List of Papers Browse by Subject Areas Author List
Abstract
Self-supervised learning (SSL) has emerged as a powerful paradigm in medical image analysis, but existing SSL frameworks for multi-sequence MRI often fail to exploit inherent cross-sequence structural representations or remain agnostic to pathological heterogeneity. In this paper, we propose Text-guided Invariant Anatomical Learning (TeInL), a novel SSL framework for multi-sequence brain MRI. TeInL decouples image representations into sequence-invariant anatomical features and sequence-specific style features, enforces structural consistency by text-image alignment, learns anatomical consistency, and performs cross-sequence reconstruction. We integrate anatomical phenotypes from brain atlases and pathological signatures from unsupervised lesion segmentation into our SSL framework via image-text alignment. Aligning anatomy and pathology-aware textual descriptions with visual features enables the model to capture fine-grained correspondences between complex structural phenotypes and semantic representations. Extensive experiments across six benchmark datasets covering multi-granularity brain parcellations, stroke lesions, and brain tumors demonstrate that TeInL consistently outperforms state-of-the-art SSL methods and reduces annotation efforts by at least 40%. Code and pre-trained models will be available at https://github.com/MICCAI2026/MICCAI2026_2265.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2265_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/liangjianwen01/MICCAI2026_TeInL
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiaJia_Learning_MICCAI2026,
author = { Liang, Jianwen AND Cheng, Pujin AND Tang, Xiaoying},
title = { { Learning Invariant Anatomical Representations from Multi-Sequence MRI for Brain Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a self-supervised framework for multi-sequence brain MRI that learns sequence-invariant anatomical representations by combining cross-sequence reconstruction, anatomical consistency learning, and text-guided image-text alignment. The key idea is to use atlas-derived and pathology-aware textual descriptions, together with pseudo labels from unsupervised lesion segmentation, to supervise pre-training and improve downstream brain structure and lesion segmentation.
- 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 is well written and has a clear motivation: reducing reliance on dense voxel-level annotation while leveraging the complementary information in multi-sequence MRI. The method is technically rich and clinically relevant, and the reported gains are strong across six benchmarks, with consistent improvements in both structure and lesion segmentation.
- 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 claim that the method reduces annotation cost needs clearer justification. From the description, the training pipeline relies on pseudo labels generated by LF-SynthSeg for lesion-aware supervision, so it is not fully clear in what sense the annotation burden is reduced compared with a standard supervised or teacher-student setup. In particular, it would help to explain whether the reduction refers to fewer manual labels, weaker labels, or less expert annotation effort overall. Moreover, the use case of the model is also not fully clear to me. Since the method is presented as a pre-training framework for downstream segmentation, the paper should better explain how the learned representations are intended to be used in practice after training, and what concrete clinical or research workflow they enable. Without that, the practical value feels somewhat indirect even though the technical results are strong.
- 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
1.The paper states that TeInL reduces annotation cost. Could the authors clarify exactly what annotation cost is being counted here, and how this is measured? 2.Since LF-SynthSeg is used to produce pseudo labels, should TeInL be viewed as a teacher-student-style framework, or is the intended interpretation different? 3.What is the main practical use case of the learned model after pre-training? Is the goal mainly initialization for downstream segmentation? 3.How sensitive are the results to the quality of the pseudo labels and to the pathology-aware text descriptions?
- 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?
Although some parts of the annotation-cost claim and downstream use case are not fully convincing to me, the overall contribution is still strong and the empirical evidence is reasonably convincing.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The rebuttal clarifies the role of LF-SynthSeg pseudo labels and the intended use of the framework as a pre-training initialization strategy. The practical impact relative to the added methodological complexity is still up for debate. Additionally, for a foundation model, including a benchmarking under varying %s of labeled downstream training data would be a strong way to demonstrate label efficiency and the value of the learned representations.
Review #2
- Please describe the contribution of the paper
The paper presents a pretraining strategy for multi-sequence images that disentangles anatomical content from sequence-specific style information. The method uses single-modality image masking and learnable style prototypes to pretrain the model by cross-sequence reconstruction, where the target style is injected to reconstruct the anatomical content. To preserve anatomical consistency across sequences, the method minimizes the distance between anatomical features extracted from different sequences. Finally, the image features are aligned with text descriptions of the anatomical structures, taking into account the health status of the structure. All of these strategies result in a stronger pretraining that improves downstream performance on brain structure and lesions segmentation tasks.
- 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 presents an interesting idea. Multi-sequence studies are very common but there are not many methods that can leverage these different images or that can handle some of them being missing.
- The design choices are well motivated. The proposed method combines several strategies to learn a representation space that disentangles anatomical content from sequence-specific style while aligning image features with textual descriptions.
- The results show that the proposed method improves performance over state-of-the-art pretraining strategies, and the ablations demonstrate that all the components of the approach contribute to this improvement.
- The explanation of the pretraining method is clear and easy to follow, and the results are in general well organized and presented.
- 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.
- Some important methodological details are missing: – The method is pretrained using crops from a single image sequence, but it’s not clear how it is fine-tuned for downstream tasks or how it is evaluated. As stated in the introduction, one of the limitations of existing multi-sequence approaches is that they rely on a full set of sequences and, based on the proposed pretraining, it could be expected that this approach doesn’t suffer from that limitation. However, it is not clear how many sequences are used after pretraining or how the model would behave if some were missing. – It is not clear if the text descriptions and style prototypes are used only during pretraining or if they are included in the next steps as well. – How are the style prototypes initialized? – The style matching loss is not defined but described as being similar to some recently introduced loss. However, it is not clear if it refers to the matching loss (eq. 6) or the anatomy-text alignment loss (eq. 4).
- Figure 1 is quite complex and has no description to help understand it.
- (minor) The results in Tables 1 and 2 follow the same dataset ordering, but Figure 3 presents the datasets in a different order.
- 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.
(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 presents an interesting approach to multi-sequence pretraining and the presented method demonstrates good performance, superior to other state-of-the-art strategies. The main weakness of the paper is that some important explanations are missing, making it unclear how the method is applied to downstream tasks and how the results are obtained. I understand that this might be caused by the limited space in a MICCAI paper, but I believe some of those details are critical for understanding the approach. Therefore, if those clarifications could be included in the manuscript, I would be willing to raise my score.
- 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 addressed most of the concerns. With the extra clarifications, this work could represent a valuable contribution for MICCAI
Review #3
- Please describe the contribution of the paper
This paper proposes a novel framework for self-supervised learning to supplement brain segmentation tasks. The proposed framework incorporates a wide range of approaches to produce generalizable and steerable representations by considering style-matching, anatomical consistency, and contrastive image-text learning.
- 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 has significant technical components which are combined into a cohesive framework. The resulting framework produces improvements over strong baselines, and the ablation studies show that the added complexity is justifiable. The study is well designed and has a thorough evaluation, considering segmentations for brain anatomy for healthy/diseased patients, tumor, and stroke lesions. It was particularly impressive that improved performance could be reached with significantly less annotation efforts.
- 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 is very digestible, but some key details are unclear. I see that the intensity head is changed for a segmentation head after pretraining, but I think this detail should not be tucked under the Datasets/Implementation section, since it is easy to miss. Additionally, this should be made more explicit in Figure 1, either within the Figure, or at least in the caption. The style for Figure 3 is correct, but the fonts are too small. Future work should evaluate on low-data domains such as BraTS-SSA where pretraining is posed to provide even more advantages.
- 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?
The paper had a strong technical contribution through a complex framework that incorporated insights from prior work. The results demonstrate that the framework has improved results over the compared baselines, and the ablation studies motivated each component. The presentation was well-organized and the paper was easy to read, with a few improvements required.
- 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 experimental design is strong, and all reviewers suggested minimal improvements. Suggestions were properly addressed by the authors.
Author Feedback
We sincerely thank the reviewers for their valuable feedback and encouraging comments on the methodological design, pre-training framework, downstream segmentation performance, and clarity of presentation of our work. Below, we address the main concerns point by point: R1&MR: Annotation cost reduction. Similar to most existing foundation models, TeInL leverages large-scale multi-sequence brain MRI pre-training to learn robust anatomical and pathology-aware representations. This allows downstream segmentation on six benchmark tasks to achieve high performance using only limited annotated data, substantially reducing expert labeling effort. R1: Use of LF-SynthSeg pseudo labels. The LF-SynthSeg pseudo labels are used only during pre-training to provide pathology-aware textual guidance, ensuring tumor-containing samples are described in a clinically consistent manner. LF-SynthSeg is an unsupervised method and does not use any tumor labels for training. These pseudo labels do not constitute a teacher-student framework and are not used in downstream tasks. We will clarify this in the final version. R1&MR: Practical use of the learned model. The pre-trained TeInL model is intended solely for initializing network parameters in downstream tasks. We will clarify this in the final version. R1: Sensitivity to pseudo label and text quality. While the tumor pseudo labels generated in an unsupervised manner do not guarantee precise tumor boundaries, they provide sufficient localization for generating pathology-aware textual descriptions. Moreover, the pre-trained text encoder is robust to diverse medical text formats. We will perform comprehensive sensitivity analysis in our journal extension version. R2&MR: Important methodological details. We will clarify these important technical issues in the final version. 1.Unlike existing multi-sequence pre-training methods that require a full set of imaging sequences, TeInL only requires each sample to have at least two sequences to enable the anatomical consistency constraint. For downstream segmentation tasks, there is no requirement of the total number of input sequences (a single sequence also works). 2.The text descriptions and style prototypes are used only during pre-training to guide representation learning. For downstream tasks, only the pre-trained model weights are used for initialization. 3.The style prototypes are randomly initialized and updated during pre-training via a momentum-based strategy. 4.The style matching loss has the same form as the anatomy-text matching loss (Eq. 6), but operates on style prototypes and their corresponding textual descriptions. Specifically: LMS=-1/K ∑k [ys log(σ(D(pk,ht(Ak))))+(1-ys)log(1-σ(D(pk,ht(Ak))))], where K is the number of style prototypes, Ak is the textual description of the k-th MRI sequence’s intensity style, ys is the style label, D is the similarity function, and σ is the sigmoid activation function. R2: Figure 1 clarity. We will provide detailed descriptions in the caption of Figure 1 to improve readability and facilitate understanding. R2: Dataset ordering consistency. We will adjust the arrangement of results in Figure 3 to align with the dataset ordering presented in Table 1 and Table 2.R3: Segmentation head detail. We will move the description of replacing the intensity head with the segmentation head from the Datasets/Implementation section to the Methods section to improve readability. We will also update Figure 1 and its caption to explicitly reflect this change. R3: Font size in Figure 3.We will increase the font size in Figure 3 to improve readability. R3: Evaluation on low-data domains. We will include BraTS-SSA as an additional downstream segmentation task in our journal extension version to further validate the performance of our pre-training framework.
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 presents a self-supervised learning framework for multi-sequence brain MRI representation learning. It is technically rich and generally well written (R1R3), with strong experimental validation across multiple datasets and tasks (R1R2R3). The rebuttal should address remaining unclarities in the method description, including how the framework is used for downstream tasks (R1R2), how the method will behave under missing-sequence settings (R2), and the justification for the claimed reduced annotation cost (R1).
- 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.
Two of the three reviewers accept the paper, and most initial concerns have been clearly addressed in the rebuttal. Regarding annotation-cost efficiency, I agree that benchmarking with varying percentages of labeled data would provide stronger validation, perhaps in an extended version. However, Figure 3 already provides some insight into this point.
Additional post-rebuttal feedback: the paper appears to be a solid technical extension of MCLIM [12] to the multi-sequence MRI setting. The baselines, as well as the proposed model itself, seem to struggle in the multi-sequence setting compared with MCLIM. The scratch performance in Tables 1 and 2 and the caption of Figure 3 should also be revised. However, these issues are not considered as fatal flaws. Overall, I recommend acceptance.
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 presents a technically rich self-supervised pre-training framework for multi-sequence brain MRI, with a strong empirical evaluation across multiple benchmarks. Reviewers initially raised concerns about missing methodological details, particularly how the pre-trained model is used downstream, the justification for reduced annotation cost, and behavior under missing sequences. In rebuttal, the authors clarified that the model only provides weight initialization, that pseudo-labels from an unsupervised lesion segmentation method serve solely for pre-training text guidance, and that the framework can operate with as few as two sequences during pre-training and even with a single sequence downstream. Reviewers #2 and #3 found these clarifications adequate and raised their decisions to Accept, citing the solid experimental design and the value of the contribution. Reviewer #1 remained unconvinced, questioning the practical impact relative to methodological complexity and noting the absence of a label-efficiency benchmark with varying percentages of labeled data. Despite this reservation, the paper’s technical strengths, comprehensive ablation studies, and the resolution of major clarity issues support its acceptance. I recommend accepting the paper.
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.
Although there are some flaws in the method and experiments, the novelty and motivation of the proposed method were recognized by reviewers.
