List of Papers Browse by Subject Areas Author List
Abstract
One of the primary challenges in brain tumor segmentation arises from the uncertainty of voxels close to tumor boundaries. However, the conventional process of generating ground truth segmentation masks fails to treat such uncertainties properly. Those ``hard labels’’ with 0s and 1s conceptually influenced the majority of prior studies on brain image segmentation. As a result, tumor segmentation is often solved through voxel classification. In this work, we instead view this problem as a voxel-level regression, where the ground truth represents a certainty mapping from any pixel to the border of the tumor. We propose a novel ground truth label transformation, which is based on a signed geodesic transform, to capture the uncertainty in brain tumors’ vicinity. We combine this idea with a Focal-like regression L1-loss that enables effective regression learning in high-dimensional output space by appropriately weighting voxels according to their difficulty. We thoroughly conduct an experimental evaluation to validate the components of our proposed method, compare it to a diverse array of state-of-the-art segmentation models, and show that it is architecture-agnostic. The code of our method is made publicly available (\url{https://github.com/Oulu-IMEDS/SiNGR/}).
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/2261_paper.pdf
SharedIt Link: pending
SpringerLink (DOI): pending
Supplementary Material: https://papers.miccai.org/miccai-2024/supp/2261_supp.pdf
Link to the Code Repository
https://github.com/Oulu-IMEDS/SiNGR
Link to the Dataset(s)
https://www.med.upenn.edu/cbica/brats2020/data.html
https://ieee-dataport.org/documents/brain-mri-segmentation
BibTex
@InProceedings{Dan_SiNGR_MICCAI2024,
author = { Dang, Trung and Nguyen, Huy Hoang and Tiulpin, Aleksei},
title = { { SiNGR: Brain Tumor Segmentation via Signed Normalized Geodesic Transform Regression } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15009},
month = {October},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces a novel approach to brain tumour segmentation by viewing the task as voxel-level regression instead of the conventional classification task. Authors propose a novel ground truth transformation, signed geodesic transformation, in order to better capture the uncertainty of labels around tumour boundaries. The proposed approach incorporates a Focal-like regression L1-loss, which weights voxels based on their difficulty, allowing for more effective learning from hard-to-learn regions. A thorough experimental evaluation is conducted to show the effectiveness of the proposed approach.
- 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 method offers a simple yet effective solution to an important problem in brain tumour segmentation.
-
Experimental results demonstrate a significant improvement in segmentation accuracy with the proposed approach.
-
The method appears to be architecture-agnostic, as it delivers consistent results across a range of different network architectures.
-
The approach allows a simple architecture like UNet3D to outperform more complex models, thereby reducing both the complexity of training and the time required for successful segmentation results.
-
The proposed approach can be easily adapted to different types of data, making it flexible and broadly applicable.
-
- 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 behavior of the Focal-L1 loss depicted in Fig. 2b might be confusing. For voxels with ground truth values other than -1 or 1, Fig 2b. indicates proposed loss function has two local minima. For example, if S_i is equal to 0.1 (a foreground voxel), Focal-L1 has a local minima at 0.1 and another local minima at -1. This suggests that as S_i - Z_i increases, the loss function does not exhibit a non-decreasing behaviour, potentially leading to misinterpretation of the loss design. The authors should address this ambiguity and clarify the expected behavior of the Focal-L1 loss. - The literature review lacks depth, overlooking key studies that treat segmentation as regression of Euclidean distance. Some notable works are:
1) A Pixel-Wise Distance Regression Approach for Joint Retinal Optical Disc and Fovea Detection, Meyer et al., MICCAI 2018 2) Enforcing connectivity of 3D linear structures using their 2D projections, Oner et al., MICCAI 2022 3) Adjusting the ground truth annotations for connectivity-based learning to delineate, Oner et al., TMI 2022
To avoid confusion, the authors should emphasise that the main novelty of their approach is the use of signed geodesic distance transform, not the regression concept itself. The differences between the proposed method and these earlier studies should be discussed. It would be beneficial to see a justification for using geodesic distance over Euclidean distance, with both qualitative and quantitative results. Additionally, a comparison between SiNG, GeoLS, and Euclidean distance regression in Table 3 would offer valuable insights.
-
Fig. 3 shows that UNet3D achieves a significantly greater performance boost compared to more complex architectures. The authors should explain why this simpler architecture outperforms others in this context, providing more detailed insights.
-
The reported scores for GeoLS in Table 3 are substantially different from those in reference [20]. The authors should clarify whether these discrepancies are due to variations in the experimental setup or other factors.
-
- 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
While the paper proposes an effective solution to an important problem, two critical points need to be clarified and improved.
-
The proposed loss function seems to have two local minima for certain voxel values, which could lead to unpredictable behaviour during training. It should be made clear that the confusing behaviour seen in Fig 2b. is caused by the sample weighting and since the weighting does not have an effect on back-propagation, it will not lead to unstable behaviour during training.
-
The paper overlooks key studies that approach segmentation using regression of Euclidean distance. In order to strengthen the context of the paper, the authors should incorporate these references and clarify how the use of a signed geodesic distance transform offers a unique advantage with visuals and experimental results.
-
- 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?
Despite the aforementioned drawbacks, the paper’s strengths—such as improved performance, architecture-agnosticism, and flexible application—suggest that the method has potential. Given this, a weak accept is recommended, with the expectation that the authors will address the identified issues in future revisions.
- 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
Review #2
- Please describe the contribution of the paper
This paper proposed unsigned geodesic distance transform to approximate modeling the segmentation and introduce a Focal-L1 loss to handle the imbalance of foreground and background.
- 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-organized and proposed method is easy to reproduce.
- 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.
As distance map has been a hot topic in the past few years, a lot of papers are proposed to train segmentation models with the help of distance map. Especially, the paper named “Learning geodesic active contours for embedding object global information in segmentation CNNs” is much similar to this paper. The novelty of this paper is limited. In experiment part, authors need to point out the differences between the algorithm in this article and the algorithms in the following literature, and what are the advantages of the algorithm in this article compared to their methods. [1]Zhang, Minghui, Guang-Zhong Yang, and Yun Gu. “Differentiable topology-preserved distance transform for pulmonary airway segmentation.” arxiv preprint arxiv:2209.08355 1.2 (2022). [2]Ma, Jun, Jian He, and X, Yang. “Learning geodesic active contours for embedding object global information in segmentation CNNs.” IEEE Transactions on Medical Imaging 40.1 (2020): 93-104. [3]Wang, Yan, et al. “Deep distance transform for tubular structure segmentation in ct scans.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.
- 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
Authors need to point out the differences between the algorithm in this article and the algorithms in the following literature
- 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 paper is well-organized and proposed method is easy to reproduce. The novelty of this paper is limited. Author should further illustrate their novelty comparing to other distance map based methods.
- 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 paper formulated the semantic segmentation problem as voxel regression, and proposed a novel ground truth label transformation, which is based on a signed geodesic transform, capturing the uncertainty of brain tumors’ vicinity. The paper also introduced Focal-L1 loss to handle the imbalance of foreground and background voxel. The experimental evaluation shows that our method is beneficial for various DL architectures.
- 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 proposed a unique ground truth label transformation technique that relies on a signed geodesic transform. This approach effectively captures the uncertainty surrounding brain tumors, providing a more nuanced and accurate representation of the tumor’s vicinity. This is a valuable contribution to the field of medical image analysis, as it allows for a more precise understanding of tumor boundaries, which is crucial for accurate diagnosis and treatment planning. -The paper introduced the novel SiNG transform, a technique that converts traditional 0-1 annotated masks into soft labels. This transformation accounts for the inherent uncertainty in the labeling process, resulting in more robust and reliable labels. -The paper introduced the Focal-L1 loss, a novel loss function that effectively weights voxels based on their difficulty. This approach allows the model to focus more on challenging voxels during training, improving the overall performance of the deep learning models. -The paper conducted standardized and thorough experiments on two widely used datasets, BraTS and LGG FLAIR. The results demonstrate that the proposed method consistently enhances performance across different deep learning architectures. This is a strong indicator of the method’s generalizability and applicability to a wide range of medical image analysis tasks.
- 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 lacks visualization results, which are crucial for intuitively demonstrating the improvement in edge segmentation. Without these visualizations, it becomes difficult for readers to fully grasp the effectiveness of the proposed method in enhancing the accuracy of tumor boundary detection. -The paper lacks ablation experiments to demonstrate the individual contributions of the SiNG transform and Focal-L1 loss to the model’s performance. Ablation studies are essential for understanding the relative importance of different components of a complex system like the one proposed in this article.
- 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
-The paper should include visualization results to intuitively demonstrate the improvement in edge segmentation. By providing visualizations of the output from the proposed method, readers can better appreciate the enhanced accuracy in tumor boundary detection. These visualizations could include contour overlays on the original images, 3D renderings, or any other appropriate representation that clearly illustrates the improvements. -The paper should include more comprehensive ablation experiments to demonstrate the individual contributions of the SiNG transform and Focal-L1 loss to the model’s performance. The authors should conduct experiments where they remove or replace each component and compare the results to the baseline. By analyzing the performance differences, they can provide evidence of the specific impact of SiNG and Focal-L1 on the overall accuracy and reliability of the system. It is essential to present and discuss these results clearly, explaining the reasons for any observed improvements or degradations.
- 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?
This paper proposes a novel segmentation method that can capture the uncertainty of the target edge, and the experimental results also prove that the proposed method has good performance and generalizability.
- 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
We thank all the reviewers for their constructive feedback and appreciate the high-quality reviews! We would like to address the comments in a point-to-point manner below.
R1
Visualization results: We provided the qualitative results in the supplementary materials (see Fig. S2).
Ablation experiments: Ablation studies are illustrated in Fig. 3, where we demonstrated that the combination of SiNG with a margin of 0.5 and Focal-L1 yielded the best results.
Experiments: We plan to add more experiments to the journal version of the paper.
R2
Focal-L1 loss: Thank you for your very insightful observations. We acknowledge that we do not use backpropagation on the sample weighting. This, otherwise can indeed cause unexpected behavior during training. We realize that we can simplify the denominator to only “abs(gt)” to achieve our intended purpose of the loss while avoiding this issue.
Literature review: We acknowledge that our literature review can be improved. We will add these to camera-ready, and will also incorporate relevant work proposed by R3.
UNet3D: We will study this in the journal version of the paper.
GeoLS: We will provide the detailed configurations of this method in our experiments and clarify the potential differences in experimental setup in camera-ready.
R3
We would like to address the concern related to the novelty of our method here.
Firstly, the focus of our paper is on brain tumor segmentation (BTS), which is an established problem in the MICCAI community and of high clinical relevance on its own. None of the papers listed by R2 and R3 solve BTS, and in the BTS literature, we have a novel method and results.
Secondly, our goal in using the geodesic transformation was to produce uncertainty maps for the ground truth annotations. The novelty of our method is in predicting uncertainty maps directly, instead of having it as an additional learning objective, as done in other papers (i.e. those listed by the reviewer).
Thirdly, our method is purely a regression approach, which replaces the conventional pixel-wise classification learning. This is difficult because simple normalized distance values of other approaches can be too close to the class boundaries. For example, the difference between 0.01 and -0.01 is minimal, although they belong to 2 different classes. The sign-checking part in the sample weighting of our loss function, as well as the transformation in the soft label generation, is meant to tackle this issue. We demonstrate that this pixel-wise regression approach can achieve better performance than combining pixel-wise and region-aware classification losses such as BCE and Dice loss.
Meta-Review
Meta-review not available, early accepted paper.