Datasets:
Tasks:
Text Classification
Modalities:
Text
Formats:
text
Languages:
English
Size:
10K - 100K
License:
{"forum": "H1gMYCQxg4", "submission_url": "https://openreview.net/forum?id=H1gMYCQxg4", "submission_content": {"title": "Adversarial Pseudo Healthy Synthesis Needs Pathology Factorization", "authors": ["Tian Xia", "Agisilaos Chartsias", "Sotirios A. Tsaftaris"], "authorids": ["tian.xia@ed.ac.uk", "agis.chartsias@ed.ac.uk", "s.tsaftaris@ed.ac.uk"], "keywords": ["pseudo healthy synthesis", "GAN", "cycle-consistency", "factorization"], "TL;DR": "We propose a pseudo healthy synthesis method using adversarial learning and pathology factorization.", "abstract": "Pseudo healthy synthesis, i.e. the creation of a subject-specific `healthy' image from a pathological one, could be helpful in tasks such as anomaly detection, understanding changes induced by pathology and disease or even as data augmentation. We treat this task as a factor decomposition problem: we aim to separate what appears to be healthy and where disease is (as a map). The two factors are then recombined (by a network) to reconstruct the input disease image. We train our models in an adversarial way using either paired or unpaired settings, where we pair disease images and maps (as segmentation masks) when available. We quantitatively evaluate the quality of pseudo healthy images. We show in a series of experiments, performed in ISLES and BraTS datasets, that our method is better than conditional GAN and CycleGAN, highlighting challenges in using adversarial methods in the image translation task of pseudo healthy image generation.", "pdf": "/pdf/2c5eef74a56af650fb6ed038f18bca434c14ae80.pdf", "code of conduct": "I have read and accept the code of conduct.", "remove if rejected": "(optional) Remove submission if paper is rejected.", "paperhash": "xia|adversarial_pseudo_healthy_synthesis_needs_pathology_factorization", "_bibtex": "@inproceedings{xia:MIDLFull2019a,\ntitle={Adversarial Pseudo Healthy Synthesis Needs Pathology Factorization},\nauthor={Xia, Tian and Chartsias, Agisilaos and Tsaftaris, Sotirios A.},\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=H1gMYCQxg4},\nabstract={Pseudo healthy synthesis, i.e. the creation of a subject-specific `healthy' image from a pathological one, could be helpful in tasks such as anomaly detection, understanding changes induced by pathology and disease or even as data augmentation. We treat this task as a factor decomposition problem: we aim to separate what appears to be healthy and where disease is (as a map). The two factors are then recombined (by a network) to reconstruct the input disease image. We train our models in an adversarial way using either paired or unpaired settings, where we pair disease images and maps (as segmentation masks) when available. We quantitatively evaluate the quality of pseudo healthy images. We show in a series of experiments, performed in ISLES and BraTS datasets, that our method is better than conditional GAN and CycleGAN, highlighting challenges in using adversarial methods in the image translation task of pseudo healthy image generation.},\n}"}, "submission_cdate": 1544728186506, "submission_tcdate": 1544728186506, "submission_tmdate": 1561398531641, "submission_ddate": null, "review_id": ["S1ecHtPkVE", "HJl1TEij7E", "SkxThtvi7N"], "review_url": ["https://openreview.net/forum?id=H1gMYCQxg4¬eId=S1ecHtPkVE", "https://openreview.net/forum?id=H1gMYCQxg4¬eId=HJl1TEij7E", "https://openreview.net/forum?id=H1gMYCQxg4¬eId=SkxThtvi7N"], "review_cdate": [1548872002267, 1548625078650, 1548609972895], "review_tcdate": [1548872002267, 1548625078650, 1548609972895], "review_tmdate": [1548872002267, 1548856745594, 1548856743457], "review_readers": [["everyone"], ["everyone"], ["everyone"]], "review_writers": [["MIDL.io/2019/Conference/Paper91/AnonReviewer3"], ["MIDL.io/2019/Conference/Paper91/AnonReviewer1"], ["MIDL.io/2019/Conference/Paper91/AnonReviewer2"]], "review_reply_count": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "review_replyto": ["H1gMYCQxg4", "H1gMYCQxg4", "H1gMYCQxg4"], "review_content": [{"pros": "This paper presents an interesting work on pathological image synthesis by developing adversarial learning models. The paper is well-written and organized for readers to follow. Experimental results show that the proposed model performs better than other baseline methods: conditional GAN and CycleGAN. ", "cons": "While the authors show that the proposed method has better synthesized image quality than other baseline algorithms, the synthesized image quality is substantially worse than the original image (as shown in figure 4). Many of the \u2018healthy\u2019 parts of the image either introduce artifacts, or with over-smoothed structures. In particular, artifacts that seem like \u2018check-board pattern\u2019 appear in the pathological areas. \n \nBased on the experimental results, I am wondering whether this problem could be solved just by segmentation itself since reconstruction on pathological areas is not considered. It is trivial to adjust the contrast of the segmented pathology to better match the intensity distribution of healthiness. In this case, you shall perfectly preserve the image information on healthy parts. ", "rating": "3: accept", "confidence": "2: The reviewer is fairly confident that the evaluation is correct"}, {"pros": "A well-written paper proposing an adversarial network for pseudo-healthy image synthesis by explicitly separating the healthy and pathology domain. Method has been compared to two baseline methods (CycleGan and conditional GAN) on two publicly available datasets with superior results evaluated using their newly proposed metrics for 'healthiness' (measuring size of predicted pathology) and for 'identity' (measuring similarity of non-pathological regions). This is very interesting approach to an important problem, particularly by the lack of desired image pairs (image with and without pathology of same patient) and potential wide number of applications. ", "cons": "- Implicitly, their approach does a detection and segmentation of pathology. Why not evaluating their method on how well this part has been done, and not only on healthiness and identity?\n- how were both cycles (Figure 3) trained? Does the following order matter?\n- it is unclear if the method was trained and evaluated using the images as 3D or 2D slices. I suppose as 2D otherwise the number of patient data seem too limited\n- I suggest to have expert reader or radiologist to evaluate how well the healthy synthesis has succeeded\n", "rating": "3: accept", "confidence": "2: The reviewer is fairly confident that the evaluation is correct", "oral_presentation": ["Consider for oral presentation"]}, {"pros": "The paper proposes a deep learning framework for pseudo healthy synthesis based on the factorization of pathological and anatomical information. The network is trained following two different settings namely, the paired and unpaired. To enable quantitative performance evaluation the \u201chealthiness\u201d and the \u201cidentity\u201d metrics are proposed. The method has been validated on two different datasets and its performance has been compared to two baselines, the conditional GAN and the CycleGAN.\n\nThis is an interesting work which fits well to the scope of the conference. The paper is well written and easy to follow. The contributions of the paper have been clearly defined. The presented work is of sufficient technical novelty and seems technically and theoretically sound. The references are adequate. The figures could be improved as it is explained below in detail.\n", "cons": "Suggestions for revision\n\n1. In the 4th paragraph in Section 3.3, it is mentioned that \u201ca pathology mask for a real healthy image cannot be defined\u201d. It is not clear why a black mask can not be used in this case. \n\n2. In the experimental results, why has the ISLES dataset been divided so unevenly into the training (22 volumes) and testing sets (6 volumes)?\n\n3. Figures 2 and 4 are quite small, making it difficult to distinguish the differences between the subfigures. The size of these figures should be increased.\n", "rating": "4: strong accept", "confidence": "2: The reviewer is fairly confident that the evaluation is correct"}], "comment_id": ["rJlhDRotNE", "H1gL9l2F4N", "SkggnJnFNN"], "comment_cdate": [1549545059578, 1549545614210, 1549545383822], "comment_tcdate": [1549545059578, 1549545614210, 1549545383822], "comment_tmdate": [1555946021997, 1555946021780, 1555946021565], "comment_readers": [["everyone"], ["everyone"], ["everyone"]], "comment_writers": [["MIDL.io/2019/Conference/Paper91/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper91/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper91/Authors", "MIDL.io/2019/Conference"]], "comment_reply_content": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "comment_content": [{"title": "Response to Reviewer3", "comment": "First of all, we would like to thank you for your positive remarks and appreciation of our work. Below we address the points raised.\n\nThe \u2018check-board\u2019 artifacts are a known problem in computer vision. They are discussed in a recent ICLR paper (J. Oramas et al., 2019) and an earlier distil article by Google Brain (Odena et al., 2016). Both point to the use of deconvolutional layers when downsampling, as the main reason, the use of which we have avoided. However, other operations, such as max pooling or strided convolution in GAN discriminators, are also linked with \u2018check-board\u2019 artifacts. \n\nChecker-board patterns may also have an interplay with our smoothed results. We only use L1 reconstruction loss between an input pathological image and its reconstruction x \u0302_p. The use of L1 reconstruction loss alone can result in a blurry reconstructed image x \u0302_p. Thus, it is possible that checkerboard patterns may \u201cfoul\u201d the discriminator, in the meantime, they are also not penalized by the reconstruction loss (of the cycle). \n\nWe are actively working to improve on these. First, we believe that moving to 3D will help alleviate some of this problem. We also plan to add an adversarial loss on x \u0302_p to make x \u0302_p sharp and, consequently, sensitive to the quality of synthetic x \u0303_h. We expect that adding this adversarial loss will help to alleviate the \u2018check-board\u2019 artifacts. Yet we do note that we need to identify more suitable discriminators to be sensitive to these artifacts.\n\nYour suggested method of adjusting the contrast of just the segmented pathology is a valid idea, and similar in spirit with the work of Simon et al. (2018), although, we do note that the \u2018check-board pattern\u2019 still appears in their synthetic images. More importantly, as we mention in the paper, some pathologies have a global effect on the brain, for example by deforming brain shape. Such a global effect cannot be fixed by just adjusting the intensities within pathological areas. Thus, our proposed method is more general as it has the potential to correct deformations or other global effects. In fact, fixing such global effects of pathology is part of our future plan of extending our model.\n\n \n\nReferences:\nJose Oramas, Kaili Wang, and Tinne Tuytelaars. \"Visual Explanation by Interpretation: Improving Visual Feedback Capabilities of Deep Neural Networks.\" ICLR 2019\n\nOdena, et al., \"Deconvolution and Checkerboard Artifacts\", Distill, 2016. http://doi.org/10.23915/d\n\nSimon Andermatt, Antal Horv \u0301ath, Simon Pezold, and Philippe Cattin. Pathology Segmentation using Distributional Differences to Images of Healthy Origin. Brain-Lesion workshop (BrainLes). MICCAI, 2018. "}, {"title": "Response to Reviewer2", "comment": "First of all, we would like to thank you for your positive remarks and appreciation of our work. Below we address the points raised.\n\nQuestion1: In the 4th paragraph in Section 3.3, it is mentioned that \u201ca pathology mask for a real healthy image cannot be defined\u201d. It is not clear why a black mask can not be used in this case. \n\nResponse: We apologize for the confusion. Indeed, we use a black mask with real healthy images. What we wanted to say is that we do not want to give a random pathology mask to a real healthy image and let the generator synthesize a synthetic pathological image. Therefore, we chose to use a black mask and a healthy image to synthesize an identical healthy image in Cycle H-H. We will rewrite this sentence to express our idea better.\n\nQuestion2: In the experimental results, why has the ISLES dataset been divided so unevenly into the training (22 volumes) and testing sets (6 volumes)?\n\nResponse: The number of volumes in the ISLES dataset is limited. Therefore, in order to make sure that we have enough training images to learn and enough test images for testing we divided the ISLES dataset in ~80% volumes for training and ~20% for testing. We have adopted a similar split into training and testing sets as in V. Sevetlidis (2016) and H. Shin (2018).\n\nQuestion3: Figures 2 and 4 are quite small, making it difficult to distinguish the differences between the subfigures. The size of these figures should be increased.\n\nResponse: We apologize for this. We presented Figures 2 and 4 in this size due to limited page space. We will enlarge the Figures and we will include larger versions of these images in the Appendix of our revised paper.\n\n\n\nReferences:\nVasileios Sevetlidis, Mario Valerio Giuffrida, and Sotirios A. Tsaftaris. \"Whole image synthesis using a deep encoder-decoder network.\" International Workshop on Simulation and Synthesis in Medical Imaging. Springer, Cham, 2016.\n\nHoo-Chang Shin, et al. \"Medical image synthesis for data augmentation and anonymization using generative adversarial networks.\" International Workshop on Simulation and Synthesis in Medical Imaging. Springer, Cham, 2018."}, {"title": "Response to Reviewer1", "comment": "First of all, we would like to thank you for your positive remarks and appreciation of our work. Below we address the points raised.\n\nQuestion1: -Implicitly, their approach does a detection and segmentation of pathology. Why not evaluating their method on how well this part has been done, and not only on healthiness and identity?\n\nResponse: Our approach performs segmentation and synthesis at the same time. However, since the focus of our paper is the synthesis of \u2018pseudo-healthy\u2019 images, we only presented the evaluation results related to synthesized \u2018healthy\u2019 images. To address your comment, we numerically evaluated the trained models on segmentation performance on pathology segmentation using the Dice coefficient. We found that our proposed method (unpaired), our proposed method (paired) and a single U-net (as a comparison) obtain 0.61, 0.80 and 0.82 respectively.\n\nQuestion2:- how were both cycles (Figure 3) trained? Does the following order matter?\n\nResponse: In every epoch, we started by first training Cycle P-H using a batch of pathological images, then trained Cycle H-H using a batch of healthy images, and so on and so forth. At the end of every epoch the images are reshuffled into batches, so the order that the images are presented (for every cycle) changes per epoch. \n\nThe training order may matter in the first few iterations. Suppose that we started with Cycle H-H, where the task is to synthesize an identical image and thus encourage to learn the identity mapping when conditioned on an empty pathology mask. Because the generator uses residual connections, it could be possible that the network uses the residual skips to learn the identity function and nullify many weights, and pushes the solution towards local minima. However, starting with Cycle P-H avoids this potential issue. \n\nQuestion3:- it is unclear if the method was trained and evaluated using the images as 3D or 2D slices. I suppose as 2D otherwise the number of patient data seem too limited\n\nResponse: Our method was trained on 2D slices. In Section 4.1 we mention that \u201cwe choose the middle 60 slices from each volume\u201d, although we do not explicitly specify that our model is trained and evaluated on 2D slices. We will make this clarification early on in the introduction section in our revised paper to avoid confusion.\n\nQuestion4:- I suggest to have expert reader or radiologist to evaluate how well the healthy synthesis has succeeded\n\nResponse: We thank you for your valuable suggestion. We are currently working on an extension of our work from 2D slices to 3D volumes for the task of pseudo healthy synthesis, also considering global pathology effects (e.g. deformations of the brain shape caused by certain pathologies), and plan to involve experts to evaluate our synthetic results.\n"}], "comment_replyto": ["S1ecHtPkVE", "SkxThtvi7N", "HJl1TEij7E"], "comment_url": ["https://openreview.net/forum?id=H1gMYCQxg4¬eId=rJlhDRotNE", "https://openreview.net/forum?id=H1gMYCQxg4¬eId=H1gL9l2F4N", "https://openreview.net/forum?id=H1gMYCQxg4¬eId=SkggnJnFNN"], "meta_review_cdate": 1551356571369, "meta_review_tcdate": 1551356571369, "meta_review_tmdate": 1551881973364, "meta_review_ddate ": null, "meta_review_title": "Acceptance Decision", "meta_review_metareview": "Reviewers agree on acceptance and high quality/originality. Rebuttals are solid and supported with citations.\n\nPros include: well-written, interesting, good results\nCons include: some artifacts in results, which authors are working on now and is known in the field\n\nOne reviewer recommended Oral Presentation. As this is the top paper in my stack I will recommend accept as oral presentation.", "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=H1gMYCQxg4¬eId=H1gXjMIB8V"], "decision": "Accept"} |