List of Papers Browse by Subject Areas Author List
Abstract
Implicit neural representations (INRs) have emerged as a powerful framework for continuous image representation learning. In Functa-based approaches, each image is encoded as a latent modulation vector that conditions a shared INR, enabling strong reconstruction performance. However, the structure and interpretability of the corresponding latent spaces remain largely unexplored.
In this work, we investigate the latent space of Functa-based models for ultrasound videos and propose Low-Rank-Modulated Functa (LRM-Functa), a novel architecture that enforces a low-rank adaptation of modulation vectors in the time-resolved latent space. When applied to cardiac ultrasound, the resulting latent space exhibits clearly structured periodic trajectories, facilitating visualization and interpretability of temporal patterns. The latent space can be traversed to sample novel frames, revealing smooth transitions along the cardiac cycle, and enabling direct readout of end-diastolic (ED) and end-systolic (ES) frames without additional model training.
We show that LRM-Functa outperforms prior methods in unsupervised ED and ES frame detection, while compressing each video frame to as low as rank k=2 without sacrificing competitive downstream performance on ejection fraction prediction. Evaluations on out-of-distribution frame selection in a cardiac point-of-care dataset, as well as on lung ultrasound for B-line classification, demonstrate the generalizability of our approach.
Overall, LRM-Functa provides a compact, interpretable, and generalizable framework for ultrasound video analysis. The code is available at https://github.com/JuliaWolleb/LRM_Functa.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1186_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/1186_supp.zip
Link to the Code Repository
https://github.com/JuliaWolleb/LRM_Functa
Link to the Dataset(s)
EchoNet Dynamic dataset: https://echonet.github.io/dynamic/
BibTex
@InProceedings{WolJul_LowRankModulated_MICCAI2026,
author = { Wolleb, Julia AND Baloescu, Cristiana AND Durrer, Alicia AND Tagare, Hemant D. AND Papademetris, Xenophon},
title = { { Low-Rank-Modulated Functa: Exploring the Latent Space of Implicit Neural Representations for Interpretable Ultrasound Video Analysis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes LRM-Functa, a low-rank extension of Functa-based implicit neural representations for ultrasound videos. By constraining time-resolved modulation vectors to lie in a learnable low-rank subspace, the method aims to produce more compact, structured, and interpretable latent trajectories. A particularly interesting contribution is that the learned latent trajectories can be used to identify clinically relevant cardiac phases, such as ED/ES frames, without training an additional dedicated prediction model. The paper also validates the framework on multiple ultrasound tasks, including cardiac and lung ultrasound analysis.
- 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 this work is its clear motivation and meaningful problem setting. Rather than only improving reconstruction performance, the paper focuses on structuring the latent space of ultrasound video representations, which is both novel and relevant. The proposed method is simple yet elegant, and the low-rank design is well aligned with the periodic nature of cardiac motion. In addition, the paper shows good practical value by linking latent representation learning with clinically meaningful temporal analysis.
- 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.
No weakness found
- 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?
clear and meaningful motivation, elegance of the method design, and reasonably broad empirical validation.
- 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 presents a neural implicit video representation framework for cardiac ultrasound based on a shared 2D implicit neural function modulated time and data-specific descriptors. The descriptors are obtained by multiplying a shared temporal dictionary with fitted coefficients that form the compressed representation. The coefficients can then be post-processed to extract a 1D motion signal and estimate cardiac phases.
- 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 methodology is very well explained and the results are impressive. While the studied modality may be a particularly good fit for a low-rank approach (quasi-periodic cardiac motion as the main temporal variation), the model is generic enough that it could be adapted to a broad range of video settings. In particular, the interpretability of the motion traces gives an a posteriori justification for treating time as a modulation rather than using it as an additional input to the implicit neural function.
- 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 dense and some missing details can make it hard to read.
For example, it would be good to have a short explanation of the meta-learning schedule rather than just referring to it since it is the main training method of the paper.
In Fig. 5, Table 1 and 2, it is not clear what the rank k means for the baseline methods that do not contain a low-rank decomposition. If an equivalent compression rate is being used, then that should be in legends rather than k.
- 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
The paper states q=256 and k goes from 2 to 512.For k > q, doesn’t the model become redundant instead of low-rank?
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The method is solid and has good potential to be applied in other fields, so it should be relevant to a big part of the MICCAI attendees.
- 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 #3
- Please describe the contribution of the paper
This work introduces Low-Rank-Modulated Functa (LRM-Functa), a novel Implicit Neural Representation (INR) framework designed for ultrasound video analysis. Building upon the VidFuncta architecture, the proposed method enforces a low-rank constraint on the time-resolved latent modulation vectors. This structural constraint forces the model to learn highly compact and interpretable latent representations of temporal dynamics. When applied to cardiac ultrasound videos, the latent space naturally organizes into periodic trajectories (e.g., spirals or lines) that correspond to the cardiac cycle. This allows for the unsupervised identification of end-diastolic (ED) and end-systolic (ES) frames without any additional training. The authors evaluate their method on the EchoNet-Dynamic dataset, an out-of-distribution (OOD) POCUS dataset, and a Lung Ultrasound (LUS) dataset, demonstrating state-of-the-art unsupervised frame detection and robust downstream performance (ejection fraction and B-line classification) even at extreme compression rates (k=2).
- 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 major strengths are: 1.High Novelty and Interpretability: While INRs and the Functa framework have shown promise for medical image compression and reconstruction, their latent spaces are typically treated as black boxes. Imposing a low-rank constraint to extract physically meaningful, continuous temporal trajectories (cardiac cycles) is a highly elegant and novel contribution to medical imaging computing. 2.Strong Unsupervised Clinical Application: The ability to perform unsupervised ED and ES frame detection simply by traversing the 1D filtered signal of the latent trajectory is practically highly valuable. Achieving an ED MAE of 2.26, which is competitive with fully supervised methods, is an impressive result. 3.Robustness to High Compression: The paper investigates the compression-reconstruction trade-off. Demonstrating that downstream tasks (like Ejection Fraction prediction) maintain strong performance (MAE of 5.29) even when the temporal dynamics are compressed to just k=2 components per frame highlights the efficiency of the learned representations. 3.Comprehensive Evaluation: The inclusion of an OOD POCUS dataset and a completely different clinical task (LUS B-line classification) strongly validates the generalizability of the proposed architecture.
- 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 major weaknesses are: 1.Lack of Analysis of Computational Cost and Time. The manuscript lacks a discussion on the computational cost and wall-clock time required to encode a new video at test time. 2.Orthogonal vs. Basic Variants: The paper introduces two variants: LRM-Functa_b (basic) and LRM-Functa_o (orthogonal). However, the results are somewhat mixed. For instance, in Table 1 (EchoNet, k=2), the basic variant outperforms the orthogonal variant, whereas in Table 2, the orthogonal variant is superior. The manuscript would benefit from a deeper analysis or intuition regarding when the orthogonal constraint is beneficial or not.
- 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?
Please refer to the major strengths and weaknesses.
- 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
Author Feedback
We thank all reviewers for their positive feedback and helpful comments.
Response to Reviewer 1: We thank the reviewer for highlighting the clarity in motivation and meaningful problem setting, as well as the elegance of the method design and empirical validation.
Response to Reviewer 2: We thank Reviewer 2 for the positive evaluation in terms of applicability to a broad range of video settings, and interpretability of our method.
To address the point made by Reviewer 2 to provide more training details, in the camera-ready version, we will add a short explanation of the meta-learning schedule in Section 2.2: “We follow the meta-learning strategy used in [28]. The outer loop optimizes the shared parameters θ of M_θ and β of B_β, while the inner loop performs G steps of optimization over the modulation vectors v and {ϕ_t}_{t=1}^b. For each batch of b frames from a training video, the latent vectors v and ϕ are initialized to zero, followed by G=10 inner-loop updates of v and ϕ, and finally one outer-loop update of θ and β.”
In Fig. 5 and Tables 1–2, we will clarify the meaning of k. Across all methods, k denotes the dimensionality of the compressed representation per frame, ensuring equivalent compression rates. We will clarify this in the camera-ready version by adapting the legends accordingly. Thanks for pointing out that for k > q, the matrix B becomes redundant rather than providing a low-rank structure. We considered values as high as k=512 because this was the baseline setting for the VidFuncta architecture, which serves as a basis for this method.
Response to Reviewer 3: We thank Reviewer 3 for highlighting the high novelty and interpretability, the strong unsupervised clinical application, the robustness even for high compression rates, and the comprehensive evaluation.
As for the points raised by Reviewer 3: 1.To provide additional insight into the computational cost, we will add the following sentence in the camera-ready version in Section 4.2: “Encoding a new video at test time requires 2.28 GB of GPU memory and 27.5 seconds for a video of length T=200.”
2.We thank Reviewer 3 for this insightful observation about the orthogonal vs. basic variants. We will strengthen the manuscript by adding the following paragraph for further intuition regarding the differences between LRM-Functa_b (basic) and LRM-Functa_o (orthogonal) in Section 4.2 in the camera-ready version: “As shown in Fig. 3, LRM-Functa_b collapses the latent trajectory to an almost one-dimensional line, suggesting that cardiac motion can largely be represented with k=1 value per frame. This produces a clean projection onto the principal motion direction p, resulting in a clear filtered 1D signal s_filt and facilitating ED/ES frame detection. This explains the stronger performance of LRM-Functa_b in Table 1.In contrast, LRM-Functa_o preserves greater variability in the latent space through the orthogonal constraint, leading to slightly improved reconstruction fidelity, as shown in Figure 5.This becomes beneficial for downstream tasks on reconstructed videos, where preserving fine-grained visual information is important, explaining the improved performance of LRM-Functa_o in Table 2.”
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.
The core idea—enforcing a low-rank structure on an INR’s latent space to yield an interpretable, compact representation—is both novel and elegant. The unsupervised clinical application is practically valuable. Possible improvement could be analysis of computational cost and the mixed results of the two variants.