{"forum": "ByxLSoblgV", "submission_url": "https://openreview.net/forum?id=ByxLSoblgV", "submission_content": {"title": "Deep Learning Approach to Semantic Segmentation in 3D Point Cloud Intra-oral Scans of Teeth", "authors": ["Farhad Ghazvinian Zanjani", "David Anssari Moin", "Bas Verheij", "Frank Claessen", "Teo Cherici", "Tao Tan", "Peter H. N. de With"], "authorids": ["f.ghazvinian.zanjani@tue.nl", "david@promaton.com", "bas@promaton.com", "frank@promaton.com", "teo@promaton.com", "t.tan1@tue.nl", "p.h.n.de.with@tue.nl"], "keywords": ["Deep learning", "3D point cloud", "intra-oral scan", "semantic segmentation"], "TL;DR": "Deep learning approach to Semantic Segmentation in 3D Point Cloud IOS", "abstract": "Accurate segmentation of data, derived from intra-oral scans (IOS), is a crucial step in a computer-aided design (CAD) system for many clinical tasks, such as implantology and orthodontics in modern dentistry. In order to reach the highest possible quality, a segmentation model may process a point cloud derived from an IOS in its highest available spatial resolution, especially for performing a valid analysis in finely detailed regions such as the curvatures in border lines between two teeth. In this paper, we propose an end-to-end deep learning framework for semantic segmentation of individual teeth as well as the gingiva from point clouds representing IOS. By introducing a non-uniform resampling technique, our proposed model is trained and deployed on the highest available spatial resolution where it learns the local fine details along with the global coarse structure of IOS. Furthermore, the point-wise cross-entropy loss for semantic segmentation of a point cloud is an ill-posed problem, since the relative geometrical structures between the instances (e.g. the teeth) are not formulated. By training a secondary simple network as a discriminator in an adversarial setting and penalizing unrealistic arrangements of assigned labels to the teeth on the dental arch, we improve the segmentation results considerably. Hence, a heavy post-processing stage for relational and dependency modeling (e.g. iterative energy minimization of a constructed graph) is not required anymore. Our experiments show that the proposed approach improves the performance of our baseline network and outperforms the state-of-the-art networks by achieving 0.94 IOU score.", "pdf": "/pdf/3d4a6a90945abedf66e75160858b9bc703664ea7.pdf", "code of conduct": "I have read and accept the code of conduct.", "remove if rejected": "(optional) Remove submission if paper is rejected.", "paperhash": "zanjani|deep_learning_approach_to_semantic_segmentation_in_3d_point_cloud_intraoral_scans_of_teeth", "_bibtex": "@inproceedings{zanjani:MIDLFull2019a,\ntitle={Deep Learning Approach to Semantic Segmentation in 3D Point Cloud Intra-oral Scans of Teeth},\nauthor={Zanjani, Farhad Ghazvinian and Moin, David Anssari and Verheij, Bas and Claessen, Frank and Cherici, Teo and Tan, Tao and With, Peter H. N. de},\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=ByxLSoblgV},\nabstract={Accurate segmentation of data, derived from intra-oral scans (IOS), is a crucial step in a computer-aided design (CAD) system for many clinical tasks, such as implantology and orthodontics in modern dentistry. In order to reach the highest possible quality, a segmentation model may process a point cloud derived from an IOS in its highest available spatial resolution, especially for performing a valid analysis in finely detailed regions such as the curvatures in border lines between two teeth. In this paper, we propose an end-to-end deep learning framework for semantic segmentation of individual teeth as well as the gingiva from point clouds representing IOS. By introducing a non-uniform resampling technique, our proposed model is trained and deployed on the highest available spatial resolution where it learns the local fine details along with the global coarse structure of IOS. Furthermore, the point-wise cross-entropy loss for semantic segmentation of a point cloud is an ill-posed problem, since the relative geometrical structures between the instances (e.g. the teeth) are not formulated. By training a secondary simple network as a discriminator in an adversarial setting and penalizing unrealistic arrangements of assigned labels to the teeth on the dental arch, we improve the segmentation results considerably. Hence, a heavy post-processing stage for relational and dependency modeling (e.g. iterative energy minimization of a constructed graph) is not required anymore. Our experiments show that the proposed approach improves the performance of our baseline network and outperforms the state-of-the-art networks by achieving 0.94 IOU score.},\n}"}, "submission_cdate": 1544719165537, "submission_tcdate": 1544719165537, "submission_tmdate": 1561397974358, "submission_ddate": null, "review_id": ["HkgcroPJNE", "r1eT2kXq7E", "HklgkoI6X4"], "review_url": ["https://openreview.net/forum?id=ByxLSoblgV&noteId=HkgcroPJNE", "https://openreview.net/forum?id=ByxLSoblgV&noteId=r1eT2kXq7E", "https://openreview.net/forum?id=ByxLSoblgV&noteId=HklgkoI6X4"], "review_cdate": [1548872514171, 1548525493308, 1548737239915], "review_tcdate": [1548872514171, 1548525493308, 1548737239915], "review_tmdate": [1548872514171, 1548856735310, 1548856686544], "review_readers": [["everyone"], ["everyone"], ["everyone"]], "review_writers": [["MIDL.io/2019/Conference/Paper65/AnonReviewer2"], ["MIDL.io/2019/Conference/Paper65/AnonReviewer1"], ["MIDL.io/2019/Conference/Paper65/AnonReviewer3"]], "review_reply_count": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "review_replyto": ["ByxLSoblgV", "ByxLSoblgV", "ByxLSoblgV"], "review_content": [{"pros": "- This paper proposes the first end-to-end solution to the problem of semantic segmentation of teeth from intra-oral 3D scans. This is an interesting application of deep learning to 3D point clouds, as opposed to the more traditionally encountered image-based segmentation.\n\n- The paper is clearly written. In particular I found the work to be strongly motivated in the introduction by a clear presentation of the specific properties and challenges of the addressed application.\n\n- The proposed method and evaluation seem sound.\n\n- The two contributions introduced in the paper are well validated, both against external baselines and individually through an ablation study. The evaluation confirms the positive impact of each contribution on the segmentation performance.", "cons": "- One of the two contributions, namely the addition of a discriminative network during training to structure the prediction, is insufficiently discussed in comparison to the previous works. The idea of using an adversarial training which differentiates realistic from unrealistic label configurations was already introduced in several works, starting (I believe) with Luc et al [a]. Multiple additional references for medical applications are for example available in the introduction and related work sections of the paper (Ghafoorian et al, 2018). However, although (Ghafoorian et al, 2018) is mentioned in the Methods section, none of these works are mentioned in the introduction or the related work, which I found to be misleading regarding the novelty of this contribution.\n\n- The adversarial network is based on simple features extracted from the predicted labels (mean and variance of 3D voxel positions for each class). While this is technically indeed a novelty, I find this aspect to go a bit against the end-to-end claim, since it amounts to handcrafting features in the label space before applying a multilayer perceptron. This contrasts for example with [a] where a network is trained end-to-end to learn how a realistic label prediction should look like, possibly discovering high-level criteria related for example to object shapes, etc. Even if using these features might be perfectly sound for this application, I find that this is not motivated enough in comparison to the existing end-to-end approaches (which also goes back to the previous point), and rather in contradiction with the narrative of the paper.  The limitations of handcrafted features are indeed regularly pointed out in \u201cRelated work\u201d.\n\n- The lack of universal coordinate system is mentioned as a challenge for this clinical application, for example in \u201cDiscussion and conclusion\u201d. However, the features in the label space used to discriminate realistic from unrealistic labels include the means of 3D coordinates. Does not this possibly break the invariance with respect to the choice of coordinate system? I wonder whether using pairwise distances between classes, i.e. pairwise differences of means instead of the means directly, would be more suitable to guarantee an invariance to the choice of coordinates directly in the feature representation.\n\n\n[a]  Luc et al, Semantic Segmentation using Adversarial Network, NIPS Workshop 2016\n\n\nMinor comments:\n\n- If the extracted statistical features are sufficient, it might be worth training a simpler and more interpretable adversarial classifier than a multilayer perceptron \u2013 maybe simply a logistic regression? This would give a better understanding on how the notion of realistic segmentation output is encoded.\n\n- I wonder if the discriminator could not be used at prediction time as well, in addition to structuring the training. Could it for example provide a confidence measure on the output?\n\n- In general, I believe it is better to avoid using an ArXiV reference when the work was published and peer-reviewed, as is for example the case with the PointCNN paper (Li et al 2018).", "rating": "3: accept", "confidence": "2: The reviewer is fairly confident that the evaluation is correct"}, {"pros": "Summary\n\nThis paper proposes a new method on segmenting point cloud of intra-oral scans (IOS). This method contains three components: 1. Applying convolution neural networks (CNNs) on teeth semantic segmentation; 2. Proposing a non-uniform resampling strategy for better spatial learning; 3. Training loss combines with the auxiliary adversarial loss.\n\nThe proposed method achieves very good performance.", "cons": "1. The idea of this paper is not novel, segmentation network is based on PointCNN and dDiscriminator network is identical to the part of Point-Net. However, the author utilizes CNNs on teeth semantic segmentation, which is important.\n\n2. The comparison with different method and settings are straightforward.  \n\n3. Applying non-uniform resampling strategy will generate different segmentation mask according to the chosen fovea point. How to get the final completed segmentation result should be specified.\n\nMinor:\n-----------\n1. Figure. 1 shows the model framework, but this figure is hard to follow, please explain it properly in the paper.\n\n2. In equation (6), the author applies an adaptive weight between segmentation loss and adversarial loss. However, the network is not shared between segmentation and discriminative network. The weight may not be necessary here. No ablation experiment is done to identify how that affects performance.", "rating": "3: accept", "confidence": "2: The reviewer is fairly confident that the evaluation is correct"}, {"pros": "The paper applies point set classification on intra-oral scan of teeth. The proposed method includes a non-uniform resampling mechanism, and also a point-wise classification loss with an adversarial loss. \n\nThe paper is basically well written. The challenges are discussed, and the contributions are summarized. The whole paper is easy to follow. \n\nThe proposed method out-performed three state-of-the-art algorithms. ", "cons": "Minor comments: the two losses are not shown very well in Figure 1", "rating": "3: accept", "confidence": "2: The reviewer is fairly confident that the evaluation is correct"}], "comment_id": ["HJe3m4DsNN", "rJlNQXDi4N", "S1x-OVDsN4", "HketcBPiVV"], "comment_cdate": [1549657123870, 1549656859883, 1549657193048, 1549657489043], "comment_tcdate": [1549657123870, 1549656859883, 1549657193048, 1549657489043], "comment_tmdate": [1555946002336, 1555946002059, 1555946001840, 1555946001627], "comment_readers": [["everyone"], ["everyone"], ["everyone"], ["everyone"]], "comment_writers": [["MIDL.io/2019/Conference/Paper65/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper65/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper65/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper65/Authors", "MIDL.io/2019/Conference"]], "comment_reply_content": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "comment_content": [{"title": "Response to review comments", "comment": "We thank the reviewer for the detailed comments and appreciation shown. We would like to answer the reviewer\u2019s comments as follows:\n\n- We are agree with the reviewer that using an adversarial setting in a supervised segmentation approach is not explained in detail. For sake of space in the submitted version, we only referred the reader to such an approach in a recent publication (which is the work of Ghafoorian et al.). Taking this comment in consideration, in the final version of the paper, we will expand upon the background of this approach by including the prior works in image analysis which motivated us to employ such a framework on point cloud analysis.\n\n- We are agree with the reviewer that training the discriminator directly on the predicted labels and ground truth has advantages in comparison with training on the abstracted data (i.e. computed statistics). However, as we mentioned in the paper, this requires employing a deep network as discriminator (which might be as complex as the segmentation network). Because of the practical implications, such as instability of convergence and hardware limitations, using a deep network was considered to be prohibitive in our method. We found that employing simple statistics (mean and variance of points belonging to each label class) is sufficient for constraining the segmentation network to generate a more realistic prediction. In total, computing the low-order statistics (1st and 2nd moments) might be considered generic enough in the context of point cloud analysis as they are widely employed even as computational module in a CNN (e.g. average pooling or batch normalization operators). Hence, training the discriminator with low-order statistics of predicted labels might still be considered to fall within an end-to-end learning scheme.\n\n- The training data are augmented by applying random rotations. Using 3D coordinates of points as input to the discriminator was handled by employing a transformation network (called T-Net) inside the discriminator. The T-Net performs a learnable affine transformation on the input coordinates. Indeed the reviewer's suggestion regarding the use of a pair-wise distance seems interesting and worth trying but it should also be noted that although such a pair-wise encoding benefits from relational information instead of an absolute value of various coordinates, missing teeth are expected to require dedicated handling in a pair-wise encoding.\n\n- In our proposed method, the discriminator consist of a 3 layers MLP network excluding the T-Net. Such an architecture can be considered a shallow (i.e. simple) classifier. The suggestion of replacing the discriminator with a simpler classifier (e.g. a logistic regression) would be applicable in the case of employing rotation-invariant features which would aid in removal of the T-Net in the proposed architecture. Such an approach would require the design of other features which would break an end-to-end learning scheme as the reviewer mentioned in earlier comment.\n\n- The network probabilities, either generated by the segmentation or the discriminator can be used as an indication of uncertainty but such probability values need to be calibrated (Guo et al. 2017). Overall, the estimated uncertainty might not be as reliable as one would expect from a Bayesian approach.\n\n- In the final version of the paper, we will exchange the cited ArXiv papers for their published peer-reviewed references."}, {"title": "General Reply to Reviews", "comment": "We would like to thank all of the reviewers for their positive comments about the paper and their constructive suggestions."}, {"title": "Response to review comments", "comment": "We thank the reviewer for appreciation. In the final version of the paper, We will modify Figure (1) to show the losses as would be more representative of the proposed framework."}, {"title": "Response to review comments", "comment": "We thank the reviewer for the comments. We would like to address the comments below:\n\n- We agree with the reviewer that the architecture of the employed networks for IOS segmentation is not novel. One of the main contributions, as stated in the paper, is introducing a training framework for a deep model on the original resolution of a point cloud by introducing a non-uniform re-sampling mechanism based on Monte Carlo sampling and using a compatible loss function.\n\n- Since this is the first study into building end-to-end learning for IOS segmentation, we compare our proposed model only with three other state-of-the-art models for point cloud analysis. We also performed an ablation study for employing an adversarial setting and the non-uniform re-sampling.\n\n- In page 8, subsection \"Inference on the whole point cloud\", was explained that the complete segmentation is obtained by extracting a number of subsets (i.e. non-uniform resampled point sets) out of a given point cloud and aggregating the predictions on all subsets. Although in the submitted version of the paper, such a procedure has been demonstrated in detail by pseudocode in Algorithm 2 of the appendix, by considering this comment, in the final version, we will clarify in text how these subsets are extracted and how the prediction results are aggregated. In short, prediction on whole point cloud is started by randomly positioning of the fovea on a point and extracting the first subset of points. Applying the network on the given subset, the predicted probability vectors are assigned to those points. Thereafter, the fovea is randomly positioned on the points which are not within the previously processed subset(s). Here, the probability vector of a repeated point (which is visited more than once) is added to its previous values. In the end, by using an argmax operator, the most likely label is assigned to each point. In practice, after a few number of iterations (< 10), all points in the point cloud have been drawn at least once and the complete prediction is reported.\n \n- We will modify Figure (1) to be more representative of our proposed approach and will expand upon the explanation of what it shows after receiving the comments from the reviewers.\n\n- In the proposed framework, the segmentation network adapts its parameters to minimize two losses simultaneously, similar to what is defined in a multi-task learning framework. Since the output of the segmentation network (after computing mean and variance of each class prediction) is passed as input to the discriminator, one might consider these two networks as cascaded networks and resembling the concept of sharing weights. In previous works, different values for such a contribution weight have been proposed. For example, in work of Luc et al. (2016), a value equal to 2 (lambda=2) was suggested by employing a grid search and evaluating the performance on a validation set and in the work of Ghafoorian et al. (2018), it was adjusted to unity (lambda=1). We included the ablation test for employing the discriminator as well as for the proposed non-uniform re-sampling as reported in table 1. However, for using an adaptive weight between two losses, we have not performed such an ablation test as it was out of the scope of this paper and not considered as our contribution."}], "comment_replyto": ["HkgcroPJNE", "ByxLSoblgV", "HklgkoI6X4", "r1eT2kXq7E"], "comment_url": ["https://openreview.net/forum?id=ByxLSoblgV&noteId=HJe3m4DsNN", "https://openreview.net/forum?id=ByxLSoblgV&noteId=rJlNQXDi4N", "https://openreview.net/forum?id=ByxLSoblgV&noteId=S1x-OVDsN4", "https://openreview.net/forum?id=ByxLSoblgV&noteId=HketcBPiVV"], "meta_review_cdate": 1551356588318, "meta_review_tcdate": 1551356588318, "meta_review_tmdate": 1551881980838, "meta_review_ddate ": null, "meta_review_title": "Acceptance Decision", "meta_review_metareview": "Reviewers have recognized a well-motivated paper facing an important and interesting application. There were some concerns raised about the novelty of the contribution, which were answered to some extent by the authors, putting forward a novel non-uniform resampling mechanism. The final version of the paper should clarify both the novelty with respect to prior work in adversarial segmentation and the positioning of the paper regarding the end-to-end vs handcrafted feature approaches. \n\n", "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=ByxLSoblgV&noteId=rklE2G8rIV"], "decision": "Accept"}