List of Papers Browse by Subject Areas Author List
Abstract
Accurate segmentation in low-dose CT scans remains a challenge in medical imaging, primarily due to the high annotation costs. This study introduces LUCIDA, a Low-dose Universal-tissue CT Image Domain Adaptation model operating under an unsupervised protocol without requiring LDCT annotations. It uniquely incorporates the Weighted Segmentation Reconstruction (WSR) module to establish a linear relationship between prediction maps and reconstructed images. By enhancing the quality of reconstructed images, LUCIDA improves the accuracy of prediction maps, facilitating a new domain adaptation framework. Extensive evaluation experiments demonstrate LUCIDA’s effectiveness in accurately recognizing a wide range of tissues, significantly outperforming traditional methods. We also introduce the LUCIDA-Ensemble model, demonstrating comparable performance to supervised learning models in organ segmentation and recognizing 112 tissue types.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/0562_paper.pdf
SharedIt Link: https://rdcu.be/dZxdO
SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72111-3_37
Supplementary Material: N/A
Link to the Code Repository
https://github.com/YixinChen-AI/LUCIDA
Link to the Dataset(s)
N/A
BibTex
@InProceedings{Che_LUCIDA_MICCAI2024,
author = { Chen, Yixin and Meng, Xiangxi and Wang, Yan and Zeng, Shuang and Liu, Xi and Xie, Zhaoheng},
title = { { LUCIDA: Low-dose Universal-tissue CT Image Domain Adaptation For Medical Segmentation } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15008},
month = {October},
page = {393 -- 402}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper develops a low-dose universal tissue segmentation model from NDCT with domain adaptation. The approach uses weighted segmentation reconstruction to adapt NDCT segmentor for LDCT.
- 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 paper is well-written. The evaluation is comprehensive, 1000+ cases with 100+ structures.
- 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 paper’s novelty is somewhat restricted. While the utilization of weighted segmentation reconstruction as a training loss interesting, the implication of significant intensity overlaps among various structures in CT images remains unclear.
- 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 provide sufficient information for 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
The need for domain adaptation is ambiguous, given that the disparity between NDCT and LDCT images is relatively minor compared to numerous other computer vision applications. It is possible that the NDCT model could perform effectively with LDCT data. The authors might consider including an ablation study involving the direct application of the NDCT model to LDCT images.
- 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 Reject — could be rejected, dependent on rebuttal (3)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the paper novelty is limited, with certain design questions, the evaluation presented also falls short of rendering it acceptable as an application paper.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Weak Reject — could be rejected, dependent on rebuttal (3)
- [Post rebuttal] Please justify your decision
The key question of why the NDCT segmenter could help LDCT segmentation via reconstruction loss, especially on structures with overlapping intensity still remains unclear.
Review #2
- Please describe the contribution of the paper
The paper introduces the LUCIDA model, which includes the Weighted Segmentation Reconstruction module to enhance the quality of reconstructed images and improve the accuracy of prediction maps without the need for annotations. Additionally, the paper utilizes Fourier-based domain adaptation for frequency domain adaptation of LDCT images, optimizing parameters to minimize errors between reconstructed and input images. The study evaluates denoising methods to enhance downstream task performance, highlighting the novelty of the approach in improving domain adaptation in the context of LDCT denoising.
- 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 supplement the NDCT TotalSeg dataset with seven additional categories, resulting in a more comprehensive dense label dataset. This involves a significant amount of annotation work, which will be of great help to future intensive downstream tasks.
- The paper is well-motivated. The problem of previous methods is obvious, and the proposed solution is reasonable to me. The illustrations and figures are clear.
- The innovation is novel. Map the predicted label to the reconstructed image by designing a Weighted Segmentation Reconstruction. And characterize the performance of UDA by measuring the difference between the reconstructed image and the original image. It provides new ideas for UDA in medical fields.
- 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 writing and expression are relatively implicit, and some details are difficult to follow, such as why WSR can reflect the performance of UDA, why U-net indirectly matches the patterns of the source and target domains, etc.
- For experiments, there is a lack of ablation studies to quantitatively demonstrate the effectiveness of WSR and Fourier based DA strategies. In addition, the lack of comparison with some of the latest UDA methods may lead to unconvincing experimental results.
- For formulas 2 and 5, it may be “argmin” instead of “argmax”.
- For the testing phase, the article does not specify which models or modules the target domain image needs to pass through.
- 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.
- Do you have any additional comments regarding the paper’s reproducibility?
None
- 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 WSR module is a linear mapping module, like a MLP. This simple mapping may make the shape and relationship of the tissue in the reconstructed image correct, but it cannot reconstruct the texture of a specific tissue. There is a lack of theoretical support and reasonable explanation for whether constraining the distance between this “non-realistic” reconstructed image and the original image can truly improve the reconstruction performance of WSR.
- The traditional Fourier based UAD method usually exchanges the amplitude spectra of the source domain image and the target domain image. In your method, utilize Unet to convert the amplitude and phase spectra of the target domain into those of the source domain, but the specific principle, supporting technologies and formulas are not provided. Is it possible to replace it with another non-UNet structure?
- For formulas 2 and 5, it may be “argmin” instead of “argmax” because the model hopes that the WSR reconstructed image is very close to the original image, so that in the UDA stage, the quality of the reconstructed image can reflect the model’s prediction quality.
- During the testing phase, does the target domain image only need to pass through the frozen U-Net and Segmentor?
- 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?
This paper is well-motivated. The problem of previous methods is obvious, and the proposed solution is novel and effective, which examined the UDA problem from the perspective of data reconstruction. Additionally, authors added new category annotations on the basis of publicly available datasets, making it be a more comprehensive dense label dataset. However, the article is somewhat difficult to follow, and the writing and structure still need to be optimized. And some core modules did not provide theoretical explanations
- 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
Weak Accept — could be accepted, dependent on rebuttal (4)
- [Post rebuttal] Please justify your decision
I still stand by my previous decision (i.e., weak accept).
Review #3
- Please describe the contribution of the paper
The paper introduces LUCIDA, a unsupervised domain adaptation method for medical image segmentation in low-dose CT scans. It enables apply segmentation model trained on NDCT to apply on LDCT without the need for LDCT annotations.
- 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.
This paper presents a novel UDA method by using a novel module, Weighted Segmentation Reconstruction (WSR), and a Fourier-based Domain Adaptation Unet. It is a new strategy different from GAN-style methods. The WSR is a simple linear perceptron, which is in theory suitable for CT images, to reconstruct the segmentation mask to CT images and the results validate the effectiveness of it. In addition, it also provides extensive experiments on using most popular segmentation models and validate on 21 types of tissues.
- 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 is no discussion on the quality of image reconstruction. The WSR was first trained using NDCT and was directly applied on the segmentation results of the transformed target domain image (LDCT). They did not provide details about the quality of the transformed LDCT, only mentioning that the predicted segmentation map can reflect the quality of the transformed LDCT.
- 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?
no
- 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
- In the methodology section, the equation (2) and (5) should use min instead of max. The goal is to minimize the difference between the image and the reconstructed image right?
- There is no discussion on the quality of image reconstruction. The WSR was trained to reconstruct NDCT. But It was directly applied for LDCT to train the Unet, will this cause issues?
- 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?
The method is novel and they provide extensive experiments to show LUCIDA’s effectiveness.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
Dear Reviewers and Chairs,
Thank you for all your insightful comments.
WSR Texture Issue for Reviewer #1 & Reviewer #4 Due to the physical imaging characteristics of CT, the values in a CT image reflect tissue density. Within the same Region of Interest (ROI), density is typically uniform, especially considering our detailed division into 112 distinct areas. Therefore, the WSR module does not reconstruct texture, which does not significantly impact the reconstruction loss; differences in density between different ROIs are substantial, which more critically influences reconstruction loss. The primary goal of the WSR module is not to achieve perfect reconstruction, but rather to use reconstruction to facilitate higher accuracy in identifying ROIs, which represent purely structural information.
WSR applied for LDCT Issue Reviewer #1 & Reviewer #4 The parameters of the WSR module reflect the physical density properties of ROIs. Since both NDCT and LDCT operate on the same imaging principle and yield consistent density measurements for the same structures, the WSR module trained on NDCT can be directly applied to LDCT without adaptation issues.
U-Net Alternative for Reviewer #1 In domain adaptation tasks, loss functions are crucial for effective adaptation and a good model architecture fully leverages the potential of these loss functions. For instance, GANs employ adversarial loss and FVP utilizes semi-supervised cross-entropy loss. Our innovation lies in proposing a new reconstruction loss tailored for the universal model. We chose a Fourier-based U-Net due to its impressive performance in recent FVP study, ensuring a fair comparison of results. While other neural network architectures are feasible, they do not compromise the novelty of the WSR. During testing, target domain images pass through the frozen segmentor and the frozen WSR, and U-Net parameters are optimized by minimizing the reconstruction loss.
Lack of comparison with SOTA Issue for Reviewer #1 In our manuscript, we included comparisons with recent UDA methods such as AIGAN and FVP, both introduced last year. AIGAN has achieved SOTA performance among GAN-based and traditional methods, while FVP has set a new benchmark in source-free UDA tasks. Our experimental results demonstrate that our method outperforms these latest UDA methods, affirming its effectiveness and innovation in the field.
Lack of Ablation Issue for Reviewer #3 The ablation experiments you mentioned are included in Table 1. This table shows that the NDCT model performs well on LDCT images without adaptation (first row), due to our enhanced dense label annotation. However, for more challenging tissues like vessels and abdominal organs, our LUCIDA model significantly improves performance from 84.0% and 87.7% to 91.7% and 93.8%, respectively. These enhancements are critical as the higher model accuracy benefits various downstream tasks, including report generation and kinetic analysis of PET/CT, thus facilitating further research in medical imaging.
** Why do LDCT and NDCT adaptation for Reviewer #3**
- Domain adaptation between LDCT and NDCT has been well-researched and remains an active field, as seen with AIGAN (2023) and FVP (2023).
- Adapting LDCT to NDCT significantly enhances the diagnostic quality of low-dose CT scans, making them clearer and more detailed while maintaining low radiation exposure. This adaptation reduces the need for repeated scans due to poor image quality, optimizing patient safety and healthcare efficiency. It also broadens the clinical applications of LDCT, suhc as in pediatric care and early cancer screening.
Formulas 2 and 5 for Reviewer #1 & Reviewer #4 We appreciate your comments and will correct this typo in the manuscript.
Sincerely, The Authors
Meta-Review
Meta-review #1
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
The paper develops a low-dose universal tissue segmentation model from NDCT with domain adaptation. The reviewers are generally in favor of the paper. However, some minor concerns still remain after the rebuttal. The authors shall carefully polish the paper to address the remaining concerns in their final version.
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
The paper develops a low-dose universal tissue segmentation model from NDCT with domain adaptation. The reviewers are generally in favor of the paper. However, some minor concerns still remain after the rebuttal. The authors shall carefully polish the paper to address the remaining concerns in their final version.