built on the TextBraTS baseline with three novel components: (1) a text-modulated soft cascade decoder that predicts WT->TC->ET in a coarse-to-fine manner consistent with their anatomical hierarchy. (2) sub-region-aware prompt tuning, using learnable soft prompts with a LoRA-adapted BioBERT encoder to generate specialized representations for each sub-region; (3) text-semantic channel modulators that convert these representations into channel-wise refinement signals, enabling the decoder to emphasize features aligned with clinically described patterns. Results on the TextBraTS dataset demonstrate consistent improvements across all sub-regions, outperforming state-of-the-art methods by 1.7% and 6.2% on Dice and HD95, respectively. Code is available at: https://github.com/Bahram-Mohammadi/TextSCP.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5067_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Bahram-Mohammadi/TextSCP
Link to the Dataset(s)
N/A
BibTex
@InProceedings{MohBah_Hierarchical_MICCAI2026,
author = { Mohammadi, Bahram AND Huy, Ta Duc AND Sheikholeslami, Afrouz AND Chen, Qi AND Phan, Vu Minh Hieu AND White, Sam AND To, Minh-Son AND Zhang, Xuyun AND Beheshti, Amin AND Zhou, Luping AND Qi, Yuankai},
title = { { Hierarchical Text-Guided Brain Tumor Segmentation via Sub-Region-Aware Prompts } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a novel framework for hierarchical text-guided brain tumor segmentation. It aims to address the limitation in current image-text multimodal segmentation methods, where the text is only aligned with the global image, overlooking the specific interactive characteristics between the text and individual target sub-regions. Innovatively, this study utilizes LoRA to fine-tune a BERT model to extract specialized text representations tailored for the multi-level sub-regions in brain tumor segmentation, such as WT (Whole Tumor), ET (Enhancing Tumor), and TC (Tumor Core). Furthermore, it designs a multi-stage cascaded fusion and prediction module to integrate the text-guided visual features, effectively improving the segmentation accuracy across all regions compared to the baseline.
- 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.
I believe this paper has several key strengths:
Precise Text-to-Region Alignment: The authors accurately identify a major limitation in current text-guided segmentation: the reliance on global text fusion, which overlooks the specific impact of complex textual features on individual target sub-regions. To address this, they propose a highly effective method that utilizes multi-level and multi-target independent text features for fusion.
Parameter-Efficient Hierarchical Fusion: The methodology is both elegant and computationally practical. By cleverly utilizing LoRA for hierarchical text fine-tuning and introducing a multi-level progressive fusion scheme, the approach significantly improves segmentation performance across various regions without imposing a heavy parameter burden.
Rigorous Experimental Validation and Clarity: The paper is exceptionally well-structured and logically sound. Furthermore, the proposed approach is strongly corroborated by extensive and comprehensive experimental results, making the claims highly convincing.
- 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.
I believe this paper has the following weaknesses:
Lack of Fine-Grained Semantic Alignment: While applying LoRA-based prompt fine-tuning to globally extracted text features yields strong results, this approach overlooks the explicit semantic correspondence between clinical reports and distinct lesion regions. This renders the text-image fusion process somewhat of a black box, hindering further integrated clinical analysis.
Absence of Robustness Evaluation: Although the ablation studies effectively validate individual modules, the study lacks critical robustness experiments. For instance, it remains unclear whether masking or degrading specific textual features during training or inference would severely impact the corresponding segmented regions.
Limited Reproducibility: Although the authors provide basic implementation hyperparameters in Section 3.1, they do not provide open-source code. Attempting to fully reproduce the complex multi-stage cascade and prompt-tuning mechanisms based solely on the paper’s textual descriptions and the high-level architecture in Figure 1 would still be highly challenging.
- 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 does not provide sufficient information for reproducibility.
- 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?
I recommend that this paper be accepted directly, without the need for a rebuttal. To address the common limitation in text-guided segmentation where text is typically only aligned with the global image, this paper proposes a method that utilizes a fine-tuned BERT to extract localized textual information, coupled with a multi-stage cascaded hierarchical fusion and prediction scheme. This approach effectively improves segmentation accuracy across all specific regions. The paper is well-structured and logically sound, and the experimental section provides compelling evidence demonstrating the effectiveness of the proposed modules. Although there are still some minor shortcomings regarding interpretability and robustness evaluation, it remains an excellent piece of work. Therefore, I strongly recommend this paper for acceptance.
- 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 TextCSP, a hierarchical text-guided brain tumor segmentation framework built on the TextBraTS baseline. The paper introduces three main components: (1) a text-modulated soft cascade decoder that predicts tumor sub-regions in a hierarchical order WT to TC to ET, (2) sub-region-aware prompt tuning with a shared LoRA-adapted BioBERT encoder to generate specialized text representations, and (3) text-semantic channel modulators that inject branch-specific linguistic priors into the decoder features. The motivation is that existing text-guided methods use a single global text embedding and ignore the anatomical containment hierarchy among tumor sub-regions. The paper reports improved Dice and HD95 performance on the TextBraTS dataset relative to prior methods and the TextBraTS baseline.
- 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 paper addresses an interesting and relevant problem, namely how to use radiological text more effectively in multimodal brain tumor segmentation. 2). The idea of explicitly modeling the hierarchy among WT, TC, and ET is intuitive and clinically meaningful. 3). Another strength is that the paper moves beyond a single shared text representation and attempts to tailor textual information to different sub-regions. 4). The method improves both average Dice and HD95, and the gains are consistent across all three tumor sub-regions.
- 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). My main concern is the limited strength of the ablation evidence for the claimed key contribution. The paper strongly emphasizes sub-region-aware prompt tuning as one of its central novelties, but in Table 2, adding this component on top of the soft cascade improves Dice only from 85.9 to 86.4, which is a relatively small gain. LoRA and text modulation add only further modest increments. This raises concern about whether the text-specific contribution is as significant as claimed. 2). The second main concern is about the computational efficiency. As adding or removing the text-specific module does not contribute significantly towards performance gain. I am concerned about how much computation did these module add in the overall network. 3). My third main concern is the no of token as mentioned in Table 3.First is there are criteria of selecting these no of learnable tokens as going from 1 to 4 performance is getting better but when it is further increased from 4 to 10 performance get worsen why? As per my understanding 10 is not big number in terms of leranable tokens why perfromance get worsen. Is there any trend if goes beyond 10 is still performance drops or not? As based on this ablation which appears to be vague it can not be concluded that as no of tokens increases beyond 4 performance will always go down.
- 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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?
My decision is mainly driven by the gap between the claimed contribution and the strength of the supporting ablation evidence. The paper presents sub-region-aware prompt tuning as a major novelty, but the actual isolated gain from that component is modest. The more convincing contribution appears to be the hierarchical cascade design. Because of this, I am not fully convinced that the paper’s text-specific novelty is strong enough for acceptance in its current form. Secondly, there is no discussion about the computational complexity of adding or removing the textual modules in the overall network. So, it is vague to conclude beacause if removing these textual modules reduce overall model size and computation significantly why they are necessary to add in the system as they are not improving performnace significantly if added in network.
- 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 #3
- Please describe the contribution of the paper
The paper explicitly models the anatomical hierarchy (WT–>TC–> ET), leading to more consistent and clinically meaningful segmentation results. In addition, the introduction of sub-region-aware prompt tuning enables the model to extract specialized textual features for each tumor sub-region, thereby improving semantic alignment between text and image. Furthermore, the proposed framework demonstrates clear and consistent performance gains over strong baselines in both Dice and HD95 metrics, with ablation studies validating the effectiveness of each component.
- 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.
The paper explicitly models the anatomical hierarchy (WT–>TC–>ET), leading to more consistent and clinically meaningful segmentation results. In addition, the introduction of sub-region-aware prompt tuning enables the model to extract specialized textual features for each tumor sub-region, thereby improving semantic alignment between text and image. Furthermore, the proposed framework demonstrates clear and consistent performance gains over strong baselines in both Dice and HD95 metrics, with ablation studies validating the effectiveness of each component.
- 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) Most techniques (prompt tuning, LoRA, attention-based modulation) are existing methods, making the contribution largely incremental. (2) Experiments are conducted only on the TextBraTS dataset, raising concerns about generalizability to other datasets or modalities. (3) The framework introduces multiple components (cascade decoder, prompt tuning, modulators), which may increase implementation complexity and computational overhead.
- 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 does not provide sufficient information for reproducibility.
- 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.
(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 paper explicitly models the anatomical hierarchy (WT–>TC–>ET), leading to more consistent and clinically meaningful segmentation results. In addition, the introduction of sub-region-aware prompt tuning enables the model to extract specialized textual features for each tumor sub-region, thereby improving semantic alignment between text and image. Furthermore, the proposed framework demonstrates clear and consistent performance gains over strong baselines in both Dice and HD95 metrics, with ablation studies validating the effectiveness of each component. However, it has some weakness: (1) Most techniques (prompt tuning, LoRA, attention-based modulation) are existing methods, making the contribution largely incremental. (2) Experiments are conducted only on the TextBraTS dataset, raising concerns about generalizability to other datasets or modalities. (3) The framework introduces multiple components (cascade decoder, prompt tuning, modulators), which may increase implementation complexity and computational overhead.
- Reviewer confidence
Somewhat confident (2)
- [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
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 proposes a method that utilizes a fine-tuned BERT to extract localized textual information, coupled with a multi-stage cascaded hierarchical fusion and prediction scheme. This approach effectively improves segmentation accuracy across all specific regions. The proposed framework demonstrates clear and consistent performance gains over strong baselines in both Dice and HD95 metrics, with ablation studies validating the effectiveness of each component.