Datasets:

mfromm
/

AMSR

{"forum": "95R22aaruC", "submission_url": "https://openreview.net/forum?id=gIff2jm4xB", "submission_content": {"keywords": ["aortic dissection", "segmentation", "convolutional neural networks", "deep learning", "3D reconstruction"], "TL;DR": "We introduced a variation on the common U-net architecture, where stacks of slices are inputted into the network instead of individual 2D slices, allowing the network to better learn from neighboring slices.", "track": "full conference paper", "authorids": ["[email protected]", "[email protected]", "[email protected]", "[email protected]", "[email protected]", "[email protected]", "[email protected]", "[email protected]", "[email protected]", "[email protected]"], "title": "Context Aware Convolutional Neural Networks for Segmentation of Aortic Dissection", "authors": ["Bradley Feiger", "Erick Lorenzana-Saldivar", "Roarke Horstmeyer", "Colin Cooke", "Muath Bishawi", "Julie Doberne", "G. Chad Hughes", "David Ranney", "Soraya Voigt", "Amanda Randles"], "paper_type": "both", "abstract": "Three-dimensional (3D) reconstruction of patient-specific arteries is necessary for a variety of medical and engineering fields, such as surgical planning and physiological modeling. These geometries are created by segmenting and stacking hundreds (or thousands) of two-dimensional (2D) slices from a patient scan to form a composite 3D structure. However, this process is typically laborious and can take hours to fully segment each scan. Convolutional neural networks (CNNs) offer an attractive alternative to reduce the burden of manual segmentation, allowing researchers to reconstruct 3D geometries in a fraction of the time. We focused this work specifically on Stanford type B aortic dissection (TBAD), characterized by a tear in the descending aortic wall that creates two channels of blood flow: a normal channel called a true lumen and a pathologic new channel within the wall called a false lumen. While significant work has been dedicated to automated aortic segmentations, TBAD segmentations present unique challenges due to their irregular shapes, the need to distinguish between the two lumens, and patient to patient variability in the false lumen contrast. Here, we introduced a variation on the U-net architecture, where small stacks of slices are inputted into the network instead of individual 2D slices. This allowed the network to take advantage of contextual information present within neighboring slices. We compared and evaluated this variation with a variety of standard CNN segmentation architectures and found that our stacked input structure significantly improved segmentation accuracy for both the true and false lumen by more than ~12%. The resulting segmentations allowed for more accurate 3D reconstructions which closely matched our manual results.", "paperhash": "feiger|context_aware_convolutional_neural_networks_for_segmentation_of_aortic_dissection", "supplementary_material": "/attachment/224133ee5a726e1c19bbef572da4d4ef9d7f0cc9.zip", "pdf": "/pdf/a9b07493eaf444e64eeac867f41b0ea4aace6b57.pdf", "_bibtex": "@misc{\nfeiger2020context,\ntitle={Context Aware Convolutional Neural Networks for Segmentation of Aortic Dissection},\nauthor={Bradley Feiger and Erick Lorenzana-Saldivar and Roarke Horstmeyer and Colin Cooke and Muath Bishawi and Julie Doberne and G. Chad Hughes and David Ranney and Soraya Voigt and Amanda Randles},\nyear={2020},\nurl={https://openreview.net/forum?id=gIff2jm4xB}\n}"}, "submission_cdate": 1579955751505, "submission_tcdate": 1579955751505, "submission_tmdate": 1587172170913, "submission_ddate": null, "review_id": ["Rdamh15eM-", "0YgPmzt2U", "GebZlkNGH", "fJmUnWsq6Z"], "review_url": ["https://openreview.net/forum?id=gIff2jm4xB&noteId=Rdamh15eM-", "https://openreview.net/forum?id=gIff2jm4xB&noteId=0YgPmzt2U", "https://openreview.net/forum?id=gIff2jm4xB&noteId=GebZlkNGH", "https://openreview.net/forum?id=gIff2jm4xB&noteId=fJmUnWsq6Z"], "review_cdate": [1584188341942, 1584048741597, 1582562049284, 1582277117967], "review_tcdate": [1584188341942, 1584048741597, 1582562049284, 1582277117967], "review_tmdate": [1585229675686, 1585229675146, 1585229674651, 1585229674149], "review_readers": [["everyone"], ["everyone"], ["everyone"], ["everyone"]], "review_writers": [["MIDL.io/2020/Conference/Paper247/AnonReviewer2"], ["MIDL.io/2020/Conference/Paper247/AnonReviewer1"], ["MIDL.io/2020/Conference/Paper247/AnonReviewer3"], ["MIDL.io/2020/Conference/Paper247/AnonReviewer4"]], "review_reply_count": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "review_replyto": ["95R22aaruC", "95R22aaruC", "95R22aaruC", "95R22aaruC"], "review_content": [{"title": "Differences with U-Net should be better explained", "paper_type": "methodological development", "summary": "The authors propose a deep learning architecture based on the well established U-Net architecture where they propose to stack neighboring images of the image to segment to add context. They evaluate their method 21 images of type B aorta dissection and compare their results with the state of the art techniques.", "strengths": "* Good comparison with other state of the art methods\n* Clear explanation of the problem\n* Discussion on the limitations of the method. The authors talk about the problem of manual annotations coming from different users. This is an interesting point. I wonder if they experimented with this. ", "weaknesses": "* The contributions of the paper should be better explained as there are very subtle differences with U-Net. It should be made very clear what makes this method different from U-Net: \n 1) The added value of 3D SU Net is not clear. Adding the fourth channel seems trivial as has no real effect.\n 2) Isn't the 2D SU Net a sort of variation of the standard 3D UNet?\n* The ground truth is obtained from an automated tool. Why don't just using it?", "questions_to_address_in_the_rebuttal": "You say you use an existing tool to obtain automatically the segmentations you use for ground truth. Why is there the need for your method? \nThe differences between your approach and a Unet are very subtle. Please explain this better.", "rating": "2: Weak reject", "justification_of_rating": "I think this paper is well-written and has some \n\nBut the motivation has a flaw: the authors are using an already existing tool to obtain their ground truth automatically. Some times this fails and then it has to be corrected. Is there then a need for the proposed method? I assume that, if the final segmentations were gonna be used, they would have to do some manual corrections as well. Why not staying with what is there? The authors should quantify the advantages between the new method and the commercial tool.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": [], "special_issue": "no"}, {"title": "Results look good, but very weak baseline", "paper_type": "validation/application paper", "summary": "The authors propose to segment aortic dissections in CTA images using a U-Net. To provide more consistent segmentation results, the U-Net takes in multiple slices at the same time as separate channels. Results look good for the proposed method, and true lumen and false lumen are separated quite well.", "strengths": "-\tThe paper is quite well written and figures look nice.\n-\tResults for the SU-Net look good.\n-\tSegmentation of aorta dissections is a clinically relevant problem.\n- The authors provide a comparison to other CNN architectures.", "weaknesses": "-\tThe main contribution of the paper seems to be to use stacks of slices as inputs to a CNN instead of single slices, to enable smoothness between segmentations in adjacent slices. This is not a very novel idea, which has been extensively studied before 3D CNNs became widely used. Moreover, by using these slices as channels, valuable signal information along the z-axis might be missed.\n-\tThe authors used a small data set of 21 patients. The proposed method is compared to a 3D U-net and V-net, but the input sizes for the 2D and 3D networks are very different. The 3D networks take 128x128x128 voxels as input, while the 2D networks only take a couple of slices as input. Hence, the chance of overfitting is a lot larger for the 3D networks. Results in Table 1 and Figure 5 are surprisingly poor for the 3D networks. A fairer comparison would take same-sized inputs for the multi-slice and 3D networks. \n-\tResults in Fig. 4 are not very convincing, why is there a drop in performance at 5 slices? Were these experiments repeated multiple times with different seeds? \n", "questions_to_address_in_the_rebuttal": "3D U-net/V-net and SU-Net should be retrained on comparable inputs.", "detailed_comments": "-\tIt\u2019s unclear what the authors mean with \u2018context aware convolutional neural networks\u2019 in the title. \n-\tIn the introduction, the authors suggest that patient-specific artery models are usually extracted by slice-by-slice segmentation. In fact, there are much more efficient methods for that based on e.g. central lumen lines.\n-\tIt\u2019s not very logical to evaluate using both the Dice and the Jaccard coefficient, as each can be written in terms of the other.\n-\tThe description of pre-processing is unclear. What do the authors mean with \u2018brightened the image for better visualization of the aorta with Mimics\u2019?\n-\tImages were severely downsampled to 128x128 pixels in-plane resolution, why not just use the original resolution?\n", "rating": "2: Weak reject", "justification_of_rating": "The results look nice, but I think there is little novelty in the proposed method. The paper could have value as a systematic comparison of slice-stack inputs to 3D inputs, but I doubt whether the 3D CNNs were used correctly. The results that the authors provide here for 3D methods are very poor. ", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": [], "special_issue": "no"}, {"title": "2.5D U-net applied to aortic dissections, very minor contribution", "paper_type": "validation/application paper", "summary": "The \"stacked U-net\" (\"SU-net\") is just a regular U-net applied to multiple slices at once. This is typically called 2.5D input. The only very minor difference is that they also let the network output several slices (but only use the middle slice at inference time).\n\nApparently, the authors were content with their results, but they're not evaluated on a public dataset, so it is hard to judge whether the comparison is fair. The outputs of the methods compared with look surprisingly bad to me.", "strengths": "The paper is well-written and easy to understand and contains useful illustrations.\n\nThe task & dataset is interesting, but unfortunately neither large, nor public.\n\nThe authors have put less important things into an appendix.", "weaknesses": "The SU-net is hardly novel; it is mostly a standard U-net (with reduced numbers of filters) in a \"2.5D\" setting. In my personal experience, such an approach often does not perform better than a 2D U-net, let alone properly trained 3D U-nets. The multi-slice loss is a stronger supervision than the common one, but the difference is not separately evaluated.\n\nApparently, the authors used padded convolutions (they do not describe padding procedures, which would have been very important for the 3D SU-net when applied on 3 slices only), which is not a good choice for segmentation tasks, since the results depend on the position within the image.\n\nThe authors write that \"[3D U-net and V-net] input whole volumes of medical image data \u2026\" which I interpret as an additional clue that they are not familiar with a proper overlapping tile strategy. (U-nets are fully convolutional and can therefore be applied to properly padded subregions of the input image, while allowing to stitch together their outputs to a complete result without any artifacts from this tiling.)\n\nThe authors claim \"significant\" improvements, but no statistical tests were performed or described.\n\nSome space is wasted by plotting both DC and JC.\n\nFor some reason, CT slices without aortic dissection were discarded (instead of using them as negative examples).", "detailed_comments": "It is not necessary to measure both Dice and Jaccard, because they're redundant / bijectively mappable:\nJC=DC/(2-DC)\nDC=2JC/(1+JC)\nSurface distances (including the HD specified), though, give complementary information.\n\nV-net is not an adaptation of 3D U-net (but an independent development based on the 2D U-net).", "rating": "1: Strong reject", "justification_of_rating": "- The contribution is very minor.\n- The paper is somehow borderline between weak and strong reject for me, because there are not too many peer-reviewed papers on TBAD segmentation yet. On the other hand, some of the preprints and reports I could find contain more technological novelty than this one.\n- There are some technical flaws that limit the paper's contribution.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": [], "special_issue": "no"}, {"title": "Well written paper, but methodological and applicational below MIDL threshold", "paper_type": "both", "summary": "The authors work on aorta segmentation from computed tomography angiography with focus on aortic dissections of type B, where the aorta decomposes into two channels, the true and the false lumen. They present a variant of the well known U-Net where instead of one slice a stack with extra added neighboring slices is passed to the net. These neighboring slices deliver some volumetric contextual information and thus improve segmentation performance. The method is compared to other CNN segmentation approaches on 21 scans with manual ground truth segmentations and shows better performance than these.", "strengths": "The paper is well written, easy to read and its overall structure is sound. The accuracy measures comprise of set-theoretic ones (Dice and Jacard coefficient) as well as distance based ones (Hausdorff distance) allowing for a more holistic assessment of segmentation performance. A parameter study w.r.t. number of neighboring slices was carried out to justify the respective choice.", "weaknesses": "Method:\n\nThe proposed variation of the U-Net is incremental and -more important- the idea lacks novelty to some extend. Please see:\n\nAmbellan, Felix, et al. \"Automated segmentation of knee bone and cartilage combining statistical shape knowledge and convolutional neural networks\"\nMedical image analysis 52 (2019): 109-118.\n\nAmbellan et al. are also employing neighboring slices within a U-Net framework (see Sec. 3.1) with a slight difference w.r.t. the output of the net. The proposed net outputs a segmentation mask for the whole stack, whereas Ambellan et al. output a mask for the center slice only. However, for testing the proposed method also utilizes the center slice only and discards the others.\n\nValidation:\n\nThe authors mention the work of Cao et al. \n\nCao, Long, et al. \"Fully automatic segmentation of type B aortic dissection from CTA images enabled by deep learning.\" \nEuropean journal of radiology 121 (2019): 108713.\n\nin Sec. 1.1 (Related work) several times. Amongst others they state: '...In the work by Cao et al. (2019), the authors used a 3D U-net to segment the true and false lumen...'. In Sec. 4.1 (Study limitations) the authors again refer to Cao et al. who wrote that different experts achieve a dice coefficient of 0.92-0.94 when segmenting the same data. Reading this I wonder why the authors don't discuss the results of Cao et al. and/or compare to Cao et al.'s method directly (the 3D U-Net used in the comparison is not the one of Cao et al.), since they work on exact the same task. The reported dice coefficients of Cao et al. (0.93(+-0.01) for the whole aorta, 0.93(+-0.01) for the true lumen and 0.91(+-0.02) for the false lumen) are superior to the highest given in the proposed manuscript (0.92 w.a., 0.82 t.l., 0.87 f.l.), especially for the true/false lumen. This should not be ignored.", "questions_to_address_in_the_rebuttal": "In what way do you consider the remaining methodological contribution (compared to Ambellan et al.) reaching a level of novelty and/or a threshold of significance necessary for MIDL?\n\nWhy don't you compare to the work and results of Cao et al.?", "rating": "1: Strong reject", "justification_of_rating": "The paper comes with two major weaknesses that concern both aspects of the contribution. For the methodological one I hardly see any quick fix and the applicational aspect would at least require some additional evaluation that does not fit the review/rebuttal time frame, consequently my rating has to be a strong reject.", "confidence": "5: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "recommendation": [], "special_issue": "no"}], "comment_id": ["FWbC7iZCwDg", "QeY2jLrpWgi", "JwiM4Jkbcmc", "j4O9UZURxm", "N-sNV9FUSLu", "Z2J07qv-Wz0"], "comment_cdate": [1585639217753, 1585356891068, 1585356849220, 1585356790338, 1585356713920, 1585356639408], "comment_tcdate": [1585639217753, 1585356891068, 1585356849220, 1585356790338, 1585356713920, 1585356639408], "comment_tmdate": [1585639217753, 1585585762281, 1585585733082, 1585585706964, 1585585681418, 1585585670161], "comment_readers": [["everyone"], ["everyone"], ["everyone"], ["everyone"], ["everyone"], ["everyone"]], "comment_writers": [["MIDL.io/2020/Conference/Paper247/AnonReviewer3", "MIDL.io/2020/Conference"], ["MIDL.io/2020/Conference/Paper247/Authors", "MIDL.io/2020/Conference"], ["MIDL.io/2020/Conference/Paper247/Authors", "MIDL.io/2020/Conference"], ["MIDL.io/2020/Conference/Paper247/Authors", "MIDL.io/2020/Conference"], ["MIDL.io/2020/Conference/Paper247/Authors", "MIDL.io/2020/Conference"], ["MIDL.io/2020/Conference/Paper247/Authors", "MIDL.io/2020/Conference"]], "comment_reply_content": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "comment_content": [{"title": "Improvements go in the right direction, but comparison is still unfair", "comment": "Re comment 1: Asymmetric stacks add only marginal novelty, but go well with the main direction of the paper.\n\nRe comment 2: Applying 3D networks to stacks is not the same as a proper overlapping tile strategy. The main (and absolutely crucial!) point is that the fully convolutional architectures are not fed the correct context to be able to perform properly. These networks are absolutely capable of processing arbitrarily large volumes with high resolution and good performance, but the tiling is important to get right (and one must not use padded convolutions).\n\nRe comment 4: You're assuming a normal distribution then? Please check your assumption or use different statistics.\n\nMy comments have been taken into consideration, and many smaller issues are addressed, but a main remaining problem are the bad baseline implementations compared against."}, {"title": "Response to reviewer #4", "comment": "Response to reviewer #4\nReviewer comment 1: The proposed variation of the U-Net is incremental and -more important- the idea lacks novelty to some extend. Please see: Ambellan, Felix, et al. \"Automated segmentation of knee bone and cartilage combining statistical shape knowledge and convolutional neural networks\" Medical image analysis 52 (2019): 109-118. Ambellan et al. are also employing neighboring slices within a U-Net framework (see Sec. 3.1) with a slight difference w.r.t. the output of the net. The proposed net outputs a segmentation mask for the whole stack, whereas Ambellan et al. output a mask for the center slice only. However, for testing the proposed method also utilizes the center slice only and discards the others.\n\nResponse to comment 1: We added several references to papers that use these stacked U-nets, including the reference to Ambellan et al., and also renamed our stacked 2D U-net to 2.5D U-net. To add novelty to the paper, we added two new studies that looked at how asymmetric stacks impact results. Specifically, we included stacks where the slice of interest was located at the bottom or the top of the stack, instead of always being at the center. The asymmetric stack results are included in the updated paper. We also believe that part of the novelty lies in the application of TBAD which has very limited existing studies.\n\nReviewer comment 2: The authors mention the work of Cao et al. Cao, Long, et al. \"Fully automatic segmentation of type B aortic dissection from CTA images enabled by deep learning.\" European journal of radiology 121 (2019): 108713. in Sec. 1.1 (Related work) several times. Amongst others they state: '...In the work by Cao et al. (2019), the authors used a 3D U-net to segment the true and false lumen...'. In Sec. 4.1 (Study limitations) the authors again refer to Cao et al. who wrote that different experts achieve a dice coefficient of 0.92-0.94 when segmenting the same data. Reading this I wonder why the authors don't discuss the results of Cao et al. and/or compare to Cao et al.'s method directly (the 3D U-Net used in the comparison is not the one of Cao et al.), since they work on exact the same task. The reported dice coefficients of Cao et al. (0.93(+-0.01) for the whole aorta, 0.93(+-0.01) for the true lumen and 0.91(+-0.02) for the false lumen) are superior to the highest given in the proposed manuscript (0.92 w.a., 0.82 t.l., 0.87 f.l.), especially for the true/false lumen. This should not be ignored.\n\nResponse to comment 2: To address this comment, we discussed the results of Cao et al. in the discussion and reflected that many of our images contain very poorly contrasted regions whereas, in the work of Cao et al., all images shown are well-contrasted. The 3D U-net implemented by Cao et al. did have a different structure than ours, although we did test the impact of cascading networks on the best performing network and did not see an improvement.\n\nReviewer comment 3: In what way do you consider the remaining methodological contribution (compared to Ambellan et al.) reaching a level of novelty and/or a threshold of significance necessary for MIDL?\n\nResponse to comment 3: Part of the novelty of this work was the aortic dissection application as little work has previously addressed segmenting this disease. To add further novelty to the work, we added two more variations of stacked U-nets such that the stacks were asymmetric with respect to the slice of interest. Specifically, we included stacks where the slice of interest was located at the bottom or the top of the stack, instead of always being at the center. The asymmetric stack results are included in the updated paper. We believe that this addition as well as the little amount of work that exists with TBAD provides significant novelty.\n\nReviewer comment 4: Why don't you compare to the work and results of Cao et al.\n\nResponse to comment 4: See response to comment 2."}, {"title": "Response to reviewer #3", "comment": "Response to reviewer #3\nReviewer comment 1: The SU-net is hardly novel; it is mostly a standard U-net (with reduced numbers of filters) in a \"2.5D\" setting. In my personal experience, such an approach often does not perform better than a 2D U-net, let alone properly trained 3D U-nets. The multi-slice loss is a stronger supervision than the common one, but the difference is not separately evaluated.\n\nResponse to comment 1: We added several references to papers that use 2.5D U-nets and also renamed our stacked 2D U-net to 2.5D U-net. To add novelty to the paper, we added two new studies that looked at how asymmetric stacks impact results. Specifically, we included stacks where the slice of interest was located at the bottom or the top of the stack, instead of always being at the center. The asymmetric stack results are included in the updated paper. We also believe that part of the novelty lies in the application of TBAD which has very limited existing studies.\n\nReviewer comment 2: The authors write that \"[3D U-net and V-net] input whole volumes of medical image data \u2026\" which I interpret as an additional clue that they are not familiar with a proper overlapping tile strategy. (U-nets are fully convolutional and can therefore be applied to properly padded subregions of the input image, while allowing to stitch together their outputs to a complete result without any artifacts from this tiling.)\n\nResponse to comment 2: To address this comment, we added studies testing the 3D U-net and V-net networks on the stacked input data structure in addition to the whole volumes. This additional data provided a more insightful comparison with the stacked 2D U-net because the sample sizes were made consistent. These results are almost completed and will be included in the updated paper.\n\nReviewer comment 3: Some space is wasted by plotting both DC and JC.\n\nResponse to comment 3: We removed the Jaccard coefficient from the paper and included an average distance calculation. The average distance is a metric suitable for comparing 3D geometries without being highly subject to outliers.\n\nReviewer comment 4: The authors claim \"significant\" improvements, but no statistical tests were performed or described.\n\nResponse to comment 4: We will add standard deviations to all of the results and will work to include Student\u2019s t-tests to test for differences in the means.\n\nReviewer comment 5: For some reason, CT slices without aortic dissection were discarded (instead of using them as negative examples)\n\nResponse to comment 5: CT slices without aortic dissection were in fact included in the pipeline. See the first geometry in figure 6 for example that shows the ascending and lower descending aorta without any dissection. We did discard side branches from the aorta as we wanted to maintain focus on the aortic dissection.\n\nReviewer comment 6: V-net is not an adaptation of 3D U-net (but an independent development based on the 2D U-net).\n\nResponse to comment 6: We adjusted the language to reflect that V-net is not an adaptation of 3D U-net."}, {"title": "Response to reviewer #2", "comment": "Response to reviewer #2\nReviewer comment 1: The main contribution of the paper seems to be to use stacks of slices as inputs to a CNN instead of single slices, to enable smoothness between segmentations in adjacent slices. This is not a very novel idea, which has been extensively studied before 3D CNNs became widely used. Moreover, by using these slices as channels, valuable signal information along the z-axis might be missed.\n\nResponse to comment 1: We believe the novelty of the work lies partially with the aortic dissection application as very little existing work has addressed this disease. To further extend the novelty, we added two more variations of stacked U-nets such that the stacks were asymmetric with respect to the slice of interest. Specifically, we included stacks where the slice of interest was located at the bottom or the top of the stack, instead of always being at the center. The asymmetric stack results are included in the updated paper.\n\nReviewer comment 2: The authors used a small data set of 21 patients. The proposed method is compared to a 3D U-net and V-net, but the input sizes for the 2D and 3D networks are very different. The 3D networks take 128x128x128 voxels as input, while the 2D networks only take a couple of slices as input. Hence, the chance of overfitting is a lot larger for the 3D networks. Results in Table 1 and Figure 5 are surprisingly poor for the 3D networks. A fairer comparison would take same-sized inputs for the multi-slice and 3D networks.\n\nResponse to comment 2: To address this comment, we added studies testing the 3D U-net and V-net networks on the stacked input data structure in addition to the whole volumes. This gave us a better comparison with the stacked 2D U-net because the sample sizes are now the same. These results are almost completed and will be included in the updated paper.\n\nReviewer comment 3: Results in Fig. 4 are not very convincing, why is there a drop in performance at 5 slices? Were these experiments repeated multiple times with different seeds?\n\nResponse to comment 3: We did not previously test with different seeds, but we did use three-fold cross-validation. We are currently testing with different seeds in just one of the networks to see if that impacts results.\n\nReviewer comment 4: 3D U-net/V-net and SU-Net should be retrained on comparable inputs.\n\nResponse to comment 4: See response to comment 2.\n\nReviewer comment 5: It\u2019s unclear what the authors mean with \u2018context aware convolutional neural networks\u2019 in the title.\n\nResponse to comment 5: We were referring to the fact that the networks we implemented and tested use contextual information in the vertical direction. We clarified this in the introduction.\n\nReviewer comment 6: In the introduction, the authors suggest that patient-specific artery models are usually extracted by slice-by-slice segmentation. In fact, there are much more efficient methods for that based on e.g. central lumen lines.\n\nResponse to comment 6: We added a sentence to the introduction about central lumen lines and also included a citation.\n\nReviewer comment 7: It\u2019s not very logical to evaluate using both the Dice and the Jaccard coefficient, as each can be written in terms of the other.\n\nResponse to comment 7: We removed the Jaccard coefficient from the paper and include an average distance calculation. The average distance is a metric suitable for comparing 3D geometries without being highly subject to outliers.\n\nReviewer comment 8: The description of pre-processing is unclear. What do the authors mean with \u2018brightened the image for better visualization of the aorta with Mimics\u2019?\n\nResponse to comment 8: We agree that this sentence is misleading. We were simply referring to the fact that we brightened the image to assist with manual segmentation. As this action was not a part of the pre-processing phase for inputs to the networks, we removed the sentence to reduce the chance of confusion.\n\nReviewer comment 9: Images were severely down-sampled to 128x128 pixels in-plane resolution, why not just use the original resolution?\n\nResponse to comment 9: Images were down-sampled to make the training times feasible. Also, 3D U-net and V-net were constrained by memory when inputting whole volumes, and we wanted the dimensions of each slice to be consistent."}, {"title": "Response to reviewer #1", "comment": "Response to reviewer #1\nReviewer comment 1: You say you use an existing tool to obtain automatically the segmentations you use for ground truth. Why is there the need for your method? \n\nResponse to comment 1: The tool that we used can only automate small parts of the segmentation that have good contrast and regular shapes. This is particularly a problem with TBAD since many of the images have poor contrast and irregular shapes, and therefore, manual segmentation was required to get the ground truth. We have added details to clarify this in the methods section. \n\nReviewer comment 2: The differences between your approach and a Unet are very subtle. Please explain this better.\n\nResponse to comment 2: Our approach takes small stacks of images as input as opposed to the conventional 2D U-net which relies on single images. To investigate this complexity further, we have added studies that looked at asymmetric stacks such that each stack included the slice of interest and other slices in either the upward or downward direction, but not both.\n\nReviewer comment 3: But the motivation has a flaw: the authors are using an already existing tool to obtain their ground truth automatically. Some times this fails and then it has to be corrected. Is there then a need for the proposed method? I assume that, if the final segmentations were gonna be used, they would have to do some manual corrections as well. Why not staying with what is there? The authors should quantify the advantages between the new method and the commercial tool.\n\nResponse to comment 3: The reviewer is correct that some of the segmentations will need to be corrected even after the network is used. However, the number of segmentations that will need to be corrected is far fewer than if we only used the automated tools in the commercial software. Specifically, each manually segmented aortic dissection required between 8 and 10 hours of work due to the poor contrast in the false lumen. The network outputted segmentations only required a few minutes of additional edits. We added two sentences to address this comment in the discussion."}, {"title": "General response summary to all reviewers", "comment": "Summary (to all reviewers): We would like to thank the reviewers for their comments which we used to improve our manuscript. In summary, we were able to address some of the reviewer\u2019s comments with changes to the text and inclusion of a few studies. In fact, many of the reviewers\u2019 comments were items that we had already accomplished but did not include in the paper and some of the other comments are items we have been working on since submission. The two major items that we added to paper are listed below. We believe the two items add novelty and add better comparisons between the networks, which were the main comments of the reviewers. We\u2019d be happy to provide figures if requested.\n\n\u2022\tThe first study we added included testing the impacts of asymmetric stacks on segmentation results. In the submitted paper, the stacks were symmetric with respect to the slice of interest such that we predicted the center slice of each stack. We added variations with asymmetric stacks such that the stacks include slices either above the slice of interest or below the slice of interest. In cases where the stack was above the slice of interest, we predicted the segmentation results on the bottom slice, and in cases where the stack was below the slice of interest, we predicted segmentation results on the top slice. We found that stacks above the slice of interest improved segmentation results with an optimal segmentation at 11 slices in the stack. We believe this study adds another element of novelty to our work.\n\u2022\tThe second major change we made was using 3D U-net and V-net on the image stacks rather than only on the whole geometries. This allowed us to have the same sample sizes when using 3D U-net and V-net compared with our stacked 2D U-net. We also tested asymmetric stacks with these networks. Currently, our V-net networks are training and will be finished within 48 hours. So far, we found that 3D U-net produced slightly worse segmentation accuracy than 2D U-net on the image stacks, and V-net results will be analyzed as soon as the networks finish training. We believe this component adds a better comparison between networks.\n\nThe remainder of the changes we made to the manuscript are described in our response to each specific reviewer."}], "comment_replyto": ["JwiM4Jkbcmc", "fJmUnWsq6Z", "GebZlkNGH", "0YgPmzt2U", "Rdamh15eM-", "95R22aaruC"], "comment_url": ["https://openreview.net/forum?id=gIff2jm4xB&noteId=FWbC7iZCwDg", "https://openreview.net/forum?id=gIff2jm4xB&noteId=QeY2jLrpWgi", "https://openreview.net/forum?id=gIff2jm4xB&noteId=JwiM4Jkbcmc", "https://openreview.net/forum?id=gIff2jm4xB&noteId=j4O9UZURxm", "https://openreview.net/forum?id=gIff2jm4xB&noteId=N-sNV9FUSLu", "https://openreview.net/forum?id=gIff2jm4xB&noteId=Z2J07qv-Wz0"], "meta_review_cdate": 1586026542971, "meta_review_tcdate": 1586026542971, "meta_review_tmdate": 1586026542971, "meta_review_ddate ": null, "meta_review_title": "MetaReview of Paper247 by AreaChair1", "meta_review_metareview": "The reviewers agree that the paper is well written and has a number of strengths, but they also agree that it has little methodological novelty and also point to several methodological flaws. The authors have provided extensive replies to the criticism, which I really appreciate. In these replies some results and modifications are promised, while it would be stronger if concrete text was proposed. Overall, my evaluation is that the modifications unfortunately do not seem to add major novelty or significantly change the manuscript. ", "meta_review_readers": ["everyone"], "meta_review_writers": ["MIDL.io/2020/Conference/Program_Chairs", "MIDL.io/2020/Conference/Paper247/Area_Chairs"], "meta_review_reply_count": {"replyCount": 0}, "meta_review_url": ["https://openreview.net/forum?id=gIff2jm4xB&noteId=VQ8N8ZwgPo8"], "decision": "accept"}