AMSR / conferences_raw /midl19 /MIDL.io_2019_Conference_HyeuSq9ke4.json
mfromm's picture
Upload 3539 files
fad35ef
raw
history blame
27 kB
{"forum": "HyeuSq9ke4", "submission_url": "https://openreview.net/forum?id=HyeuSq9ke4", "submission_content": {"title": "On the Spatial and Temporal Influence for the Reconstruction of Magnetic Resonance Fingerprinting", "authors": ["Fabian Balsiger", "Olivier Scheidegger", "Pierre G. Carlier", "Benjamin Marty", "Mauricio Reyes"], "authorids": ["fabian.balsiger@artorg.unibe.ch", "olivier.scheidegger@insel.ch", "p.carlier@institut-myologie.org", "b.marty@institut-myologie.org", "mauricio.reyes@med.unibe.ch"], "keywords": ["Magnetic Resonance Fingerprinting", "Image Reconstruction", "Convolutional Neural Network", "Quantitative Magnetic Resonance Imaging", "Neuromuscular Diseases"], "abstract": "Magnetic resonance fingerprinting (MRF) is a promising tool for fast and multiparametric quantitative MR imaging. A drawback of MRF, however, is that the reconstruction of the MR maps is computationally demanding and lacks scalability. Several works have been proposed to improve the reconstruction of MRF by deep learning methods. Unfortunately, such methods have never been evaluated on an extensive clinical data set, and there exists no consensus on whether a fingerprint-wise or spatiotemporal reconstruction is favorable. Therefore, we propose a convolutional neural network (CNN) that reconstructs MR maps from MRF-WF, a MRF sequence for neuromuscular diseases. We evaluated the CNN's performance on a large and highly heterogeneous data set consisting of 95 patients with various neuromuscular diseases. We empirically show the benefit of using the information of neighboring fingerprints and visualize, via occlusion experiments, the importance of temporal frames for the reconstruction.", "pdf": "/pdf/fde13eed1b2b9fdef502413ab3e0757eac93e8f4.pdf", "code of conduct": "I have read and accept the code of conduct.", "remove if rejected": "(optional) Remove submission if paper is rejected.", "paperhash": "balsiger|on_the_spatial_and_temporal_influence_for_the_reconstruction_of_magnetic_resonance_fingerprinting", "_bibtex": "@inproceedings{balsiger:MIDLFull2019a,\ntitle={On the Spatial and Temporal Influence for the Reconstruction of Magnetic Resonance Fingerprinting},\nauthor={Balsiger, Fabian and Scheidegger, Olivier and Carlier, Pierre G. and Marty, Benjamin and Reyes, Mauricio},\nbooktitle={International Conference on Medical Imaging with Deep Learning -- Full Paper Track},\naddress={London, United Kingdom},\nyear={2019},\nmonth={08--10 Jul},\nurl={https://openreview.net/forum?id=HyeuSq9ke4},\nabstract={Magnetic resonance fingerprinting (MRF) is a promising tool for fast and multiparametric quantitative MR imaging. A drawback of MRF, however, is that the reconstruction of the MR maps is computationally demanding and lacks scalability. Several works have been proposed to improve the reconstruction of MRF by deep learning methods. Unfortunately, such methods have never been evaluated on an extensive clinical data set, and there exists no consensus on whether a fingerprint-wise or spatiotemporal reconstruction is favorable. Therefore, we propose a convolutional neural network (CNN) that reconstructs MR maps from MRF-WF, a MRF sequence for neuromuscular diseases. We evaluated the CNN's performance on a large and highly heterogeneous data set consisting of 95 patients with various neuromuscular diseases. We empirically show the benefit of using the information of neighboring fingerprints and visualize, via occlusion experiments, the importance of temporal frames for the reconstruction.},\n}"}, "submission_cdate": 1544690240251, "submission_tcdate": 1544690240251, "submission_tmdate": 1561399393214, "submission_ddate": null, "review_id": ["S1g6kGOdQ4", "BJxnzFaBmN", "Bke353snm4"], "review_url": ["https://openreview.net/forum?id=HyeuSq9ke4&noteId=S1g6kGOdQ4", "https://openreview.net/forum?id=HyeuSq9ke4&noteId=BJxnzFaBmN", "https://openreview.net/forum?id=HyeuSq9ke4&noteId=Bke353snm4"], "review_cdate": [1548415461459, 1548241172325, 1548692628329], "review_tcdate": [1548415461459, 1548241172325, 1548692628329], "review_tmdate": [1550053802064, 1548856720001, 1548856700759], "review_readers": [["everyone"], ["everyone"], ["everyone"]], "review_writers": [["MIDL.io/2019/Conference/Paper37/AnonReviewer1"], ["MIDL.io/2019/Conference/Paper37/AnonReviewer3"], ["MIDL.io/2019/Conference/Paper37/AnonReviewer2"]], "review_reply_count": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "review_replyto": ["HyeuSq9ke4", "HyeuSq9ke4", "HyeuSq9ke4"], "review_content": [{"pros": "- This study is well-motivated: reconstructing MR maps from MRF fast scans provide a useful tool for wide-range of medical applications.\n- Clarity: the text is very clear and understandable with enough necessary technical details.\n ", "cons": "- Limited contribution with respect to the previous work: this manuscript offers limited methodological contribution with respect to an already published paper from the same authors (Balsinger et al., 2018). The extra contributions of this manuscript are 1) application on a larger dataset (from 6 to 95); 2) a 'yet' new architecture; and 3) extra analysis on the size of receptive fields and temporal frame importance. While the first extra contribution is very valuable, the second and the third does not add much due to technical soundness, sub-optimal experimental design, and partial reporting of results (see the followings).\n- This study suggests a new CNN architecture for patch-wise MR reconstruction from spatio-temporal MRFs using the under-experiment data with limited size. This has recently become a common practice, and an unfortunate pitfall, especially in the medical imaging context due to \"Architecture-Data Bias\" problem in comparison with other existing methods. In other words, the authors tend to compare their already very-fine tuned architecture (with a massive number of parameters and hyperparameters) -on a rather small benchmarked data- with existing architectures that are again fine-tuned with \"other\" datasets. Without any surprise every time we try this we will come up with better results from every respect compared to competing approaches (same as table 2). There is a pressing need in the community to raise awareness about this problem. The author might convince me to some degree to change my mind about the presented results if they also present the results for the proposed architecture in Balsinger et al., 2018 on this dataset.\n- The authors proposed to concatenate the real and imaginary parts of input in the time dimension. This choice seems to oppose the natural spatio-temporal structure of data and blurs the temporal frame importance analysis. I would suggest to consider the real/imaginary features as a new dimension in data and use 3D-CNN. Analyzing the importance of real and imaginary parts in the reconstruction process would also be a nice research question.\n- Analyzing the influence of receptive field is very interesting but some arbitrary choice made by the authors in the analysis and reporting results put the results under a question. For example, the authors opted to only report the results for T1_H2O and not for FF and B. Another example is to use 3 \\times 3 convolution for 13 \\times 13 receptive field! Why? I would also like to see results for bigger receptive fields e.g., 20 \\times 20.\n- A minor comment is the quality of Figure 2. The left-panel is not informative as it shows similar reconstruction across different methods. The numbers in the colorbars are so tiny and not readable.", "rating": "3: accept", "confidence": "3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, {"pros": "\nThis paper proposes to use a CNN architecture to reconstruct MR Fingerprinting parametric maps. The authors test their algorithm on a dataset of 95 subjects for neuromuscular disease. They compare their method with two state of the art deep learning methods and illustrate superior performance on NRMSE, PSNR, SSIM and R2 metrics. Moreover, they have done some ablation studies to show the importance of the receptive field and temporal frames for MRF reconstruction.\n\nI believe the experiments are thorough and well designed to back the claims of the paper. The utilized network architecture can be better explained with an emphasis on specific design choices.\n\n1- This paper is well written and the message is clear to the reader. \n2- The extensive tests on a real dataset instead of phantom cases is definitely a strength of the paper. \n", "cons": "\n3- The description of the network architecture is not clear for the reader. How does the temporal and spatial blocks work? They seem to work in different dimensions of the signals. Even though the authors explain the details in the text I believe an additional illustration in each block (maybe in Appendix) might be helpful to reproduce the method in the paper for further research.\n4- How does the specifics of the network architecture influence the performance? Why do the authors reuse the input of a temporal block to its output and how does this influence the performance?\n5- How is the complex component of the signal concatenated into a channel ? Does the order of concatenation influence the results? Did the authors considered to utilize complex valued networks for this task? \n6- The quantitative results are yielded using multiple segmentation masks due to MR physics related concerns. Are the results on Table 1 heavily dependent on use of these masks? Are the results on the entire parametric maps in line with the current results? \n7- What is the number of parameters required for each method in Table 1? The reason for high performance of the proposed method can be explained with the required number of parameters to train the method. Please elaborate on this.\n8-The lack of scalability and the requirement of computational time is highlighted in the introduction and abstract. However, no quantitative comparisons are provided. I believe the computational time can be added for each method in Table 1. \n\n\nMinor suggestions\n\na- Some recent work on using the complex-valued neural networks (Virtue Patrick et al., arxiv), geometry of deep learning (Golbabaee et al., arxiv)and recurrent neural networks (Oksuz et al.,arxiv) for MRF dictionary matching can be mentioned in the literature review with their strengths and weakneses.\nb- Please explain (a.u.) term in Fig.2.\nc- Quantitative results can be mentioned in the abstract.\n", "rating": "3: accept", "confidence": "3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature"}, {"pros": "his paper addresses the issue of MR map reconstruction for magnetic resonance fingerprinting. The authors proposed a CNN based strategy to reduce the time required for reconstruction of such images. \n\nThe paper includes an extended state of the art review. One of the open issues highlighted is the limited amount of data available for this type of studies. In this study a dataset of 95 scans is included. \n \nThe proposed architecture is compared with two other deep learning architectures.", "cons": "Was experimentation performed before the proposed CNN architecture was defined? Were other optimizers and optimizer hyperparameters evaluated?\n\nMeasures utilized for evaluation. It would be beneficial to see what are the effects of such methods in the texture of the reconstructed maps. For instance homogeneity.", "rating": "3: accept", "confidence": "2: The reviewer is fairly confident that the evaluation is correct"}], "comment_id": ["HyeEc8TtVE", "r1xWbw6FNV", "rJeBL_at4E", "ryxQOOaYN4", "H1eZtMOWr4"], "comment_cdate": [1549551244436, 1549551353059, 1549551692599, 1549551723025, 1550054008759], "comment_tcdate": [1549551244436, 1549551353059, 1549551692599, 1549551723025, 1550054008759], "comment_tmdate": [1555946020488, 1555946020268, 1555946020004, 1555946019744, 1555945958930], "comment_readers": [["everyone"], ["everyone"], ["everyone"], ["everyone"], ["everyone"]], "comment_writers": [["MIDL.io/2019/Conference/Paper37/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper37/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper37/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper37/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper37/AnonReviewer1", "MIDL.io/2019/Conference"]], "comment_reply_content": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "comment_content": [{"title": "Response to AnonReviewer2", "comment": "We thank the reviewer for his comments and address his questions and concerns hereby.\n\n1) Was experimentation performed before the proposed CNN architecture was defined?\n\tWe first experimented with the architecture proposed in Balsiger et al. 2018. However, this approach did not perform as intended, which we attributed to the more difficult dataset (highly heterogeneous patients). These findings lead then to the design of the presented architecture with temporal and spatial blocks inspired by recent literature (DenseNet). Based on the concerns raised by reviewer AnonReviewer1 and for a better comparison, we will add the results of Balsiger et al. 2018 in the camera-ready version of the paper.\n\n2) Were other optimizers and optimizer hyperparameters evaluated?\n\tWe optimized the learning rate of the current setting but did not investigate other optimizer hyperparameters or optimizer types, which is mainly limited due to our available computational resources. However, we did, for instance, try MAE as the loss function, which performed worse.\n\n3) Effects of such methods in the texture of the reconstructed maps. For instance homogeneity.\n\tWe remark that the calculated evaluation metrics correspond to the commonly used metrics for medical image reconstruction (see e.g. Zbontar et al., 2018 [https://arxiv.org/abs/1811.08839]). Among these metrics is the structural similarity index measure (SSIM) which aims to assess the visual quality of the reconstruction. However, since we work with parametric maps, the visual quality or texture might be of secondary importance compared to accurately reconstructed values. For this reason, we further calculated the coefficient of determination.\n\nBased on all reviews we will update:\n1) add results for Balsiger et al., 2018 (Table 1 and Figure 2)\n2) Figure 3 with receptive field up to 21 x 21 and with results for FF and B1 in the appendix\n3) Revise paragraph about the architecture and open source our code on GitHub for reproducibility.\n4) Discuss the number of parameters between different architectures.\n5) Other suggested minor changes (revise Figure 2 for better readability, references, ...)"}, {"title": "Response to AnonReviewer1", "comment": "We thank the reviewer for his comments and address his questions and concerns hereby.\n\n1) This study suggests a new CNN architecture for patch-wise MR reconstruction from spatio-temporal MRFs using the under-experiment data with limited size. (...)\n\tWe would like to emphasize that our study is already a major improvement to what has been published in the deep learning-based MRF reconstruction literature: 1) the \u201crather small\u201d dataset is roughly 4 times larger than what has been reported previously, 2) consists only of acquired and not phantom or simulated data, 3) contains only patients with additionally a high variability in tissue appearance, and 4) contains two anatomical regions. We will add the results for Balsiger et al, 2018 to facilitate future comparisons. Further, we will open source our code on GitHub. The results for Balsiger et al. 2018 are:\n\nMR map | NRMSE | PSNR | SSIM | R2\nT1H2O | 0.032 \u00b1 0.017 | 33.27 \u00b1 3.485 | 0.972 \u00b1 0.013 | 0.949\nFF | 0.021 \u00b1 0.007 | 33.90 \u00b1 2.564 | 0.985 \u00b1 0.012 | 0.999\nB1 | 0.035 \u00b1 0.006 | 31.08 \u00b1 1.135 | 0.974 \u00b1 0.008 | 0.993\n\n2) The authors proposed to concatenate the real and imaginary parts of input in the time dimension. (...)\n\tWe initially conducted experiments with 3-D CNNs, e.g. input would be X x Y x 175 x 2 and output would be X x Y x 1 x 3. The main problem was the need to reduce the temporal dimension (T=175 down to 1), which we tried with pooling operations as well as strides > 1. However, we failed to achieve an on par performance with the proposed architecture. Therefore, we abandoned the idea and stayed the commonly used principle of concatenating real and imaginary parts. In summary, we agree with the reviewer that the optimal MRF data handling is still an open research question, which was, in fact, the main motivation of this paper. Please also refer to response 5) for AnonReviewer3, where we also elaborate on the handling of the complex-valued data.\n\n3) The authors opted to only report the results for T1_H2O and not for FF and B.\n\tOnly reporting results for the T1_H2O was motivated by the difficulty to reconstruct this map. In case of acceptance, we will include similar figures as Figure 3 for FF and B1 into the appendix (Figure 3). For FF, the best receptive field is also clearly 11 x 11. For B1, a receptive field of 1 x 1 has a slightly lower median NRMSE than 11 x 11, but the standard deviation is much larger for 1 x 1 than 11 x 11. In summary, the observations hold for FF and B1 too.\n\n4) Another example is to use 3 x 3 convolution for 13 x 13 receptive field! Why? I would also like to see results for bigger receptive fields e.g., 20 x 20.\n\tRegarding the use of 3 x 3 convolution filters, it has been shown that this is beneficial to use two 3 x 3 kernels compared to only one 5 x 5 kernel (fewer parameters, additional non-linearity; see e.g. Kamnistas et al., 2017 Med Image Anal). To further show the limitations of a larger receptive field, we will include the sizes 15 x 15, 17 x 17, 19 x 19, and 21 x 21 into Figure 3 (and the figures in the appendix for FF and B1). For all maps (T1H2O, FF, and B1), receptive fields larger than 11 x 11 perform worse. Therefore, the observation of 11 x 11 as the optimal receptive field still holds.\n\n5) A minor comment is the quality of Figure 2. The left-panel is not informative as it shows similar reconstruction across different methods. The numbers in the colorbars are so tiny and not readable.\n\tWe will improve Figure 2 by adding the results of Balsiger et al., 2018 and increasing the figure size, which will make visual comparison easier and the color bars readable.\n\nBased on all reviews we will update:\n1) add results for Balsiger et al., 2018 (Table 1 and Figure 2)\n2) Figure 3 with receptive field up to 21 x 21 and with results for FF and B1 in the appendix\n3) Revise paragraph about the architecture and open source our code on GitHub for reproducibility.\n4) Discuss the number of parameters between different architectures.\n5) Other suggested minor changes (revise Figure 2 for better readability, references, ...)"}, {"title": "Response to AnonReviewer3 - Part 1", "comment": "We thank the reviewer for his comments and address his questions and concerns hereby.\n\n3) The description of the network architecture is not clear for the reader.\n\tWe will revise the text and open source our code on GitHub for reproducibility.\n\n4) How does the specifics of the network architecture influence the performance?\n\tAn ablation of the skip-connection around the temporal block showed a slightly decreased performance. Therefore, we kept the skip-connection as proposed in J\u00e9gou et al., 2016.\n\n5) How is the complex component of the signal concatenated into a channel ? Does the order of concatenation influence the results? Did the authors consider to utilize complex valued networks for this task?\n\tWe acknowledge this question highly, which was also raised by AnonReviewer1, and think it is of utmost importance. The optimal MRF data handling is an open research question (3-D CNNs, RNNs, complex-valued networks; please see also the response 2) to AnonReviewer1).\nThe input signal is complex-valued of length 175, i.e. a tensor 175 x 2, where the second dimension contains the real and imaginary parts. We concatenate the input by [t1_real, t2_real, ..., t175_real, t1_imag, t2_imag, ..., t175_imag], arriving at 350 x 1. We also experimented with an interleaved concatenation [t1_real, t1_imag, t2_real, t2_imag, \u2026 , t175_real, t175_imag], which showed almost no difference:\n\nNRMSE T1H2O: 0.024 \u00b1 0.014\nNRMSE FF: 0.019 \u00b1 0.007\nNRMSE B1: 0.025 \u00b1 0.007\n\nRegarding complex-valued networks, we run our architecture using the implementation by Trabelsi et al., 2017 (https://arxiv.org/abs/1705.09792). We were not able to achieve a better performance:\n\nNRMSE T1H2O: 0.028 \u00b1 0.015\nNRMSE FF: 0.022 \u00b1 0.006\nNRMSE B1: 0.025 \u00b1 0.006\n\n6) The quantitative results are yielded using multiple segmentation masks due to MR physics related concerns. Are the results on Table 1 heavily dependent on use of these masks? Are the results on the entire parametric maps in line with the current results?\n\tThe use of multiple segmentation masks is motivated by: i) we threshold the foreground mask at FF < 0.65 to limit effect of highly fatty infiltrated tissue on the T1H2O, therefore NRMSE, PSNR, and SSIM of FF and B1 maps are not affected by this procedure. ii) we segment ROIs to calculate an R2, therefore it affects T1H2O, FF, and B1.\nRegarding the thresholding to limit effects of highly fatty infiltrated tissue: Based on literature (see e.g. Carlier et al., 2014 NMD, which discusses the influence on T2H2O) it does not make sense from a clinical as well as MR perspective to assess T1H2O maps on highly fatty infiltrated tissues. Furthermore, water-related parameters are extracted with a poor confidence interval in these regions. The quantitative performance is indeed worse when the same mask is used as for FF and B1. But the decrease in performance affects all methods similarly. If requested, we can add an additional table with an appropriate description of the problem into the appendix. But we think that this might add confusion since the target audience is almost certainly not familiar with the relevant literature.\nRegarding the ROI segmentation to calculate the R2. To calculate an R2, we need to have independent samples. Therefore, calculating e.g. the R2 voxel-wise is not feasible since adjacent voxels are certainly not independent of each other. To get independent measures, we calculated the mean value within the major muscle ROIs in each subject. We argue that the major muscles might be independent of each other, which is true for most neuromuscular pathologies that independently affect the different muscle groups. By pooling these values we can calculate an R2 per MR map. So the procedure to segment ROIs is statistically motivated.\n\n7) What is the number of parameters required for each method in Table 1? The reason for high performance of the proposed method can be explained with the required number of parameters to train the method. Please elaborate on this.\n\tThe number of parameters is 5659404 (proposed), 1309892 (Fang et al.), 196804 (Cohen et al.). We agree with the reviewer that the performance compared to the other methods might be explainable with the number of parameters. However, we argue that the number of parameters play a secondary role in our investigation of the spatial and temporal dimensions, which is a major contribution of our paper. Further, it would make little sense to increase the number of parameters to be equal to our method e.g. for the method of Cohen et al., which processes only a fingerprint rather than a patch of fingerprints. In that case, the network might be easily prone to overfitting. Also, the experiments with our dataset might show that the architectures in the literature were under-designed simply due to the use of phantom, simulated, or homogenous data."}, {"title": "Response to AnonReviewer3 - Part 2", "comment": "8) The lack of scalability and the requirement of computational time is highlighted in the introduction and abstract. However, no quantitative comparisons are provided. I believe the computational time can be added for each method in Table 1.\n\tWe decided to omit the computational time since we think the comparison would be unfair mainly due to the data loading. E.g., the implementation of the data loading for Cohen et al. is suboptimal, which makes a comparison difficult and unfair (the focus of the implementation was flexibility rather than speed). We think that with a proper implementation, the ranking regarding the time to reconstruct the maps would be: Fang et al. < Proposed < Cohen et al. with all being around or below one minute.\n\n9) Minor suggestions a, b, and c\n\tWe will incorporate the suggestions accordingly.\n\nBased on all reviews we will update:\n1) add results for Balsiger et al., 2018 (Table 1 and Figure 2)\n2) Figure 3 with receptive field up to 21 x 21 and with results for FF and B1 in the appendix\n3) Revise paragraph about the architecture and open source our code on GitHub for reproducibility.\n4) Discuss the number of parameters between different architectures.\n5) Other suggested minor changes (revise Figure 2 for better readability, references, ...)"}, {"title": "ACCEPT", "comment": "The modifications proposed by the authors in the rebuttal phase address my concerns to some degree."}], "comment_replyto": ["Bke353snm4", "S1g6kGOdQ4", "BJxnzFaBmN", "BJxnzFaBmN", "r1xWbw6FNV"], "comment_url": ["https://openreview.net/forum?id=HyeuSq9ke4&noteId=HyeEc8TtVE", "https://openreview.net/forum?id=HyeuSq9ke4&noteId=r1xWbw6FNV", "https://openreview.net/forum?id=HyeuSq9ke4&noteId=rJeBL_at4E", "https://openreview.net/forum?id=HyeuSq9ke4&noteId=ryxQOOaYN4", "https://openreview.net/forum?id=HyeuSq9ke4&noteId=H1eZtMOWr4"], "meta_review_cdate": 1551356587809, "meta_review_tcdate": 1551356587809, "meta_review_tmdate": 1551881980580, "meta_review_ddate ": null, "meta_review_title": "Acceptance Decision", "meta_review_metareview": "The authors address the task of image reconstruction for magnetic resonance fingerprinting (MRF), presenting a study with almost 100 cases. The reviewers comment positively on the novelty of MRF (and, hence, the related spatio-temporal reconstruction task), the size of the data set, and the thoroughness of the evaluation, while criticizing technical specificities of the proposed reconstruction approach, such as the way how the complex signal is dealt with. \n\nThey all recommend 'accept' and I would agree with they recommendation. ", "meta_review_readers": ["everyone"], "meta_review_writers": ["MIDL.io/2019/Conference"], "meta_review_reply_count": {"replyCount": 0}, "meta_review_url": ["https://openreview.net/forum?id=HyeuSq9ke4&noteId=SklN3zIBU4"], "decision": "Accept"}