Datasets:

mfromm
/

AMSR

{"forum": "BJlXpDh1gV", "submission_url": "https://openreview.net/forum?id=BJlXpDh1gV", "submission_content": {"title": "3D multirater RCNN for multimodal multi class segmentation of extremely small objects (ESO)", "authors": ["Carole H. Sudre", "Beatriz Gomez Anson", "Silvia Ingala", "Chris D. Lane", "Daniel Jimenez", "Lukas Haider", "Thomas Varsavsky", "Lorna Smith", "S\u00e9bastien Ourselin", "Rolf H. J\u00e4ger", "M. Jorge Cardoso"], "authorids": ["carole.sudre@kcl.ac.uk", "bgomeza@santpau.cat", "s.ingala@vumc.nl", "c.lane@ucl.ac.uk", "d.jimenez@ucl.ac.uk", "l.haider@ucl.ac.uk", "thomas.varsavsky@kcl.ac.uk", "lorna.smith@ucl.ac.uk", "sebastien.ourselin@kcl.ac.uk", "r.jager@ucl.ac.uk", "m.jorge.cardoso@kcl.ac.uk"], "keywords": [], "abstract": "Extremely small objects (ESO) have become observable on clinical routine magnetic resonance imaging acquisitions, thanks to a reduction in acquisition time at higher resolution. Despite their small size (usually $<$10 voxels per object for an image of more than $10^6$ voxels), these markers reflect tissue damage and need to be accounted for to investigate the complete phenotype of complex pathological pathways. In addition to their very small size, variability in shape and appearance leads to high labelling variability across human raters, resulting in a very noisy gold standard. \nSuch objects are notably present in the context of cerebral small vessel disease where enlarged perivascular spaces and lacunes, commonly observed in the ageing population, are thought to be associated with acceleration of cognitive decline and risk of dementia onset. In this work, we redesign the RCNN model to scale to 3D data, and to jointly detect and characterise these important markers of age-related neurovascular changes. We also propose training strategies enforcing the detection of extremely small objects, ensuring a tractable and stable training process.", "pdf": "/pdf/4bef70f9dbbe597accb23387cafc52eaad05edb6.pdf", "code of conduct": "I have read and accept the code of conduct.", "remove if rejected": "(optional) Remove submission if paper is rejected.", "paperhash": "sudre|3d_multirater_rcnn_for_multimodal_multi_class_segmentation_of_extremely_small_objects_eso", "_bibtex": "@inproceedings{sudre:MIDLFull2019a,\ntitle={3D multirater {\\{}RCNN{\\}} for multimodal multi class segmentation of extremely small objects ({\\{}ESO{\\}})},\nauthor={Sudre, Carole H. and Anson, Beatriz Gomez and Ingala, Silvia and Lane, Chris D. and Jimenez, Daniel and Haider, Lukas and Varsavsky, Thomas and Smith, Lorna and Ourselin, S{\\'e}bastien and J{\\\"a}ger, Rolf H. and Cardoso, M. Jorge},\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=BJlXpDh1gV},\nabstract={Extremely small objects (ESO) have become observable on clinical routine magnetic resonance imaging acquisitions, thanks to a reduction in acquisition time at higher resolution. Despite their small size (usually {\\$}{\\ensuremath{<}}{\\$}10 voxels per object for an image of more than {\\$}10^6{\\$} voxels), these markers reflect tissue damage and need to be accounted for to investigate the complete phenotype of complex pathological pathways. In addition to their very small size, variability in shape and appearance leads to high labelling variability across human raters, resulting in a very noisy gold standard. \nSuch objects are notably present in the context of cerebral small vessel disease where enlarged perivascular spaces and lacunes, commonly observed in the ageing population, are thought to be associated with acceleration of cognitive decline and risk of dementia onset. In this work, we redesign the RCNN model to scale to 3D data, and to jointly detect and characterise these important markers of age-related neurovascular changes. We also propose training strategies enforcing the detection of extremely small objects, ensuring a tractable and stable training process.},\n}"}, "submission_cdate": 1544697786655, "submission_tcdate": 1544697786655, "submission_tmdate": 1561397457599, "submission_ddate": null, "review_id": ["BJeVv4V3QV", "BkeYBWFcz4", "ByewdOKAQE"], "review_url": ["https://openreview.net/forum?id=BJlXpDh1gV&noteId=BJeVv4V3QV", "https://openreview.net/forum?id=BJlXpDh1gV&noteId=BkeYBWFcz4", "https://openreview.net/forum?id=BJlXpDh1gV&noteId=ByewdOKAQE"], "review_cdate": [1548661852309, 1547501888856, 1548814446962], "review_tcdate": [1548661852309, 1547501888856, 1548814446962], "review_tmdate": [1548856750258, 1548856706039, 1548856680398], "review_readers": [["everyone"], ["everyone"], ["everyone"]], "review_writers": [["MIDL.io/2019/Conference/Paper44/AnonReviewer1"], ["MIDL.io/2019/Conference/Paper44/AnonReviewer2"], ["MIDL.io/2019/Conference/Paper44/AnonReviewer3"]], "review_reply_count": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "review_replyto": ["BJlXpDh1gV", "BJlXpDh1gV", "BJlXpDh1gV"], "review_content": [{"pros": "-\tThe authors address the highly complex but highly relevant problem of detection and classification of perivascular spaces and lacunes. Although the problem is very difficult, due to rater uncertainty and high class imbalance, the authors achieve reasonable sensitivity.\n-\tThe authors present interesting improvements, both in network architecture and in training approach, that could be beneficial to other applications. Especially the use of the multiple raters is original. In addition, the sampling based on distance maps is interesting.\n-\tFor interpretation of the results, some baseline models are missing. I appreciate that the authors made an effort to train such models. Unfortunately, these models gave no results.\n", "cons": "-\tThe problem is complex, but the authors made it even more complex than needed by combining two tasks (EPVS and lacunes) and by incorporating multirater information. It would be useful to see the model\u2019s (and reference models\u2019) performance on the separate problems of EPVS and lacunes.\n- The authors did not assess the added value of the multirater encoding. It will be insightful to include for example a baseline method trained on a single rater. \n- The authors did not assess the added value of their dedicated sampling strategy.\n-\tThe test set is very small, consisting of only 2 subjects. The validation is therefore quite limited.\n-\tFigure 7 is hard to interpret. What is the meaning of all the blue (=uncertain) \u2018Nothing\u2019 boxes? Why displaying the blue boxes if they should be disregarded?\n\n\nAlthough validation can be improved, this work is highly relevant and gives many interesting leads for discussion at the MIDL conference.\n", "rating": "4: strong 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"]}, {"pros": "This paper presented a redesigned the RCNN model to detect and classify extremely small objects in MRI. This paper also presented an experiment results to show the good sensitivity of this methodology. ", "cons": "General comments:\n* Grammar/sentence revision and proper introduction in the background can improve the readability of this paper significantly\n* I would also suggest to position in the Figure at the beginning of the explanation rather than at the end. The picture can better guide reader understand the specific content. Such as Figure 1 and Figure 2. \n* I think the paper is too specific and maybe hard to generalize to other datasets or problem \n * The data is specific, I don\u2019t see any discussion around the resolution of the image, noise level of the image \n * The parameters are specific: to me, there are quite a few adhoc hyperparameters \n * The dataset is small. The data description and the data cleaning process is not clear to me.\n\nSpecific comments: \n* Motivation: \n * I recommend the author add more information on the significance of doing ESO detection \n * Specific examples/articles to support 'these markers reflect tissue damage and need to be accounted for to investigate the complete phenotype of complex pathological pathways' \n* Please elaborate the purpose of the Fig 1 along with the meaning of the red dots.\n* Please explain the reason why the HighResNet was selected Backbone network. \n* In the DL equation, \n * what is the meaning of the r_n? Is that a single number distance or a vector of 3 \n * what is the unit for the distance r_n?Does that cutoff change as the resolution changes? \n * Is the scale factor the solution to deal with resolution changes? \n* 'All input data was bias field corrected, skull stripped, and then z-scored to the white matter region statistics'.\n * Was the white matter region known at the point or some estimation was applied\n* On page 5, don\u2019t understand 'the skeleton maxima of the smoothed regressed distance map (p score map >0.25)\u2019 \n* I am confused with sec 3.1. I am assuming 'Out of the initial 4147 considered elements, 2442 were used as gold standards for training. \u2018 means for all 16 subjects, 2442 elements was used for training and testing. Among 2442, the elements belongs to 14 subjects were used for training, and the element belongs to 2 subjects were used for hold-out testing. Please confirm or clarity.\n* Not clear to me what is the input of this algorithm? \n* I don\u2019t understand Figure 7, where is the ground truth? The first row boxes in Lacune and undecided look exactly same to me ", "rating": "2: reject", "confidence": "2: The reviewer is fairly confident that the evaluation is correct"}, {"pros": "This paper addresses a class of difficult problems MRI neuro image analysis, which is the detection and localization of small anomalies (such as parivascular spaces and lacunes) in MRI. The features are important in studies/analysis of, e.g., cognitive decline with aging (e.g. \"vascular dementia\"). The paper is moderately well written (with a few grammar problems), clear. The evaluation is only moderately strong. ", "cons": "This is an application of a well-known neural-net architecture (RCNN). The authors extended the NN to 3D (but little is discussed on that topic) and there were some challenges in messaging the training data (e.g. dealing with asymmetries in numbers of examples), multirater ground truth, and a few other details. Overall this is a straightforward application with a moderate evaluation, suffering from a small sample size, and a lack of evaluation of different design choices. ", "rating": "3: accept", "confidence": "2: The reviewer is fairly confident that the evaluation is correct"}], "comment_id": ["Skg03qKu4E", "HJlariYu4N", "BJxpXhtuNE"], "comment_cdate": [1549470390314, 1549470533369, 1549470757411], "comment_tcdate": [1549470390314, 1549470533369, 1549470757411], "comment_tmdate": [1555946023344, 1555946023085, 1555946022873], "comment_readers": [["everyone"], ["everyone"], ["everyone"]], "comment_writers": [["MIDL.io/2019/Conference/Paper44/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper44/Authors", "MIDL.io/2019/Conference"], ["MIDL.io/2019/Conference/Paper44/Authors", "MIDL.io/2019/Conference"]], "comment_reply_content": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "comment_content": [{"title": "Answer to Reviewer 2", "comment": "We would like to thank the reviewer for the care with which the review has been performed and the detailed associated comments. \nWe appreciate the suggestion of the reviewer to introduce the problem in more details and will ensure the motivation of this work is clarified. As per references to the topic, we would like to add references to the STRIVE paper describing the hallmarks of cerebral small vessel disease by Wardlaw et al. (Wardlaw JM et al. Neuroimaging standards for research into small vessel disease and its contribution to ageing and neurodegeneration. Lancet Neurol. 2013; 12: 822\u2013838.) and to the review by Ramirez et al. (Ramirez J, et al Imaging the Perivascular Space as a Potential Biomarker of Neurovascular and Neurodegenerative Diseases.Cell. Mol. Neurobiol. 2016) that describes notably the differences between EPVS and lacunes and the difficulty of the task. \nWith respect to the figures we would like notably to adopt the suggestion to make a better use of the figures to describe the issue. Regarding the figure 1, the red dots correspond to the centre of mass of the segmented object each rater had to classify and we will add this explanation. This dot was particularly useful to indicate which element the raters had to classify when the snapshots contained more than one potential element.\nRegarding the note of the limited test dataset, we would like to refer the reviewer to our answer to reviewer 1 detailing the difficulty to obtain the gold standard segmentation. As of note, the 16 segmented subjects were chosen for their high pathology load with a number of segmented elements ranging from 135 to 550. \nAs per the reviewer request, a sentence will be added to better describe resolution and acquisition of the data. All three modalities (T1w, T2w, FLAIR) were obtained using a 3D isotropic 1x1x1 mm acquisition.\nStill at an early stage in the preparation of this prototype, the choice of HighResNet as backbone network was motivated by its ease to train thanks to the residual connections. Its versatility has empirically been shown at training stage in problems where data imbalance is strong (Sudre CH, et al Generalised Dice overlap as a deep learning loss function for highly unbalanced segmentations . 2017). The need to further optimise the architecture of the different pieces of the framework and notably the design choice of the backbone will be expressed as another area of future investigation.\nWith respect to the details of DL, the absolute error r_n for each centre of mass is evaluated in each direction and given the isotropic nature of the image resolution (in our case) is expressed in mm. The selected cut-off is evaluated in terms of absolute (mm) distance.\nWe also confirm the interpretation of the reviewer with respect to the used data. Out of the originally 4147 segmented elements, only 2442 satisfied the criterion of being larger or equal to 5mm3 and were thus subsequently used as ground truth as training data for those belonging to the 14 training subjects or test data for those of the test subjects. The sentence will be modified to avoid any misinterpretation.\nThe sentence regarding the selection of the centres of mass based on the distance map can be rephrased as follows. In order to prune the potential positions of centre of mass, we combine the information from the score map and the distance map. The score map is thresholded at 0.25 and we extract the morphological skeleton of the underlying distance map. The maxima are then taken as potential proposed centres of mass.\nRegarding the WM mask used as reference region for the intensity normalisation, the standardisation was performed using an external segmentation tool implemented for the detection of white matter hyper intensities (Sudre CH et al. Bayesian Model Selection for Pathological Neuroimaging Data Applied to White Matter Lesion Segmentation. IEEE Trans. Med. Imaging 2015; 34). The T1, FLAIR and T2 images intensity normalised with respect to the distribution over the white matter were used as input to the framework\nAlthough we agree the application is here quite specific, we believe that this framework can be largely generalised to other issues of detection of very small pathological elements and their subsequent classification.\nIn answer to the requirement for further clarification regarding Figure 7, as per our answer to reviewer 1, the ground truth reflects the probability for each element to be classified as a certain category by the different raters. The fact that, by chance, the boxes of lacune and undecided classification look very similar as noticed by the reviewer is explained by the variability in rater classification in these cases. For instance, the small element neighbouring the caudate, was classified by 1 rater as nothing, 1 rater as lacune, 1 rater as undecided and 3 raters as EPVS while for the larger one close to the cortex, it was classified by 2 rater as lacune, 3 raters as EPVS and 1 rater as undecided. "}, {"title": "Answer to Reviewer 1", "comment": "We thank the reviewer for the detailed comments and useful suggestions of improvement.\n\nAs pointed out by all three reviewers, the test set is relatively limited with only two subjects (out of 16 for which ground truth was available). It must be however underlined that the images used for ground truth were acquired in 3D and that the manual segmentation prior to the multi rater classification was performed on each slice in each orientation to ensure segmentation consistency. Performed by a rater accustomed to the use of the segmentation software, the delineation of EPVS and lacunes on a single image required on average 20h of work. The nature of the objects to segment made the task increasingly cumbersome. This aspect of the data preparation would be further highlighted in the amended version of the manuscript.\n\nAlthough it would have been natural to start this work in a less complicated setting aiming \u201conly\u201d at the detection of EPVS, the observation of the inherent rater variability in the classification and the difficulty in many cases to discriminate between lacunes and EPVS constrained us from the start to consider the problem as a whole. Limiting to EPVS only would have deteriorated the overall quality of the ground truth.\n\nWe appreciate the suggestion of the reviewer to evaluate the relevance of the sampling strategy as potential comparative baseline model. \n\nRegarding the multi-rater encoding, we are currently assessing the benefits of such modelling. \n\nWe understand the need for a clarification of Figure 7 generally underlined by the reviewers. The ground truth (first row) provides the probability for each element of being classified as one of the 4 categories according to the average of the 6 raters. Although the blue boxes do not give information about the final crisp classification of the underlying element, their presence show the probability with which another class would be given.\n"}, {"title": "Answer to Reviewer 3", "comment": "We thank the reviewer for the care with which the review was performed and the comments provided to improve the manuscript. \n\nWe appreciate the request of reviewer 3 to give more details the ways of adapting the 2D framework in a 3D setting. Here the main aspects that were modified are the solution chosen to avoid to systematically consider anchor points for the box proposals as well as the absence of pooling and interpolation in order to avoid losing information about the small elements we are attempting to detect. We would further underline these aspects in the revised manuscript.\n\nWith respect to the question of the size of test data, we would like to refer Reviewer 3 to our answers to Reviewer 1 and 2 highlighting notably the difficulty in obtaining suitable gold standard manual segmentation."}], "comment_replyto": ["BkeYBWFcz4", "BJeVv4V3QV", "ByewdOKAQE"], "comment_url": ["https://openreview.net/forum?id=BJlXpDh1gV&noteId=Skg03qKu4E", "https://openreview.net/forum?id=BJlXpDh1gV&noteId=HJlariYu4N", "https://openreview.net/forum?id=BJlXpDh1gV&noteId=BJxpXhtuNE"], "meta_review_cdate": 1551356592641, "meta_review_tcdate": 1551356592641, "meta_review_tmdate": 1551881977632, "meta_review_ddate ": null, "meta_review_title": "Acceptance Decision", "meta_review_metareview": "This paper presented a straight forward approach to segment small objects using RCNN network architecture. The novelty is moderate and need further improvement in terms of more strict evaluations. The authors have responded to most of the critiques. Paper is generally well written. Therefore, I would suggest to accept this paper in MIDL 2019 as a poster. ", "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=BJlXpDh1gV&noteId=B1gt3zIHUE"], "decision": "Accept"}