Datasets:

mfromm
/

AMSR

{"forum": "BJesxFsA1N", "submission_url": "https://openreview.net/forum?id=BJesxFsA1N", "submission_content": {"title": "Cluster Analysis in Latent Space: Identifying Personalized Aortic Valve Prosthesis Shapes using Deep Representations", "authors": ["Jannis Hagenah", "Kenneth Kuehl", "Michael Scharfschwerdt", "Floris Ernst"], "authorids": ["hagenah@rob.uni-luebeck.de", "kennethkuehl@freenet.de", "michael.scharfschwerdt@uksh.de", "ernst@rob.uni-luebeck.de"], "keywords": ["personalized medicine", "representation learning", "aortic valve", "personalized prosthetics", "unsupervised learning"], "TL;DR": "We propose to perform a cluster analysis in the latent space of an autoencoder to identify typical anatomies for personalized prosthesis shaping.", "abstract": "Due to the high inter-patient variability of anatomies, the field of personalized prosthetics gained attention during the last years. One potential application is the aortic valve. Even though its shape is highly patient-specific, state-of-the-art aortic valve prosthesis are not capable of reproducing this individual geometry. An appraoch to reach an economically reasonable personalization would be the identification of typical valve shapes using clustering, such that each patient could be treated with the prosthesis of the type that matches his individual geometry best. However, a cluster analysis directly in image space is not sufficient due to the tough identification of reasonable metrics and the curse of dimensionality. In this work, we propose representation learning to perform the cluster analysis in the latent space, while the evaluation of the identified prosthesis shapes is performed in image space using generative modeling. To this end, we set up a data set of 58 porcine aortic valves and provide a proof-of-concept of our method using convolutional autoencoders. Furthermore, we evaluated the learned representation regarding its reconstruction accuracy, compactness and smoothness. To the best of our knowledge, this work presents the first approach to derive prosthesis shapes data-drivenly using clustering in latent space.", "pdf": "/pdf/af2fc6504c381848de4ae830118f6ea470d07b6a.pdf", "code of conduct": "I have read and accept the code of conduct.", "remove if rejected": "(optional) Remove submission if paper is rejected.", "paperhash": "hagenah|cluster_analysis_in_latent_space_identifying_personalized_aortic_valve_prosthesis_shapes_using_deep_representations", "_bibtex": "@inproceedings{hagenah:MIDLFull2019a,\ntitle={Cluster Analysis in Latent Space: Identifying Personalized Aortic Valve Prosthesis Shapes using Deep Representations},\nauthor={Hagenah, Jannis and Kuehl, Kenneth and Scharfschwerdt, Michael and Ernst, Floris},\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=BJesxFsA1N},\nabstract={Due to the high inter-patient variability of anatomies, the field of personalized prosthetics gained attention during the last years. One potential application is the aortic valve. Even though its shape is highly patient-specific, state-of-the-art aortic valve prosthesis are not capable of reproducing this individual geometry. An appraoch to reach an economically reasonable personalization would be the identification of typical valve shapes using clustering, such that each patient could be treated with the prosthesis of the type that matches his individual geometry best. However, a cluster analysis directly in image space is not sufficient due to the tough identification of reasonable metrics and the curse of dimensionality. In this work, we propose representation learning to perform the cluster analysis in the latent space, while the evaluation of the identified prosthesis shapes is performed in image space using generative modeling. To this end, we set up a data set of 58 porcine aortic valves and provide a proof-of-concept of our method using convolutional autoencoders. Furthermore, we evaluated the learned representation regarding its reconstruction accuracy, compactness and smoothness. To the best of our knowledge, this work presents the first approach to derive prosthesis shapes data-drivenly using clustering in latent space.},\n}"}, "submission_cdate": 1544628466970, "submission_tcdate": 1544628466970, "submission_tmdate": 1561396228038, "submission_ddate": null, "review_id": ["SJxrcwyUXE", "HyeRDAlRm4", "rylWybGJEV"], "review_url": ["https://openreview.net/forum?id=BJesxFsA1N&noteId=SJxrcwyUXE", "https://openreview.net/forum?id=BJesxFsA1N&noteId=HyeRDAlRm4", "https://openreview.net/forum?id=BJesxFsA1N&noteId=rylWybGJEV"], "review_cdate": [1548248973351, 1548779109559, 1548849368853], "review_tcdate": [1548248973351, 1548779109559, 1548849368853], "review_tmdate": [1549875310716, 1548856682341, 1548856679292], "review_readers": [["everyone"], ["everyone"], ["everyone"]], "review_writers": [["MIDL.io/2019/Conference/Paper18/AnonReviewer2"], ["MIDL.io/2019/Conference/Paper18/AnonReviewer1"], ["MIDL.io/2019/Conference/Paper18/AnonReviewer3"]], "review_reply_count": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "review_replyto": ["BJesxFsA1N", "BJesxFsA1N", "BJesxFsA1N"], "review_content": [{"pros": "\n- This paper presents a clustering method using deep autoencoder for aortic value shape clustering. It is the first work to identify aortic value prosthesis types using a general representation learning technique.\n\n- This work has a remarkable clinical value. Clustering of aortic value prosthesis shapes has a high contribution to personalized medicine. \n\n- The entire workflow is quite clear and complete.\n", "cons": "\n- The introduction part is a little misleading for me. The authors emphasize that the objective is to cluster the geometric shape of leaflets, and it is hard to represent the shapes in high-dimensional space (last paragraph of introduction). I'm concerned that this would make the readers misunderstand the data are shape-models (point cloud dataset) before the description of dataset in Sec. 2.\n\n- One major concern is whether the results are reliable:\n 1. The experiments shown in Table 1 compare several different network settings. This kind of vertical comparison is insufficient to support the claims made in the study. Please compare to other representation learning methods such as sparse coding (e.g. spherical K-means, dictionary learning), dimension reduction (e.g. PCA, t-sne).\n 2. This study did not give a gold-standard for shape clustering (though it could be difficult). The experiments measure the recon accuracy. However, recon accuracy highly depends on decoder network. It is not convincing to claim that the clustering is correct since even a noise can be decoded into a normal image.\n\n- In the last paragraph of the introduction, authors say 'it is hard to define a feasible metric describing the similarity of the valve shape in general.'. However, authors use Jaccard coef. and Hausdorff distance to measure the recon accuracy between original image and reconstructed image. It is a self-contradictory statement.\n\nother comments:\n- The authors use 2D images to represent leaflet shapes, I'm concerned whether 2D photograph is precise enough. 3D scanner such as CT, MRI, optical scanner could be more suitable for this work? Though this is not the issue to be considered in this work.\n- The paper is not well organized. Details of training should be more clearly written. The hyper-parameters of autoencoder and the recon decoder should be more clearly stated for reproducibility. All architectures listed in Table 1 should be stated clearly in experiments section not only in method section.\n", "rating": "3: accept", "confidence": "2: The reviewer is fairly confident that the evaluation is correct"}, {"pros": "- Authors proposed clustering analysis for approximating shapes of aortic valve prostheses.\n\n- Authors proposed a representation learning method in latent space based on autoencoder for shape clustering of aortic valves, instead of pixel-based training.\n\n- It seems that the authors proposed a somewhat novel idea for a pragmatic application.\n\n- In Results and Discussion sections, the authors provide sufficient validation and discussions via observing the performance change according to the number of clusters or the structure of learning frameworks.", "cons": "- There is an insignificant difference between the performances of comparative methods, which is probably due to the small dataset. In spite of difficulties of acquiring the additional data, it can be argued that the minimum amount of data required for training the suggested learning model should be larger than the current dataset.\n\n- In experiments, the comparison was conducted mainly on similar deep learning models. Considering the small dataset mentioned above, it could be enough to construct the conventional feature-based clustering model via extracting the classic shape features, e.g. curvature and convexity, etc. Additional comparison with this conventional feature-based model would be better.\n\n- It would be also better to provide the further visual analysis of whether the shapes of same cluster data are actually similar and the shapes of different cluster data are actually different.\n", "rating": "3: accept", "confidence": "2: The reviewer is fairly confident that the evaluation is correct"}, {"pros": "- A clustering method based on features from a convolutional autoencoder is proposed to define clusters of similar aortic valve prosthesis shapes. \n- Interesting application and well described methodology.", "cons": "- It took me a while to realise what kind of imaging was used. This could be more clear from the abstract and/or title.\n- It would be good to have a more visual representation of the quality of the results, now only figures of the Jaccard coefficient and Hausdorff distance are shown.\n- Listing the resolution in mm/pixel instead of pixel/mm would be more intuitive.", "rating": "3: accept", "confidence": "2: The reviewer is fairly confident that the evaluation is correct"}], "comment_id": ["SylgZ--oVE", "BJlsiWbiVV", "HyxJrbWjNV", "SkxN-zZjN4", "r1llvOnAVV"], "comment_cdate": [1549631735774, 1549631906924, 1549631798835, 1549631996031, 1549875288455], "comment_tcdate": [1549631735774, 1549631906924, 1549631798835, 1549631996031, 1549875288455], "comment_tmdate": [1555946012989, 1555946012772, 1555946012517, 1555946012301, 1555945969775], "comment_readers": [["everyone"], ["everyone"], ["everyone"], ["everyone"], ["everyone"]], "comment_writers": [["MIDL.io/2019/Conference/Paper18/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper18/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper18/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper18/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper18/AnonReviewer2", "MIDL.io/2019/Conference"]], "comment_reply_content": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "comment_content": [{"title": "Summarized Rebuttal Letter", "comment": "Dear Reviewers,\nfirst of all, we want to thank all of you for the valuable feedback and the helpful comments! We gave our best to address your issues during the rebuttal period to improve our manuscript. Overall, we summarized two major concerns from your reviews, with which we completely agree! \n\nThe first one is about the comparison of different representation learning models. You are right that the models we included in this study are quite similar and that a comparison to a well-established standard model would be beneficial. Hence, we used the rebuttal time to benchmark our proposed method against a PCA-decomposition approach. To this end, we replaced the auto-encoder by a conventional PCA dimensionality reduction technique and performed the clustering in this space. The results were worse than our proposed method. While the mean Jaccard coefficient is about 0.03 smaller, we observed very big errors in the mean Hausdorff distance, rising up to 12 mm with a minimal error of 6 mm. We think that these results support our idea that using convolutional auto-encoders to cluster in the latent space is beneficial and outperforms classical approaches. We will include the results of our PCA-study in the results section of the manuscript.\n\nSecond, there is a lack of visual assessment of the clustering performance. We appreciate your idea of giving qualitative results within the paper. Thus, we will add images of the resulting prosthesis types after the clustering to the appendix in order to make the difference between them easily assessible for the reader.\nAdditionally, we will do some rework on the abstract and introduction to avoid misunderstandings about the kind of data we used. Especially, we will avoid terms like \u201chigh-dimensional geometric data\u201d and replace them by \u201cpixel-space of leaflet shape images\u201d or something similar. We will also include a sketch of the proposed architecture with a compact overview of the hyperparameters in the appendix. \n\nApart from these major points that we found throughout most of the reviews, we tried our best to tackle all of your individual comments. Please have a look into the individual rebuttals for each review.\n"}, {"title": "Rebuttal Letter Reviewer 2", "comment": "Dear Reviewer,\nthank you very much for your helpful feedback!\nWe agree on your concerns that the comparison of our model only takes quite similar methods into account. To get a better evaluation of our method, we spent the rebuttal time to also evaluate a PCA-decomposition for feature extraction, followed by the same clustering method. With the same number of features (n_z=20), the resulting clustering achieves worse results than our proposed auto-encoder method. We observed smaller values in the Jaccard coefficient and in the mean Hausdorff-distance. Especially here, we observed errors of more than 12 mm. We will include this analysis in the results section!\n\nOf course, a comparison to even more representation learning models (such as bag-of-words) would increase the trust in our study, but unfortunately, we think that such a broad analysis will be out of the scope of this paper. However, we believe that the comparison to the decomposition using PCA supports the findings of our study.\n\nYou are right that the reconstruction accuracy of the auto-encoder does not tell us anything about the clustering performance. With this experiment, we validated that our auto-encoder finds a feasible latent space, where (in the ideal case) all of the variance observed in our data is induced by the inter-patient-variability of heart valve anatomies. In the reconstruction experiment, we can observe that with n_z=20, we can derive a suitable representation and the decoder is capable of synthesizing a leaflet image that looks similar to the individual ground truth image. This is the difference to synthesizing a leaflet image that looks realistic, where even noise could serve as decoder input! The clustering performance analysis is done afterwards without changing the auto-encoder itself. Here, we just use the decoder to synthesize images of our cluster centers. And this should be feasible as we already showed the reconstruction accuracy of the whole auto-encoder. We hope, that this explanation made our experiments a bit clearer.\n\nWe completely agree that our statement in the introduction about finding a feasible similarity metric is quite confusing, given the evaluation of our method. What we mean is that clustering in pixel-space is quite hard, because two images with very small visual difference (e.g. slight translations) might have large differences in a pixel-wise comparison. We will definitely change the wording of this paragraph to solve this issue!\n\nThe same holds for the description of the image types. With \u201cgeometric shape\u201d and \u201chigh-dimensional data\u201d, we basically mean the leaflet shape in 2D images, that are in pixel-space relatively high-dimensional compared to the 20-dimensional latent space. We will rework the introduction and the abstract to prevent misunderstandings.\n\nTo the best of our knowledge, we included all relevant hyperparameters of our proposed architecture in the text. However, to increase the readability, we will add a sketch of the encoder-decoder-architecture with a compact overview of the hyperparameters to the appendix section. If you still miss additional parameters in the text, we would be very thankful for a specific comment on that to fix these inconsistencies.\n"}, {"title": "Rebuttal Letter Reviewer 1", "comment": "Dear Reviewer,\nthank you very much for your valuable comments!\nWe completely agree with you that the impact of our study is limited by the size of the data set. Especially in unsupervised learning of dominant valve shapes, the data set should represent the realistic variance of all possible valve shapes. Hence, one could argue that 56 valves are not sufficient for this. Additionally, 168 leaflet images are a very small data set for training a convolutional auto-encoder (even though we could show that data augmentation does not enhance the method\u2019s accuracy). We will integrate this limitation into the discussion section to make this issue clearer!\n\nWe also agree that the presented comparison only takes quite similar methods into account. To get a better evaluation of our method, we spend now also tested a PCA-decomposition for feature extraction, followed by the same clustering method. With the same number of features (n_z=20), the resulting clustering achieves worse results than our proposed auto-encoder method. Especially in the mean Hausdorff-distance, we observed errors of more than 12 mm. \nThe reason why we decided to compare our results to a PCA approach instead of the feature descriptions you proposed is that the core idea of our work is to derive the prosthesis shapes completely data-driven. By using hand-crafted features like curvature, we introduce a \u201chuman bias\u201d to the model, telling that similar valves should be similar in these specific features. Hence, the benchmark for our model should be a well-understood standard decomposition method, which led us to PCA. We will include this analysis in the results section!\n\nThe idea of including a visual representation of the results is really helpful! To this end, we will include images of the cluster centers for k=3 in the appendix section in order to make the differences between the identified prosthesis shapes easily assessible for the. \n"}, {"title": "Rebuttal Letter Reviewer 3", "comment": "Dear Reviewer,\nThank you very much for your helpful comments! \nUnfortunately, the title will become too long if we integrate the kind of images we used. However, we will make the kind of used images clearer in the introduction and in the abstract. \n\nIncluding a visual representation of the results is a great idea! We prepared images of the cluster centers for k=3 to present in the appendix. Hence, the differences between the identified prosthesis shapes can be assessed by the reader in an intuitive way. \n\nWe also agree on the fact that mm/pixel is the more intuitive way of describing the resolution as it directly gives the step-size of the discretization. We will change this in our manuscript.\n"}, {"title": "Response to rebuttal", "comment": "The authors have addressed all of my primary concerns. \nThe authors performed an additional comparison experiment and clarified several issues that were not clear in the original paper. Based on this rebuttal and future updates to manuscript, I'm deciding to change my rating from 2 to 3. Anyway, from the aspect of clinical application, this paper is worth to be published."}], "comment_replyto": ["BJesxFsA1N", "SJxrcwyUXE", "HyeRDAlRm4", "rylWybGJEV", "BJlsiWbiVV"], "comment_url": ["https://openreview.net/forum?id=BJesxFsA1N&noteId=SylgZ--oVE", "https://openreview.net/forum?id=BJesxFsA1N&noteId=BJlsiWbiVV", "https://openreview.net/forum?id=BJesxFsA1N&noteId=HyxJrbWjNV", "https://openreview.net/forum?id=BJesxFsA1N&noteId=SkxN-zZjN4", "https://openreview.net/forum?id=BJesxFsA1N&noteId=r1llvOnAVV"], "meta_review_cdate": 1551356583780, "meta_review_tcdate": 1551356583780, "meta_review_tmdate": 1551881979597, "meta_review_ddate ": null, "meta_review_title": "Acceptance Decision", "meta_review_metareview": "The paper proposes to cluster valve shapes after mapping the data to a low-dimensional space using typical auto-encorders. The application of such approach seems to be novel and the authors have properly addressed reviewers' comments. Although the idea of clustering in a low-dimensional space to address the curse of high data dimensionality is not necessarily novel, the application is quite interesting.", "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=BJesxFsA1N&noteId=Syll2MIBLV"], "decision": "Accept"}