Datasets:

mfromm
/

AMSR

{"forum": "SJgNCUbke4", "submission_url": "https://openreview.net/forum?id=SJgNCUbke4", "submission_content": {"title": "Deep Hierarchical Multi-label Classification of Chest X-ray Images", "authors": ["Haomin Chen", "Shun Miao", "Daguang Xu", "Gregory D. Hager", "Adam P. Harrison"], "authorids": ["hchen135@jhu.edu", "shwinmiao@gmail.com", "daguangx@nvidia.com", "hager@cs.jhu.edu", "adam.p.harrison@gmail.com"], "keywords": ["hierarchical multi-label classification", "chest x-ray", "computer aided diagnosis."], "TL;DR": "A hierarchical multi-label classification approach for CXR CAD that leverages and respects clinical taxonomies", "abstract": "Chest X-rays (CXRs) are a crucial and extraordinarily common diagnostic tool, leading to heavy research for computer-aided diagnosis (CAD) solutions. However, both high classification accuracy and meaningful model predictions that respect and incorporate clinical taxonomies are crucial for CAD usability. To this end, we present a deep hierarchical multi-label classification (HMLC) approach for CXR CAD. Different than other hierarchical systems, we show that first training the network to model conditional probability directly and then refining it with unconditional probabilities is key in boosting performance. In addition, we also formulate a numerically stable cross-entropy loss function for unconditional probabilities that provides concrete performance improvements. To the best of our knowledge, we are the first to apply HMLC to medical imaging CAD. We extensively evaluate our approach on detecting 14 abnormality labels from the PLCO dataset, which comprises 198, 000 manually annotated CXRs. We report a mean area under the curve (AUC) of 0.887, the highest yet reported for this dataset. These performance improvements, combined with the inherent usefulness of taxonomic predictions, indicate that our approach represents a useful step forward for CXR CAD.", "pdf": "/pdf/31b446cd17f2c4c4ec699e9dba494757ac9fce45.pdf", "code of conduct": "I have read and accept the code of conduct.", "remove if rejected": "(optional) Remove submission if paper is rejected.", "paperhash": "chen|deep_hierarchical_multilabel_classification_of_chest_xray_images", "_bibtex": "@inproceedings{chen:MIDLFull2019a,\ntitle={Deep Hierarchical Multi-label Classification of Chest X-ray Images},\nauthor={Chen, Haomin and Miao, Shun and Xu, Daguang and Hager, Gregory D. and Harrison, Adam P.},\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=SJgNCUbke4},\nabstract={Chest X-rays (CXRs) are a crucial and extraordinarily common diagnostic tool, leading to heavy research for computer-aided diagnosis (CAD) solutions. However, both high classification accuracy and meaningful model predictions that respect and incorporate clinical taxonomies are crucial for CAD usability. To this end, we present a deep hierarchical multi-label classification (HMLC) approach for CXR CAD. Different than other hierarchical systems, we show that first training the network to model conditional probability directly and then refining it with unconditional probabilities is key in boosting performance. In addition, we also formulate a numerically stable cross-entropy loss function for unconditional probabilities that provides concrete performance improvements. To the best of our knowledge, we are the first to apply HMLC to medical imaging CAD. We extensively evaluate our approach on detecting 14 abnormality labels from the PLCO dataset, which comprises 198, 000 manually annotated CXRs. We report a mean area under the curve (AUC) of 0.887, the highest yet reported for this dataset. These performance improvements, combined with the inherent usefulness of taxonomic predictions, indicate that our approach represents a useful step forward for CXR CAD.},\n}"}, "submission_cdate": 1544652492430, "submission_tcdate": 1544652492430, "submission_tmdate": 1561399817327, "submission_ddate": null, "review_id": ["ryeLVHzPXN", "HJxGGGe1X4", "SyeiWe327E"], "review_url": ["https://openreview.net/forum?id=SJgNCUbke4&noteId=ryeLVHzPXN", "https://openreview.net/forum?id=SJgNCUbke4&noteId=HJxGGGe1X4", "https://openreview.net/forum?id=SJgNCUbke4&noteId=SyeiWe327E"], "review_cdate": [1548326190303, 1547792905700, 1548693506630], "review_tcdate": [1548326190303, 1547792905700, 1548693506630], "review_tmdate": [1548856723621, 1548856710296, 1548856700122], "review_readers": [["everyone"], ["everyone"], ["everyone"]], "review_writers": [["MIDL.io/2019/Conference/Paper29/AnonReviewer3"], ["MIDL.io/2019/Conference/Paper29/AnonReviewer2"], ["MIDL.io/2019/Conference/Paper29/AnonReviewer1"]], "review_reply_count": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "review_replyto": ["SJgNCUbke4", "SJgNCUbke4", "SJgNCUbke4"], "review_content": [{"pros": "The paper investigates the importance of explicitly using taxonomical structures of labels for classification tasks in medical imaging within the application of chest-x-ray classification into a number of hierarchically related diagnostic categories.\n\nThe key contribution is the formulation of hierarchical multi-label classification within a deep-learning framework, which is a novel and generally useful idea. They authors propose an intuitive and effective two-stage optimisation scheme which first encourages the model to capture label taxonomy and then maximise the end classification accuracy. They introduce a numerically stable implementation of cross entropy loss for unconditional class probabilities. Using a large labelled chest x-ray data based, they provide a first demonstration of hierarchical multi-label classification in medical imaging.\n\nThe paper is well written and well motivated. Results show improvement in classification accuracy over relevant baselines.\n", "cons": "Despite the new loss functions that accounts for the label hierarchy, a single\n\ndistributed model is used to predict all the classes. This means that the architecture does not respect the hierarchical structure in the data e.g. different features may be desirable for detecting \u2018Abnormality\u2019, and discriminating between \u2018Pleural fibrosis\u2019 and \u2018Fluid in pleural space\u2019.\n\nThe improvements shown in Table 2 are relatively small and no error bars are provided.\n\n\n\nIt would be interesting to see the benefits of capturing label taxonomy when the size of the training data is smaller. Injecting such domain prior knowledge may improve the data efficiency.\n\nIt would be also interesting to see how the model performs in the presence of incomplete labels i.e. each image is not necessarily labeled until it reaches one of the leaf nodes (e.g. Abnormality=> Pulmonary => Opacity, but Infiltration or Major atelectasis is not known) .\n\nIt would be more informative to reorder to the disease indices in the breadth-first order. This would help to see how the level within the taxonomy affects the performance. I also wonder what the plot would look like after the first optimisation phase based on the HLCP loss.\n\nOverall, the contribution of the paper is solid in terms of technical novelty and problem formulation. However, the paper could use stronger experiments as suggested earlier to bolster its claims. \n", "rating": "4: strong accept", "confidence": "2: The reviewer is fairly confident that the evaluation is correct", "special_issue": ["Special Issue Recommendation"], "oral_presentation": ["Consider for oral presentation"]}, {"pros": "1. to present a deep hierarchical multi-label classi\fcation (HMLC) approach for CXR CAD. \n2. to model conditional probability directly and with unconditional probabilities is key in boosting performance.\n3. formulate a numerically stable cross-entropy loss function for unconditional probabilities \n4. evaluate our approach on detecting 14 abnormality labels from the PLCO dataset, which comprises 198; 000 manually annotated CXRs. We report a mean area under the curve (AUC) of 0:887, the highest yet reported for this dataset. \n", "cons": "1. There is no cross-validation, external validation, with a confidence interval for evaluating significant better method.\n", "rating": "3: accept", "confidence": "2: The reviewer is fairly confident that the evaluation is correct"}, {"pros": "The paper is well written and describes an interesting and relatively novel approach to solving multi-class classification in a clinical domain where overlap between classes is frequently a possibility.   The approach is clearly explained and the results presented are sufficient to give merit to the idea. ", "cons": "The authors could spend a little more effort on explaining the intuition behind conditional versus unconditional labels and the advantages of each.\nOnly a single (large) dataset is used, while there are many publicly available datasets that could be included for additional experiments. \nNo public implementation of the method is provided, which would be a nice extra", "rating": "3: accept", "confidence": "3: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature", "oral_presentation": ["Consider for oral presentation"]}], "comment_id": ["rJlNf0-CNE", "HJgXaAb0EV", "SkeJZJG0EV"], "comment_cdate": [1549831691710, 1549831867311, 1549831927105], "comment_tcdate": [1549831691710, 1549831867311, 1549831927105], "comment_tmdate": [1555945978627, 1555945978151, 1555945977935], "comment_readers": [["everyone"], ["everyone"], ["everyone"]], "comment_writers": [["MIDL.io/2019/Conference/Paper29/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper29/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper29/Authors", "MIDL.io/2019/Conference"]], "comment_reply_content": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "comment_content": [{"title": "Comment for the first reviewer", "comment": "We very much appreciate your constructive comments and feedback. \n\nConditional vs. Unconditional labels:  Conditional labels and unconditional labels have different advantages and we use the two step method to combine them together. Conditional labels allow the final linear layer to perform predictions without being confounded by diseases under different parents. They are also less unbalanced during HLCP training. However, unconditional labels have more negative samples to train the network. We will add these details to our revision. \n\nDataset: We agree that more datasets would strengthen the work. We chose the PLCO because at the time it was the only large-scale multilabel dataset for CAD with reliable ground truths that we were aware of. We are excited to see the release of other multi-label CXR datasets, e.g., CheXpert, which was released after we submitted. Our revision will now mention that evaluating against CheXpert, and also against other computer vision datasets, is an important aspect of future work to further characterize performance.\n\nPublic implementation: we also see the value in a public implementation and we are doing our best to gain approval for it. In the meantime, we do hope our exposition is clear and simple enough for others to reproduce. \n"}, {"title": "Comment for the third reviewer", "comment": "We very much appreciate your comments and feedback. \n\nIncomplete labels: We agree that a powerful strength of hierarchical classification is the possibility in handling incomplete data in an elegant fashion, but we felt that it was outside the scope of this work to properly explore. Given its importance, we now mention this as an aspect of future work, which we agree deserves research focus.\n\nHierarchical features: We agree that different level of diseases may certainly benefit from hierarchical features. We actually performed some preliminary experiments along this line, but found no improvement over the shared features approach. This anecdotal experience is hardly definitive, as there is actually no straightforward way to do this and there are many avenues to try. As a result, this question is a complex one that deserves much more thorough investigation than our preliminary efforts, and we now emphasize this in our future work. Even so, we would like to point out that our work demonstrates that hierarchical classification can still produce improvements when even deployed in a limited capacity with shared features, which lends evidence to the power of the approach. \n\nUsing a small dataset: Thank you for this suggestion, it would indeed be interesting to investigate this.\n\nAUC plot: Thank you for the suggestion to graph in breadth-first search. Our revision will display the graph in this way, which we agree better reveals the effects of taxonomy levels. \n\nFor reasons of clarity, we did not graph the AUCs after the first phase of optimization (HLCP). However, the HLCP results all sit in between the baseline method and the HLUP finetune method, except for two disease patterns that actually sat slightly below the baseline method. All HLCP results have lower AUCs than the HLUP finetune method. Our revision will now mention this in the body text. \n"}, {"title": "Comment for the second reviewer", "comment": "Thank you for these suggestions. Given the very large size of the PLCO dataset, which reached almost 200K CXRs, time and resources did not allow us to cross validate. Nonetheless, we did observe that validation results were very stable from run to run, which at least demonstrated that improvements were not due to the randomness of model initialization and optimization. As well, we believe that the size of the PLCO dataset, which nears that of many computer vision datasets that aren\u2019t commonly cross-validated, helps assuage some of the concerns due to lack of cross validation. Having said all this, we certainly agree that cross validation would strengthen evaluation further. \n\nWe also agree that significance tests would strengthen the evaluation. To this end, our revision will now include the results of significance tests of the AUC values in Figure 3, using the non-parametric method of DeLong et al. 10 disease patterns are seen to show significant improvements (p<0.05), which we will now highlight in Figure 3. Of interest, is that statistical significance also respects our hierarchy (if a child is significant, so is its parent).\n\nThank you for your constructive comments and feedback.  \n"}], "comment_replyto": ["SyeiWe327E", "ryeLVHzPXN", "HJxGGGe1X4"], "comment_url": ["https://openreview.net/forum?id=SJgNCUbke4&noteId=rJlNf0-CNE", "https://openreview.net/forum?id=SJgNCUbke4&noteId=HJgXaAb0EV", "https://openreview.net/forum?id=SJgNCUbke4&noteId=SkeJZJG0EV"], "meta_review_cdate": 1551356580075, "meta_review_tcdate": 1551356580075, "meta_review_tmdate": 1551881975467, "meta_review_ddate ": null, "meta_review_title": "Acceptance Decision", "meta_review_metareview": "The authors present a deep hierarchical multi-label classification approach to CAD on chest radiographs, and test it on a large public dataset of nearly 200,000 images. The manuscript is well written and organized. The presented approach addresses the situation where the different classes can overlap. In addition, it incorporates both conditional and unconditional probabilities, and demonstrates improved performance when compared to a conventionally accepted alternative approach.\n\nIn response to the reviewers\u2019 requests for clarification, the authors have provided additional information to explain their rationale for certain choices in the experimental design. There is an appropriate level of detail here to better understand their work.\n\nThe reviewers rightly point out that some additional experiments would be important for the authors to truly show that their method is an improvement over existing approaches, and the authors have responded appropriately, and where relevant, have described changes that they will make in their revision to address these comments.", "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=SJgNCUbke4&noteId=r1g2iGUrUE"], "decision": "Accept"}