List of Papers Browse by Subject Areas Author List
Abstract
Deep learning methods have proven useful in medical image segmentation when deployed on independent and identically distributed (iid) data.
However, their effectiveness in generalizing to previously unseen domains, where data may deviate from the iid assumption, remains an open problem.
In this paper, we consider the single-source domain generalization scenario where models are trained on data from a single domain and are expected to be robust under domain shifts.
Our approach focuses on leveraging the spectral properties of images to enhance generalization performance.
Specifically, we argue that the high frequency regime contains domain-specific information in the form of device-specific noise and exemplify this case via data from multiple domains.
Overcoming this challenge is non-trivial since crucial segmentation information such as edges is also encoded in this regime.
We propose a simple regularization method, Lipschitz regularization via frequency spectrum (LRFS), that limits the sensitivity of a model’s latent representations to the high frequency components in the source domain while encouraging the sensitivity to middle frequency components.
This regularization approach frames the problem as approximating and controlling the Lipschitz constant for high frequency components.
LRFS can be seamlessly integrated into existing approaches.
Our experimental results indicate that LRFS can significantly improve the generalization performance of a variety of models.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/3740_paper.pdf
SharedIt Link: https://rdcu.be/dV55s
SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72117-5_62
Supplementary Material: N/A
Link to the Code Repository
https://github.com/kaptres/LRFS
Link to the Dataset(s)
https://liuquande.github.io/SAML/
BibTex
@InProceedings{Ars_Singlesource_MICCAI2024,
author = { Arslan, Mazlum Ferhat and Guo, Weihong and Li, Shuo},
title = { { Single-source Domain Generalization in Deep Learning Segmentation via Lipschitz Regularization } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15010},
month = {October},
page = {666 -- 674}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces a novel regularization method called Lipschitz Regularization via Frequency Spectrum (LRFS), which aims to improve the generalization of deep learning models in the context of medical image segmentation. It focuses on the challenge of single-source domain generalization scenario, presenting a theoretical framework that uses the Lipschitz constant to control the sensitivity of a model to high-frequency components, which are argued to contain device-specific noise and crucial information for medical image segmentation.
- Please list the main strengths of the paper; you should write about 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 proposed LRFS method provides a new perspective on handling frequency components to improve model robustness.
- The paper provides a thorough experimental setup with multiple models and datasets, which strengthens the validity of the results.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
- The choice of Lipschitz constants and regularization weights may be sensitive and could require tuning for different models or tasks.
- There is a lack of theoretical analysis on the relationship between the Lipschitz constant, frequency components, and generalization.
- Please rate the clarity and organization of this paper
Very 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.
- Do you have any additional comments regarding the paper’s reproducibility?
N/A.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
Please consider addressing the points listed in section 6.
- 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
Weak Accept — could be accepted, dependent on rebuttal (4)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the proposed Lipschitz regularization vis frequency spectrum is simple, the regularization technique is model-agnostic, making it widely applicable to various deep learning architectures. The effectiveness of the method has been validated through rigorous experiment, and the paper is also well written and clearly presented.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #2
- Please describe the contribution of the paper
This paper introduces a novel method based on Lipschitz Regularization to tackle the single-source domain generalization segmentation problem. The method is motivated by the spectral properties of images. The authors proposed that the sensitivity of the model’s latent representations to the high-frequency components should be reduced because they argue that these high-frequency components contain domain-specific information in the form of device-specific noise. In addition, middle-frequency components are considered to pertain to the structural information that shows domain-invariant characteristics. Then a Lipschitz regularization method is raised to limit the sensitivity to the high frequency while encouraging the sensitivity to middle frequency components. They use the Lipschitz constant as an indicator of a model’s sensitivity and regularize it by means of approximations of this quantity. This is an effective approach that can be seamlessly integrated into existing frameworks.
- Please list the main strengths of the paper; you should write about 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.
Clarity of paper structure: This paper is well-organized. The proposed method is firstly overviewed in the introduction and then explained detailedly in the methodology section. The experiment section is disassembled into datasets, implementation details, and results. Each part is demonstrated in a clear way.
Clarity of problem definition and method explanation: The details are well explained in the paper. The authors use clear notations and formulas to define the problem and explain the method. They give abundant and strict arguments to prove the rationality and feasibility of the proposed method. Readers can reproduce the whole framework easily.
Innovation and depth: Improving the sensitivity of the model to middle-frequency information is innovative in comparison to existing Fourier-based methods. The authors also take a deep discussion and analysis of this issue.
Experiments result: The experiment section is robust. The authors pick some popular baseline frameworks to integrate with the proposed method as a comparison. The proposed method acquires prominent improvement.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
Visualizations: Too few figures to help understand the proposed method. Fig 1 in the paper is unclear. Readers may not identify the differences between samples. Extra description is also necessary to tell what are symbols I, v, and HFR that are mentioned in the following sections. Some more figures are suggested to demonstrate the workflow of the method.
Too few experiments: Although the experiment results are significant, there are too few experiments. Since the proposed method is not limited only to prostate MRI datasets, the authors can test their methods on some more datasets such as fundus and chest X-rays. Results only on one dataset are not persuasive.
Too few references: As a conference paper, there can be more references. Single domain generalization problem on medical images has been studied for many years, there are a large amount of methods. More references can make the paper more persuasive.
- Please rate the clarity and organization of this paper
Very 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.
- Do you have any additional comments regarding the paper’s reproducibility?
N/A
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
-
As mentioned in the weakness, why did you only test the method in one prostate MRI dataset? The Lipschitz Regularization based single DG method can be applied to all medical images theoretically.
-
Why improving the sensitivity to middle-frequency components can solve the problem? You argue that middle-frequency components pertain to the structural information that shows more domain-invariant characteristics while lacking domain-specific information. But in Fig 3, the low-level amplitude information is linear to the frequencies. So middle-frequency components still contain less domain-invariant information than high-frequency components while having more domain-specific information than low-frequency components. Why this can be trivial to the problem?
-
- 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
Weak Accept — could be accepted, dependent on rebuttal (4)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
In summary, the paper is well organized and explained in detail. The proposed method is innovative and effective. But still, some subtle problems can be improved such as more experiments and proper analysis.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #3
- Please describe the contribution of the paper
The authors address a significantly important problem of domain generalization in machine learning (specifically, deep learning) where models trained on one dataset may not translate to data from other domains. They use a regularization technique modeled on the Lipschitz function to remove high frequency components, which they identify as containing device/domain specific information, and demonstrate using 6 datasets that the model performance improves (generalizes) after their regularization process.
- Please list the main strengths of the paper; you should write about 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 provide a clear description of the problem of generalization and their observation that the domain specific/device specific properties on medical images are in the high frequency components, unfortunately so are the edges which are necessary for segmentation. So a balance needs to be maintained between medium frequency and high frequency subtraction. They demonstrate through experiments how their regularization method supports generalization. The math behind their idea is clearly described and should very accessible to most DL scientists/researchers.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
There are few weaknesses, if any. I would have liked to see other modalities, but then it might exceed the scope of a conference paper.
- Please rate the clarity and organization of this paper
Excellent
- 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.
- Do you have any additional comments regarding the paper’s reproducibility?
It is good to see that the authors will make code available.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
The authors address a very important problem of domain generalization challenges since DL methods “overfit” to identically distributed data and are less useful on identical but differently distributed data from other domains. They observe that domain specific properties are captured in high frequencies. Subtracting out these HF components and balancing for the mid frequency components in a Lipschiftz regularization framework (with appropriate pragmatic adjustments) results in demonstrable generalization. Excellent work!
- 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
Accept — should be accepted, independent of rebuttal (5)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Meaningful, practical work and novel contribution demonstrated through well conducted experiments.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Author Feedback
N/A
Meta-Review
Meta-review not available, early accepted paper.