List of Papers Browse by Subject Areas Author List
Abstract
Magnetic resonance spectroscopy (MRS) of brain tumors provides useful metabolic information for diagnosis, treatment response, and prognosis. Single-voxel MRS requires precise planning of the acquisition volume to produce a high-quality signal localized in the pathology of interest. Appropriate placement of the voxel in a brain tumor is determined by the size and morphology of the tumor, and is guided by MR imaging. Consistent placement of a voxel precisely within a tumor requires substantial expertise in neuroimaging interpretation and MRS methodology. The need for such expertise at the time of scan has contributed to low usage of MRS in clinical practice. In this study, we propose a deep learning method to perform voxel placements in brain tumors. The network is trained in a supervised fashion using a database of voxel placements performed by MRS experts. Our proposed method accurately replicates the voxel placements of experts in tumors with comparable tumor coverage, voxel volume, and voxel position to that of experts. This novel deep learning method can be easily applied without an extensive external validation as it only requires a segmented tumor mask as input.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1668_paper.pdf
SharedIt Link: https://rdcu.be/dV1WQ
SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72384-1_51
Supplementary Material: https://papers.miccai.org/miccai-2024/supp/1668_supp.pdf
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{Lee_ADeep_MICCAI2024,
author = { Lee, Sangyoon and Branzoli, Francesca and Nguyen, Thanh and Andronesi, Ovidiu and Lin, Alexander and Liserre, Roberto and Melkus, Gerd and Chen, Clark and Marjańska, Małgorzata and Bolan, Patrick J.},
title = { { A Deep Learning Approach for Placing Magnetic Resonance Spectroscopy Voxels in Brain Tumors } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15003},
month = {October},
page = {543 -- 552}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a deep-learning based approach for using MR imaging to guide the placement of a voxel in Single Voxel Magnetic resonance spectroscopy (MRS) applications. Here, the authors use brain tumors as a case study for the application of their method. Their network is trained in a supervised manner using an established training dataset of voxel placements from experts. There are two neural networks implemented in the paper. The first one is a nnUnet, which outputs a tumor mask (trained on the BRATS challenge dataset). The second neural network takes the tumor mask and an Euclidean distance map as an input and outputs the voxel placement. Experimental results by the authors showed a good agreement of the automated voxel placement generated by their algorithm compared to expert-based placements in images with brain tumors.
- 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 problem of voxel placement in MRS is challenging, as one prescribes a single large voxel encompasses the tissue to be imaged. The placement of the voxel can affect the calculation of tissue properties (gray/white/CSF etc.) and confounds due to partial volume effects. This problem is also challenging in case of brain tumors as they are often heterogeneous, have no specific morphology, and have a large variation in size, location, and the stage. Thus, the authors are indeed tackling an important problem in the field.
The authors created an expert-annotated curated dataset tailored for voxel-placements. Although it is not clear if this dataset will be released publicly.
Experimental results show low bias and higher agreement of the deep learning-based results with the manual expert-based results.
This method is 30X faster than another optimization-based method (which is also automated).
To my knowledge, deep learning-based approaches for MRS voxel placement have not been tried in the field before.
- 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 neural network architectures are not novel.
What the authors call the heuristic method that uses a hand-crafted objective function, is in fact a model-based principled approach that maximizes the overlap between the voxel volume and the tumor mask. This simple model-based approach does achieve a higher agreement with the expert placements without any training. That means, for this type of problem, deep learning could be thought of as an over-engineered approach.
One could also do better than the objective function that maximizes the overlap between the voxel and the mask. One could also perform a geometric alignment including rotations in addition to the voxel overlap and achieve even better results. The authors should comment on the expected results from this procedure.
This method requires a tumor mask to be generated and presented as an input. This is a limitation.
Usually, the voxel placement is also limited by scanner parameters. That is, there are only limited ways that a voxel can be prescribed in the scanner as the field of view, the coverage, and the dimensions of the voxel need to satisfy geometric and timing constraints. The authors do not discuss the implications of this. It seems the author’s solution could produce a voxel of wide variety of dimensions and sizes, whereas all possible voxel placements may not be practically feasible. Can the authors comment?
- 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 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
Please also plot V_H in Fig. 4 and Fig. 5, so that readers can evaluate how the model-based approach compares in bias and accuracy estimates.
The segmentation of the tumor is key and is a necessary step prior to the author’s method. Why not compute a bounding box for the tumor, and simply adapt it to fit the voxel parameters? This could be a simple geometric alignment that can be computed quickly. The tumor segmentation could still use the deep learning solution (trained on the BRATS dataset as the authors do).
The authors could provide some information and possibly some experimental data to verify if one can get a better performance from the model-based (“heuristic” as the authors call it) approach.
Finally, no mention was made as to how this solution will be integrated in the scanner. One will require a T1w structural scan to be acquired prior to MRS (which is usually the standard way). Then these T1w scans need to be downloaded from the scanner and pre-processed for tumor mask segmentation and then fed to the deep neural network that will generate the voxel. Finally, this voxel needs to be placed on the scanner console before the MRS sequence is executed. The authors could briefly comment (even if they don’t implement in this paper) on how they plan to develop an end-to-end solution that goes from a T1w structural image to an automated voxel placement.
- 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 solves a simple but a practical problem, and thus does make a potential contribution in terms of clinical translation of methodology at MICCAI. It could also be used in an interactive setting during scanner acquisition (although the authors have not outlined a plan for this). This is a strength. However, there were several details missing from the paper, and simple geometric-based approaches were overlooked in favor of data intensive deep learning solutions. It is possible that the inference stage of the deep learning pipeline be extremely fast (in the paper, they show a 30X improvement). However, simple geometric alignments could also be tried, that will be also extremely fast to compute.
- 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 #2
- Please describe the contribution of the paper
This paper proposes a deep learning model to automatically place the cubic voxels before the magnetic resonance spectroscopy acquisition in brain glioma. This model is demonstrated to be effective in tumors with comparable tumor coverage, voxel volumes, and voxel positions. The proposed approach could provide much more consistent MRS measurements for clinical use.
- 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 main strength of the paper is the idea of utilizing the convolutional neural network for automatic voxel placement in magnetic resonance spectroscopy (MRS) of the brain tumors. MRS is a promising technique for obtaining tissue metabolites in vivo. However, it needs accurate voxel placement prior to the MR scan. Previous work addressed the issue by using a discrete optimization approach. So far, deep learning-based algorithms have not been applied in this scenario. This paper designed a CNN based deep learning model to automatically generate voxel placements, in which the tumor masks from the nnUNet model are used as the inputs. Moreover, the methodology details in the paper are well documented and the experiments are well described.
- 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 main weakness of the paper is the demonstration of clinical feasibility, no MRS result is presented. It would be interesting to see whether the final spectra would be consistent between manual, heuristic and deep-learning based placements.
- 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?
It is suggested to release the source code for reproducibility.
- 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
- Fig.1 could be deleted as Fig.3 also shows different manual placements.
- The authors use the nnUNet model trained from BRATS datasets to generate the tumor masks, what are the network parameters and how are the performance? Whether the tumor masks influence the final placement results from the network?
- The DSC values for all placements are quite low, please check the results and discuss what could be the possible reason?
- In Table.1, only the f_tumor from single expert is calculated, how about another one?
- Font size in Fig.4 are too small.
- It would be interesting to see the correlation plots between the heuristic placements and the DSC.
- 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 proposes a deep learning model to automatically placing voxels in MRS acquisition, which is of great significance in clinical practise. The idea and methodology are new, and this paper is well documented. Therefore, it could be accepted if the authors address the above concerns.
- 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 used a deep learning model to define a bounding box of an area within glioma brain tumors as defined by MRI. The model was trained against expert placed bounding boxes. Cross-validation was used to validated the model predictions and compare them to those of a previously proposed heuristic approach. The trained model appears to perform on par with the expert annotations and performs better than the heuristic 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 paper used expert annotations from 5 different experts, where each expert annotated 50 of the 125 cases. This provided an analysis of the variation within experts, which was high, as well as a training and validation for the model. The paper and methodology appear to meet a real clinical need, and the authors were systematic and thorough in evaluating their model.
- 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 main limitation of the paper is the limited dataset and the lack of a separate test set. However, the authors discuss this limitation and other limitations adequately. The model architecture and training are simple and not novel, but they appear sufficient for the task.
- 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
Page 3: “int the” should be “in the”
- 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?
The methodology appears to address a relevant clinical need and that validation of the methodology is convincing. I don’t think a rebuttal would change my evaluation of the paper.
- 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
Author Feedback
[Response to Reviewer #1]
- The neural network architectures are not novel. We agree – we chose to use established architectures to focus on the application, and will add a comment on this in the methods section.
- Concern about describing the prior method as “heuristic” rather than model-based. We will rename this as the “objective function” approach in the revision, as this description would be more specific but still distinguish it from the DL approach. It is a viable competitor to the DL approach, but the DL method can more implicitly encode the behavior of the experts and incorporate other latent factors.
- Potential for improving the objective function approach. We agree that expanding the objective functions to include additional factors may enhance its performance for the voxel placement; we will do so in future work but consider that a separate project.
- Limitation of requiring a tumor mask for deep learning method. Both deep learning and heuristic methods required tumor masks for their inputs, but fortunately this is a well-studied problem with fast and high-performing solutions.
- Constraining voxel placements by scanner performance requirements. While there are some system limits on voxel center and dimensions, the experts’ placements were all within realizable parameters, so explicit limitations were not needed. This could be useful for a more general implementation.
- Add plots of V_H in figs 4 and 5. We will add these to the revision.
- Simple geometric alignment of a bounding box for the tumor. Thank you for this suggestion; we initially had started our project with this bounding box method but found that the regression method performed better.
- Further refinement of the “heuristic” approach. We agree that improvements with this approach are possible and we will explore this further in future work.
- Implementation on the scanner. We are in the process of implementing this and will add a comment on this topic to the discussion. Briefly, T2w-FLAIR and T1w scans will be acquired and automatically sent to a dedicated DICOM receiver. Image segmentation and voxel placement will be performed automatically, and coordinates will be exported to a text file on a shared drive, where it will be imported by a custom MRS pulse sequence.
[Response to Reviewer #3]
- Limited dataset and the lack of external test set. We agree that our dataset is small, limited by the need for expert labelling. We are now starting prospective acquisitions, so we can address this in future work.
- Architecture and training are not novel. We intentionally chose this approach due to the novel application.
- Misspelling. Will be corrected in revision.
[Response to Reviewer #4]
- No demonstration of clinical feasibility with spectra data. We have not done this yet, but are beginning prospective studies in patients in Summer 2024.
- Delete figure 1. We would prefer to keep it to provide more examples, as readers may not be familiar with the voxel placement problem.
- nnUNet details and impact of segmentation on performance. We will add details of the nnUNet training to the supplemental data. Our model gave DSC=0.90 on the test set of the BRATS dataset. We observed that variations in the tumor mask did not significantly affect voxel placements.
- DSC values are low. The low DSC scores are typical for this application since there is no gold-standard for correct voxel placement. We will add a citation to a recent abstract reporting expert inter-reader DSC performance, and add a supporting comment to the introduction.
- Table 1 only shows a single expert. The ftumor value is actually the mean of all expert placements; we will clarify this in the revision.
- Font size in fig 4 is too small. We will correct this in the revision.
- Show correlation plots between heuristic placements and DSC. We will add these plots to the supplemental data in the revision.
Meta-Review
Meta-review not available, early accepted paper.