List of Papers Browse by Subject Areas Author List
Abstract
In multi-organ ultrasound (US) imaging, data variability is substantial due to heterogeneous organ-specific imaging characteristics and the presence of diverse tumor types. Existing masked autoencoder (MAE) approaches struggle to disentangle tumor-related features from organ-dependent variations and do not exploit textual priors about organs, thereby limiting their capacity to learn organ-invariant tumor representations. To address these challenges, we propose a novel disentangled adversarial masked autoencoder (DA-MAE), which partitions the encoder into distinct tumor and organ branches to capture complementary information. An organ-invariant adversarial training strategy is employed to suppress organ-specific information within the tumor branch, promoting organ-invariant tumor modeling. In parallel, text prompts provide organ-level priors that implicitly guide the decoder to reconstruct the input image, facilitating more accurate learning of organ-specific representations in the organ branch. Evaluated on five-organ US benchmarks, our method achieves superior performance in downstream classification with an AUC ranging from 78.12% to 99.97% and segmentation with a Dice ranging from 69.38% to 86.12%. Compared to the current best-performing MAE-based method, our method consistently improves classification AUC by 0.17–1.93% and segmentation Dice by 0.58–1.51%. Our method validates the effectiveness of disentangled representation learning with adversarial and text-guided strategies for accurate tumor analysis in multi-organ US imaging scenarios. The source code will be made publicly available via GitHub upon acceptance.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2499_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/yXiangXiong/DA-MAE
Link to the Dataset(s)
ExpUNet: https://plos.figshare.com/articles/dataset/The_datasets_and_the_code_of_the_expanded_U-Net_/14786491?file=28415533
BUSIS: http://cvprip.cs.usu.edu/busbench/
BUSI-WHU: https://data.mendeley.com/datasets/k6cpmwybk3/1
US3M: https://www.kaggle.com/datasets/timesxy/multimodal-breast-ultrasound-dataset-us3m
UDIAT: https://helward.mmu.ac.uk/STAFF/M.Yap/dataset.php
BCMID: https://zenodo.org/records/14970848
BUS-UCLM: https://data.mendeley.com/datasets/7fvgj4jsp7/1
BibTex
@InProceedings{XioXia_OrganInvariant_MICCAI2026,
author = { Xiong, Xiangyu AND Jia, Dengqiang AND Sun, Yue AND Yang, Hao AND Huang, Jiaju AND Zhang, Zhuoneng AND Tan, Tao},
title = { { Organ-Invariant Tumor Representation Learning via Disentangled Adversarial MAE for Multi-organ Ultrasound Imaging } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper presents a method for organ-invariant representation learning, using self-supervised learning with masked autoencoders. The self-supervised pretraining uses separate tumor and organ encoders, attempting to separate the organ-specific features using an organ-invariant adversarial loss. A reconstruction loss is supported with a CLIP-based text embedding. Afterwards, the features are evaluated in several ultrasound classification and 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 explains the proposed method fairly well: the components are clearly explained and there are some good arguments to support the proposed method.
-
The evaluation is quite extensive: several classification and segmentation tasks, and there is an ablation study to investigate the components. The proposed method outperforms the baselines.
-
I especially appreciated the analysis of some of the parameters of the method, such as the effect of the masking ratio (Fig. 3).
-
- 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.
-
While the method is evaluated on a number of datasets, there is essentially only one self-supervised training part that is used for all downstream tasks.
-
One could argue that the proposed method is a combination of several somewhat-related components: some orthogonal loss, some feature disentanglement, some adversarial learning, some prompt guidance, and together that seems to work for this application. It is difficult to point at a clear novel contribution: besides the finding that this specific approach works for this specific problem, what can we learn from reading this paper?
-
- 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
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?
While the paper is easy to follow and the proposed approach seems to outperform the other methods, it also feels like a combination of several reasonably standard things: the paper is missing an exciting idea, analysis, or conclusion.
- 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.
I’d like to thank the authors for the rebuttal. I think their answers do not remove the weaknesses of the paper, but the overall submission is sufficient to recommend acceptance.
Review #2
- Please describe the contribution of the paper
The authors proposed DA-MAE that self-learns organ and tumor features independently in the pre-training phase, and then it learns to classify tumors, invariant of organs, and to produce a tumor mask.
- 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.
- Use of separate encoders, aiming to learn organ-invariant tumor and organ characteristics.
- Use of many datasets in validation, involving 5 organs.
- 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.
- Differences between the proposed text prompt and that of URFM are not clear.
- The ablation study shows better performance by about 2% for using Distentangling Dual-Encoder, which could be a result of many intrinsic hyperparameter-related reasons other than due to Dual-Encoder. Without seeing a t-SNE or UMAP plot of features, it cannot be accepted as a certain advantage.
- 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?
- Doubtful about the performance of the distentangling dual-encoder (major novelty claim).
- 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.
I am convinced that the 2% gain is attributed to the disentangling dual-encoder, which was my main concern.
Review #3
- Please describe the contribution of the paper
The authors tackle the disentanglement of tumour commonalities from organ-specific features by proposing feature decorrelation and organ-invariant adversarial training, which suppresses organ-specific information in the tumour encoder’s feature extraction using a masked autoencoder. The authors performed extensive experimental evaluations, including comparative evaluations with SOTA methods, using a five-organ pre-training dataset and downstream task-specific datasets. Across different downstream tasks, such as benign and normal tumour classification and segmentation, and thyroid nodule segmentation, the proposed tumour encoder, which extracts disentangled tumour features, improves both classification and segmentation accuracy.
- 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 authors challenge to disentanglement of tumour commonalities from organ-specific features. -A novel combination of decorrelation and adversarial training is interesting. -The authors proposed orthogonal loss for decorrelation of tumour commonalities and organ-specific features. -Extensive evaluation, including an ablation study, with a five-organ pre-training dataset and four downstream task-specific datasets. -The authors made a five-organ pre-training dataset by collecting 21 publicly available datasets. -The authors made downstream task-specific datasets for breast, thyroid, ovary, and gallbladder datasets by using several publicly available datasets. -Comprehensive figures, tables -Convincing experimental results. -Fair and detailed presentation for the work’s repeatability.
- 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.
-About a gradient reversal layer, the authors should explain what the layer actually does (e.g., inversion by multiplying by a negative scalar). -Slight improvement of classification and segmentation accuracy
- 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 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.
(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?
Based on the above-mentioned major strengths, I think the submission is already ready for the MICCAI presentation regardless of the major weaknesses.
- 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’ rebuttal responses disabuse me of my misinterpretations, which I had listed as the major weaknesses. Now, based on the listed strengths of this work, I conclude that the submission is ready for acceptance.
Author Feedback
We thank all reviewers (R) for their valuable feedback. We response as follows with a Q&A format.
Q1: A single self-supervised training scheme for all downstream tasks. (R1) A1: This is our advantage to disentangle common tumor patterns from organ-specific styles in one shared feature space. If separate self-supervised parts were used per organ, the model could not capture shared semantic invariance and would lose cross-organ transferability. Existing mainstream methods, such as USFM (Jiao et al. MedIA 2024), DeblurringMIM (Kang et al. MedIA 2024) and Uni4Eye++ (cai et al. TMI 2024), also use only one self-supervised training part. On downstream datasets, our method surpasses USFM by an average of 2.35% (AUC).
Q2: Combining existing modules with no clear novelty. Aside from working for this specific task, what insights can be learned. (R1) A2: All components in our method are not simply combined modules, but complementary designs centered on dual-encoder feature disentanglement. Each module plays an indispensable role in disentangling task-shared pathological invariants from domain-specific anatomical confounders, and ablation experiments verify their progressive complementary effects. Individual and incremental integration of our core designs brings consistent performance gains, with AUC improved by 1.71%–6.31% and Dice increased by 1.13%–4.96% across different module combinations. Beyond the specific ultrasound analysis task, our common-specific disentanglement paradigm is generic and transferable. It can be readily extended to CT, MR and other clinical imaging scenarios that inherently contain shared pathological characteristics and domain-specific anatomical variations.
Q3: Differences from URFM’s text prompt. (R2) A3: Firstly, we utilize organ-level semantic priors to disentangle organ and tumor features. This solves the issue of entangled feature representation in URFM. Secondly, we integrate text prompts into the decoder with dual-branch features to enable organ-aware pixel-level reconstruction. This avoids the feature aliasing problem caused by single-branch text guidance of URFM. Finally, our prompts are tightly coupled with disentangled encoders, orthogonal loss and adversarial training. This addresses the limitation of isolated text guidance in URFM. Benefiting from these differences, our method achieves 0.17%–1.93% higher classification AUC than URFM, with average AUC improved by 1.29%, across five ultrasound benchmarks.
Q4: Without t-SNE/UMAP visualization, the performance gain of Disentangling Dual-Encoder in ablation may come from hyperparameter tuning. (R2) A4: All ablation experiments adopt identical hyperparameters and training settings, so the 2% gain is attributed to the Disentangling Dual-Encoder. We will add UMAP feature visualization if page allows. The UMAP shows the inter-organ feature distance and intra-tumor feature compactness. Compared with baseline MAE, our dual-encoder design reduces intra-class tumor feature dispersion and enlarges inter-organ feature distribution overlap in the tumor representation space.
Q5: Explanation of gradient reversal layer’s mechanism. (R3) A5: We have explained this in Page 5, Section “Adversarial Loss”: “During the forward pass, the GRL acts as …, During backpropagation, …”. GRL performs an identity mapping in the forward pass, while in the backward pass it inverts the gradient flow by multiplying the gradient with a negative scalar of −1.0 in our implementation.
Q6: Slight improvement of classification and segmentation. (R3) A6: On five classification tasks, our method on average outperforms SOTA URFM (IScience, 2025), by 1.29% (AUC), 4.02% (F1), 4.94% (Recall), and 3.26% (Precision). Specifically, on the typical BUSG and BUSI datasets, our method outperforms URFM by 7.91% higher Recall, 7.80% higher F1, and 8.22% higher Precision. For segmentation tasks, specifically on the OKUS dataset, our method outperforms SOTA UA-AE (BIBM, 2025) by 1.12% higher Dice.
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 work proposes organ-invariant tumor representation learning for multi-organ ultrasound image analysis by developing disentangled adversarial MAE. The recommendation is mixed among reviewers. The main disagreements are on empirical evaluation, methodology novelty and effectiveness, and lack of sufficient evidences and explanations for some claims. Based on my reading, I think that the reviewers raised effective concerns. Please try to fix these concerns in the rebuttal.
- 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.
All three reviewers agree to accept. I read the rebuttal and reviewers’ comments after the rebuttal. I think that the rebuttal solves main concerns, and found not inappropriate comments from the reviewers.
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.
Reviewers have a consensus on the positive merits of this 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.
All of the 3 reviewers agree to accept in the post-rebuttal period.