List of Papers Browse by Subject Areas Author List
Abstract
Fetal cardiac segmentation frequently suffers from performance degradation under domain shifts induced by heterogeneous scanners and rapid physiological motion. Standard test-time adaptation (TTA) methods typically rely on generic consistency or entropy minimization, which lack explicit anatomical constraints and fail to preserve biological plausibility under complex cardiac deformations. We propose DeformTTA, a physiology-driven TTA framework that bolsters segmentation robustness by explicitly modeling cardiac dynamics. Unlike domain-agnostic augmentations, DeformTTA leverages a deformation network to simulate realistic fetal heart movements, integrated with a generative model for authentic texture synthesis. Furthermore, a morphology-aware shape prior is introduced to regularize the adaptation process, ensuring the structural integrity of the predicted segmentations. By optimizing for motion-conditioned consistency and morphological alignment, DeformTTA effectively generalizes to target domains during inference. Multi-center experiments demonstrate that DeformTTA significantly outperforms state-of-the-art baselines, particularly in segmenting complex pathological cases. Code is available at https://github.com/SplinterLi/DeformTTA.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1543_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/SplinterLi/DeformTTA
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiZhi_DeformTTA_MICCAI2026,
author = { Li, Zhihao AND Liu, Zelong AND Liu, Juhua AND Xu, Yongchao AND Du, Bo},
title = { { DeformTTA: Morphology-Aware Controllable Deformation for Test-Time Adaptation in Fetal Cardiac Ultrasound Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16889},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
(1) Proposes DeformTTA, a physiology-driven TTA framework for fetal cardiac ultrasound segmentation. It integrates explicit anatomical constraints with generative models to address cross-center domain shifts and pathological variations. (2) Introduces a morphology-aware regularization mechanism using a BR coefficient. This serves as a mathematical prior to penalize biologically implausible topological deformations during unsupervised inference. (3) Develops a structure-centric, controllable morphological deformation module. By operating on Gaussian-smoothed distance fields and utilizing an anatomical gating network, it simulates non-rigid physiological movements for data augmentation. (4) Integrates a ControlNet-based conditional generative model into the TTA pipeline. It uses the generated deformation masks to synthesize target-domain style images, reducing interpolation artifacts from standard spatial warping and allowing optimization on consistent cross-view pseudo-samples.
- 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) Utilizes the BR coefficient as a morphological constraint to maintain the topological integrity of the ventricles and atria, mitigating the boundary collapse issues often observed in standard entropy minimization methods. (2) Isolates high-frequency speckle noise from deformation calculations by employing one-hot Gaussian-smoothed fields to guide displacement field generation. (3) Implements an effective “deform-then-generate” strategy. By using a fine-tuned ControlNet, it converts geometric deformations into realistic ultrasound images for adaptive optimization. (4) Demonstrates strong quantitative performance on a 650-case multi-center dataset. The method performs favorably against established domain generalization algorithms and recent TTA frameworks (including TEGDA), achieving a Mean Dice of 76.38% and reducing HD95 to 4.888 mm.
- 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) High Inference Latency: The generative diffusion pipeline (which generates 4 augmented samples and performs 2 gradient steps per test image) adds approximately 3 to 5 seconds of processing time per case. This limits its feasibility for real-time intraoperative clinical applications, restricting its use primarily to offline analysis.
(2) Cold-Start Dependency: The TTA process relies heavily on the accuracy of the initial segmentation prediction from the pre-trained baseline model (nnU-Net). In scenarios with severe domain shifts (drastic probe angle changes or heavy acoustic shadowing), the initial mask may fail entirely. The framework currently lacks an explicit rejection sampling or self-correction mechanism to prevent subsequent error propagation.
(3) Underutilization of Temporal Dynamics: The current approach performs independent adaptation based on 2D static images, treating end-diastolic and end-systolic frames as isolated instances. Despite having a deformation module that simulates motion, the framework does not leverage the inherent temporal consistency present in continuous ultrasound video frames.
- 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?
(1) Methodological Justification: The paper presents a logically sound TTA pipeline by combining anatomical priors (BR coefficient) with conditional generative models (ControlNet). Estimating displacement via Gaussian distance fields effectively bypasses speckle noise and models physiological myocardial scaling.
(2) Empirical Justification: The method demonstrates measurable robustness against pathological domain shifts, notably reducing the HD95 metric and performing well on complex congenital heart disease samples (HLHS) compared to recent baselines.
- 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.
After reading the authors’ rebuttal, I believe they have adequately addressed my main concerns. Overall, I remain positive about the work and support acceptance.
Review #2
- Please describe the contribution of the paper
The author proposed a new framework to tackle the problem of data distribution shift for the problem of fetal echocardiography segmentation. The core idea is to make test-time adaptation anatomy-aware by combining a morphology prior, controllable deformation, and ControlNet-based image synthesis before per-test optimization. The proposed method enables performance gain for segmentation models.
- 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 problem is important and well motivated, where the different aspect of fetal cardiac ultrasound may affect the i.i.d assumption.
- The method is domain-informed rather than generic. Using a boundary-regularity shape descriptor, Gaussian-smoothed structural fields, controllable deformation, and motion-conditioned synthesis is more tailored to fetal heart dynamics than plain entropy minimization
- The ablation is directionally useful and suggests the deformation module is important, the qualitative figure on page 7 is focused on clinically relevant abnormal cases rather than only easy examples.
- 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 find the dataset is a bit confusing, the dataset contains scans from two hospitals and also with normal/pathological. THis makes it really hard to tell where the performance gain actually come from.
- The ablation text states ControlNet removal changes HD95 from 24.31 to 59.61 mm, but Table 2 reports 4.888 to 11.92 mm. Those cannot both be correct.
- “MedSAM [7]” is mentioned, but reference [7] is the LoRA paper, not MedSAM. That is a concrete reference error.
- The table 2 shows w/o shape prior regularization, but the objective function doesn’t have any term for this one.
- 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.
(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 good application, a sensible domain-specific idea, and strong-looking reported gains. But at the current state, I do not think it clears the rigor bar for a top-tier acceptance because too many important pieces are either under-specified or internally inconsistent. With a cleaner and fully specified method section, a disentangled experimental design, corrected tables/references, and stronger statistical evaluation, this could become a competitive MICCAI paper.
- Reviewer confidence
Somewhat confident (2)
- [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 rebuttal addresses my concerns.
Review #3
- Please describe the contribution of the paper
1.The paper proposes DeformTTA, a novel physiology-driven test-time adaptation (TTA) framework designed for fetal cardiac ultrasound segmentation under domain shifts. 2.It introduces a Boundary Regularity (BR) coefficient to explicitly quantify anatomical morphology, which serves as a standardized shape prior to regularize the adaptation process. 3.It develops a Controllable Morphological Deformation module that operates on structure-centric Gaussian-smoothed fields. This simulates realistic fetal heart movements while avoiding texture-based overfitting. 4.It integrates a ControlNet-based generative model to synthesize high-fidelity, target-style imagery conditioned on the deformed masks, allowing the model to optimize for motion-conditioned consistency and morphological alignment.
- 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 integration of explicit physiological cardiac dynamics into the TTA process is innovative. Moving away from traditional domain-agnostic generic augmentations or purely entropy-based TTA methods, this work elegantly uses biological priors to prevent catastrophic failure under complex deformations. 2.The introduction of the Boundary Regularity coefficient provides a specific, quantifiable morphological constraint. Formulated as a radial coefficient of variation, it establishes a standardized shape prior that explicitly prevents the adaptation process from yielding biologically implausible shapes. 3.The framework integrates a ControlNet-based generative model into the TTA pipeline to resolve interpolation artifacts common in traditional geometric warping. Conditioned on the deformed masks, the generative model synthesizes artifact-free, target-style images.
- 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.As acknowledged by the authors in the conclusion, the generative TTA process is computationally expensive, adding 3-5 seconds of processing time per case. In the context of ultrasound, which is inherently a real-time imaging modality, this computational bottleneck severely limits immediate clinical deployment. 2.The framework relies on an initial mask to generate structural fields. If the initial model prediction is flawed, the morphology-aware shape prior and the subsequent deformations will be misguided. The paper lacks a fallback or self-correction mechanism to recover from catastrophic initial failures. 3.The methodology does not provide sufficient detail on the pre-training or fine-tuning process of the ControlNet and the Deformation Network. 4.The comparison method is somewhat outdated, with only one from 2025.
- 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.
(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 proposed DeformTTA framework introduces a valuable conceptual shift by explicitly incorporating physiology-driven constraints for fetal cardiac ultrasound segmentation under domain shifts. Furthermore, empirical results on a multi-center dataset demonstrate solid improvements over state-of-the-art test-time adaptation baselines. However, there are unaddressed practical limitations: the significant computational overhead restricts its deployment for real-time ultrasound, and the strict dependency on initial prediction quality poses unresolved robustness concerns.
- 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.
N/A
Author Feedback
We sincerely thank all reviewers for their constructive feedback, and address the concerns below.
1.[R1-W1/R3-W1] Inference latency The computational overhead of our TTA process comes from two parts: 1) Image generation, which is the most time-consuming part. In practice, we can optimize this part by using xFormers + DDIM 2-step + fp16 + torch.compile, reducing time from 1172.7 ms to 451.8 ms per image (2.60x). For 4 representative frames, the runtime drops from about 4.8 s to 1.8 s, with negligible accuracy change (Delta = -0.01%). 2) Model parameter updating, which requires 142 ms for our 2 optimization steps. Furthermore, we can also adapt our method to new domain under per-domain TTA setting. Precisely, we can collect a set of images from the new domain without annotation, and adopt our DeformTTA once offline before online usage on the new domain. This helps to improve the performance on new domain without extra online cost.
2.[R1-W2/R3-W2] Cold-Start Dependency We further statistically analyze the results by stratifying the initial masks into three categories based on Dice score: Low (<50%), Medium (50%-75%), High (>75%). Our DeformTTA improves the performance for all tiers, even with more gains on the Low tier. This is probably because of two reasons about the generation part in our DeformTTA: 1) The deformation network has chance to make the pseudo-labels closer to optimal segmentation; 2) The diffusion model with ControlNet can generate samples with similar intensity distribution with the new domain, improving the generalization ability of the segmentation network. This explains why our DeformTTA is also effective when initial pseudo-labels are relatively poor.
3.[R1-W3] Temporal dynamics We agree video-level temporal consistency is important. The current 2D design is constrained by available labels (annotated ED/ES frames without dense temporal correspondence). We will explicitly state this limitation and discuss optical-flow-guided video adaptation as future work.
4.[R2-W1] Dataset design clarity Our split is intentionally challenging clinical and two-axis shifted: source training set = 200 normal 4CH scans from Hospital-1 (screening center); test set = 450 CHD 4CH scans from Hospital-2 (referral center). Shift comes from both center (scanner/protocol) and pathology (unseen structural abnormality).
5.[R2-W2] HD95 inconsistency: 24.31->59.61 vs 4.888->11.92 We apologize for this mistake. The in-text values were pixel-space HD95 before physical calibration; Table 2 reports clinical HD95 in mm. We will correct this inconsistency and report HD95 consistently in mm.
6.[R2-W3] Citation error (MedSAM vs LoRA) You are correct. The citation for “fine-tuned MedSAM” should be MedSAM (Ma et al., 2024); LoRA should be cited separately for the fine-tuning strategy. We will correct this into “fine-tuned MedSAM [Ma et al., 2024] with LoRA [7]”, and review the remaining references.
7.[R2-W4] Table 2 vs method naming mismatch The row “w/o Shape Prior Regularization” actually means removing the smooth loss in Eq. (3) that implements the BR-coefficient-based shape prior. We will rename it to “w/o Smooth Loss (Shape Prior Reg.)” and align wording in the method section for better clarity.
8.[R3-W3] Missing training details The details of pre-training the deformation network and fine-tuning ControlNet are as follows:
- Deformation network: lightweight U-Net (3 encoder/decoder stages, 32 base channels), predicts displacement (u_x, u_y) from Gaussian-smoothed one-hot distance fields, 100 epochs, Adam lr=3e-4, weight decay 1e-4.- ControlNet: SD1.5 initialization, 5,000 steps, source-domain image-mask pairs (200 cases), batch 4, AdamW lr=1e-5, 256x256, RGB-encoded 7-class mask condition. We will incorporate these details in the revision.
9.[R3-W4] Comparison with more methods Excellent suggestion. We have compared to the most related TEGDA [31]. Per rebuttal policy, we cannot provide more comparative results in this response.
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 proposes a physiology-driven test-time adaptation framework for fetal cardiac ultrasound segmentation. Reviewers find the problem clinically important and appreciate the domain-informed design and strong multi-center results. However, they raise concerns about high inference latency that limits real-time use and the strong dependence on initial mask quality. They also note inconsistencies between the tables and text, a misattributed reference, and missing training details. Given these mixed assessments, the paper is invited to rebuttal for the authors to address the concerns.
- 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 reviewers initially accepting and one weakly rejecting due to internal inconsistencies, missing training details, and a reference error. The rebuttal effectively clarified these issues, including the HD95 inconsistency, citation error, and cold-start robustness, leading all three reviewers to accept post-rebuttal.
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.
This paper proposes a test-time adaptation framework for fetal cardiac ultrasound segmentation. All reviewers agree that the method has merit and that the authors addressed all raised concerns in the rebuttal. I agree with them and recommend acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The rebuttal was effective: the authors satisfyingly answered most concerns and addressed a few points that needed clarifications. I therefore recommend acceptance. However, although reducing generation time from 1 to 0.5s is better, it’s not going to permit intra-operative or real time applications, where the time threshold is usually around 200ms for perceptual visual continuum. Moreover, it’s a bit cheap to say “we have run a new baseline but cannot show the results”. This type of claim should be banned because they are not supported by evidence. Nonetheless, as said above, I still find the rebuttal convincing, and the general framework is appealing and well evaluated. I will thus recommend highlighted poster.
