![\"\"](\"https://f.easyuploader.app/20210718170843_6f656f5a.png\")
Having a look at the data, I can see we have 672 subjects (585 train + 87 test) which presumably all feature a brain tumour. Our task is to classify these tumours for the presence or absence of a genetic biomarker - providing a non-invasive screening test to direct treatment for glioblastoma multiforme (GBM, which also goes by the name 'glioblastoma').
\nMy proposed pipeline is to perform per-voxel classification on the tumour, then produce an output based on a majority voting over the voxel-wise classes. This is similar to the approach described in this paper. Interested to hear what other approaches Kagglers are thinking about.
", "messages": [ { "id": 1387381, "postDate": "2021-07-14T06:26:50.643Z", "content": "Having a look at the data, I can see we have 672 subjects (585 train + 87 test) which presumably all feature a brain tumour. Our task is to classify these tumours for the presence or absence of a genetic biomarker - providing a non-invasive screening test to direct treatment for glioblastoma multiforme (GBM, which also goes by the name 'glioblastoma').
\nMy proposed pipeline is to perform per-voxel classification on the tumour, then produce an output based on a majority voting over the voxel-wise classes. This is similar to the approach described in this paper. Interested to hear what other approaches Kagglers are thinking about.
", "rawMarkdown": "Having a look at the data, I can see we have 672 subjects (585 train + 87 test) which presumably all feature a brain tumour. Our task is to classify these tumours for the presence or absence of a genetic biomarker - providing a non-invasive screening test to direct treatment for glioblastoma multiforme (GBM, which also goes by the name 'glioblastoma').\n\nMy proposed pipeline is to perform per-voxel classification on the tumour, then produce an output based on a majority voting over the voxel-wise classes. This is similar to the approach described in [this paper](http://www.ajnr.org/content/42/5/845). Interested to hear what other approaches Kagglers are thinking about.", "votes": 26 }, { "id": 1391565, "postDate": "2021-07-17T17:59:58.500Z", "content": "Right now we are playing around with an end2end approach, just feeding a 4-channel video to a 3D CNN and see what happens. So far we know that 3D convolution takes quite a long time, even after reducing the dynamic range of the images (and thus reducing the dataset size to ~8GB). We haven't even managed to train a net yet l0l.
\nAnother idea is to try a seq2seq approach, feeding a CNN the successive frames of a video, then sending the embedding to an LSTM and keeping the classification probability at the last time step. We'll see.
I think that's a good point 😊
\nThe method you mentioned at the end above comment will be almost the same as the second place solution in the previous RSNA brain MRI competition.
\n👉 https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/117228
However, I think the main difference from the previous RSNA competition is that the labels are not provided for each image, but for each patient, so it is not clear which cross-sectional image contributes to the 0/1 label. Therefore, I think that this task will be a kind of weakly supervised learning.
\nIn addition, as shown in the histogram below (x-axis: the number of images in each patient), the number of each of the four types of images is different for each patient, so we need to consider how to deal with this.
Thanks, @maxwell110 for your insights, SO how do you think we should approach this problem. Are 3D convolution models the only way?
", "rawMarkdown": "Thanks, @maxwell110 for your insights, SO how do you think we should approach this problem. Are 3D convolution models the only way?", "votes": 1 }, { "id": 1391954, "postDate": "2021-07-18T08:53:23.840Z", "content": "Yes, we only have a single annotation per video, not per frame, and the amount of frames per video and per patient is variable.
\nTo deal with the annotation thing, we can try at first labeling all the frames with the label of the video, so if a video is labeled 0, then annotate all its frames as 0. I don't know how weakly supervised learning could play out here, so please tell us a bit, it would be very helpful.
\nRegarding the variable length of the videos, we can try padding with zeros to a fixed length, or perhaps interpolating the z-axis to a certain length.
I'm still trying to figure out how to do this, so I can't say anything for sure.
\nBut, for example, there has been a similar competition on Kaggle in the past with audio data instead of images. In that competition, audio of arbitrary length (e.g., 5 - 120 seconds) was given a label of the birdsong present in the audio. However, similar to this competition, there was no information about which part of the audio contained the birdsong, so I remember that the top solutions of the competitors used a weakly supervised learning method called Sound Event Detection (SED).
\nFor an overview of SED, I think it is best to read the following kernel.
\n👉 https://www.kaggle.com/hidehisaarai1213/introduction-to-sound-event-detection
To put it simply, it uses attention blocks to calculate the loss between the given aggregated labels and the predicted labels for each speech segment by taking the sum of the predicted labels.
\nThis explanation may not be enough for you to understand, so please read the kernel if you are interested.
However, unlike audio data, this time we are dealing with image data, so the dimension of the data will increase by one. Therefore, it is unclear whether a method like SED is feasible in terms of computational resources.
\nIn any case, I think that one of the most creative and meaningful moments in Kaggle is when we are thinking about how to apply existing and useful methods to new data and problems. This competition has very little data, which makes me a little nervous about the LB shake, but I hope we can both enjoy it until the end.
\nHappy Kaggling!
", "rawMarkdown": "I'm still trying to figure out how to do this, so I can't say anything for sure.\n\nBut, for example, there has been a similar competition on Kaggle in the past with audio data instead of images. In that competition, audio of arbitrary length (e.g., 5 - 120 seconds) was given a label of the birdsong present in the audio. However, similar to this competition, there was no information about which part of the audio contained the birdsong, so I remember that the top solutions of the competitors used a weakly supervised learning method called **Sound Event Detection (SED)**.\n\nFor an overview of SED, I think it is best to read the following kernel.\n👉 https://www.kaggle.com/hidehisaarai1213/introduction-to-sound-event-detection\n\nTo put it simply, it uses attention blocks to calculate the loss between the given aggregated labels and the predicted labels for each speech segment by taking the sum of the predicted labels.\nThis explanation may not be enough for you to understand, so please read the kernel if you are interested.\n\nHowever, unlike audio data, this time we are dealing with image data, so the dimension of the data will increase by one. Therefore, it is unclear whether a method like SED is feasible in terms of computational resources.\n\nThanks, @maxwell110 , those are really valuable insights
", "rawMarkdown": "Thanks, @maxwell110 , those are really valuable insights", "votes": 2 }, { "id": 1392331, "postDate": "2021-07-18T14:58:29.240Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1393041, "postDate": "2021-07-19T10:39:22.113Z", "content": "@maxwell110 Judging by other discussions, there are tumor segmentation models. In theory, we can filter tumor-free images
", "rawMarkdown": "@maxwell110 Judging by other discussions, there are tumor segmentation models. In theory, we can filter tumor-free images", "votes": 1 }, { "id": 1511152, "postDate": "2021-09-13T06:57:32.960Z", "content": "@pavelponenkov Would you please send some of those discussions?
", "rawMarkdown": "@pavelponenkov Would you please send some of those discussions?" } ] }, { "id": 1394377, "postDate": "2021-07-20T10:22:24.540Z", "content": "I think you are overthinking it. There is no video, and no variable number of images or missing labels per image. We need to classify each subject. Each subject has exactly four 3D images, one for each imaging modality, i.e, T1, T1ce, T2, and FLAIR. These 3D images have different resolutions, so different number of slices through the head were recorded. In dicoms, you see one image per recorded slice, so a variable number of images, but it is all just parts of the same 3D image.
", "rawMarkdown": "I think you are overthinking it. There is no video, and no variable number of images or missing labels per image. We need to classify each subject. Each subject has exactly four 3D images, one for each imaging modality, i.e, T1, T1ce, T2, and FLAIR. These 3D images have different resolutions, so different number of slices through the head were recorded. In dicoms, you see one image per recorded slice, so a variable number of images, but it is all just parts of the same 3D image. ", "votes": 2, "replies": [ { "id": 1394431, "postDate": "2021-07-20T10:56:35.580Z", "content": "So you are suggesting to look at say, the FLAIR of a patient as a 3D image. Then you would have one FLAIR per patient, with each FLAIR having its own resolution (due to each FLAIR having its own number of images).
\nBut what type of model can accommodate for the varying number of slices of the 3D image without forcing all 3D images (e.g. FLAIR) to have the same resolution (e.g. by interpolation)?
Good question, I don't know. I plan to start by realigning and resampling all 3d images to a common space, segment the tumor using some available tools, for which the images have to be in a same space anyway, and I don't know what I will do such data next :D
", "rawMarkdown": "Good question, I don't know. I plan to start by realigning and resampling all 3d images to a common space, segment the tumor using some available tools, for which the images have to be in a same space anyway, and I don't know what I will do such data next :D " } ] }, { "id": 1464586, "postDate": "2021-08-10T16:16:31.863Z", "content": "From the rules:
\nSubmissions may not use or incorporate information from hand labeling or human prediction of the validation dataset or test data records.
\nMy reading of this is that you can annotate the train dataset (say to develop an algorithm that finds which images have masses). But you cannot hand label the public test data. Also, you don't have access to the private test data, so you cannot hand label that.
", "rawMarkdown": "From the rules:\n\nSubmissions may not use or incorporate information from hand labeling or human prediction of the validation dataset or test data records.\n\nMy reading of this is that you can annotate the train dataset (say to develop an algorithm that finds which images have masses). But you cannot hand label the public test data. Also, you don't have access to the private test data, so you cannot hand label that.", "replies": [ { "id": 1464792, "postDate": "2021-08-10T17:55:54.897Z", "content": "I can't find where it is written
", "rawMarkdown": "I can't find where it is written" } ] }, { "id": 1394120, "postDate": "2021-07-20T07:27:50.920Z", "content": "My first try was to to take a mean of all pictures of a patient for every folder (FLAIR, T1w, T2, T1Gd) an train a CNN on that mean, but it did not work
", "rawMarkdown": "My first try was to to take a mean of all pictures of a patient for every folder (FLAIR, T1w, T2, T1Gd) an train a CNN on that mean, but it did not work" }, { "id": 1388395, "postDate": "2021-07-14T22:37:00.093Z", "content": "Will you also be intending on \" using only T2WI.\"? Thanks for sharing your approach thoughts. New to this type of competition, so it's nice to learn from everyone, bit by bit.
", "rawMarkdown": "Will you also be intending on \" using only T2WI.\"? Thanks for sharing your approach thoughts. New to this type of competition, so it's nice to learn from everyone, bit by bit." }, { "id": 1388065, "postDate": "2021-07-14T16:14:16.483Z", "content": "For segmentation don't we require masks for training? how will we get that
", "rawMarkdown": "For segmentation don't we require masks for training? how will we get that", "replies": [ { "id": 1388136, "postDate": "2021-07-14T17:11:13.313Z", "content": "Yeah, if I get time I will make the masks and share them! But if you have a look at a collection of what these cancers look like (here for example), they're pretty easy to spot even for the layperson
", "rawMarkdown": "Yeah, if I get time I will make the masks and share them! But if you have a look at a collection of what these cancers look like ([here for example](https://radiopaedia.org/articles/glioblastoma)), they're pretty easy to spot even for the layperson" }, { "id": 1388155, "postDate": "2021-07-14T17:21:45.703Z", "content": "Thanks for that link, seems cool.
\nCan you tell me how would you make masks of the images
This is my favourite tool for the job: https://www.medseg.ai/
", "rawMarkdown": "This is my favourite tool for the job: https://www.medseg.ai/", "votes": 2 }, { "id": 1388789, "postDate": "2021-07-15T08:22:57.157Z", "content": "Check https://fets-ai.github.io/Challenge/ I think the organization uploaded the tool they used to obtain the train tumor mask.
", "rawMarkdown": "Check https://fets-ai.github.io/Challenge/ I think the organization uploaded the tool they used to obtain the train tumor mask. ", "votes": 1 }, { "id": 1462561, "postDate": "2021-08-09T22:40:57.827Z", "content": "is manual annotation allowed in this contest?
", "rawMarkdown": "is manual annotation allowed in this contest?" }, { "id": 1466405, "postDate": "2021-08-11T12:55:08.300Z", "content": "that will my question as well, I thought manual annotation is not allowed?
", "rawMarkdown": "that will my question as well, I thought manual annotation is not allowed? " } ] } ], "comments": [ { "id": 1391565, "author_name": "Alvaro Francesc Budria Fernández", "author_url": "", "post_date": "2021-07-17T17:59:58.500000", "content": "Right now we are playing around with an end2end approach, just feeding a 4-channel video to a 3D CNN and see what happens. So far we know that 3D convolution takes quite a long time, even after reducing the dynamic range of the images (and thus reducing the dataset size to ~8GB). We haven't even managed to train a net yet l0l.
\nAnother idea is to try a seq2seq approach, feeding a CNN the successive frames of a video, then sending the embedding to an LSTM and keeping the classification probability at the last time step. We'll see.
I think that's a good point 😊
\nThe method you mentioned at the end above comment will be almost the same as the second place solution in the previous RSNA brain MRI competition.
\n👉 https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/117228
However, I think the main difference from the previous RSNA competition is that the labels are not provided for each image, but for each patient, so it is not clear which cross-sectional image contributes to the 0/1 label. Therefore, I think that this task will be a kind of weakly supervised learning.
\nIn addition, as shown in the histogram below (x-axis: the number of images in each patient), the number of each of the four types of images is different for each patient, so we need to consider how to deal with this.
Thanks, @maxwell110 for your insights, SO how do you think we should approach this problem. Are 3D convolution models the only way?
", "votes": 1, "replies": [] }, { "id": 1391954, "author_name": "Alvaro Francesc Budria Fernández", "author_url": "", "post_date": "2021-07-18T08:53:23.840000", "content": "Yes, we only have a single annotation per video, not per frame, and the amount of frames per video and per patient is variable.
\nTo deal with the annotation thing, we can try at first labeling all the frames with the label of the video, so if a video is labeled 0, then annotate all its frames as 0. I don't know how weakly supervised learning could play out here, so please tell us a bit, it would be very helpful.
\nRegarding the variable length of the videos, we can try padding with zeros to a fixed length, or perhaps interpolating the z-axis to a certain length.
I'm still trying to figure out how to do this, so I can't say anything for sure.
\nBut, for example, there has been a similar competition on Kaggle in the past with audio data instead of images. In that competition, audio of arbitrary length (e.g., 5 - 120 seconds) was given a label of the birdsong present in the audio. However, similar to this competition, there was no information about which part of the audio contained the birdsong, so I remember that the top solutions of the competitors used a weakly supervised learning method called Sound Event Detection (SED).
\nFor an overview of SED, I think it is best to read the following kernel.
\n👉 https://www.kaggle.com/hidehisaarai1213/introduction-to-sound-event-detection
To put it simply, it uses attention blocks to calculate the loss between the given aggregated labels and the predicted labels for each speech segment by taking the sum of the predicted labels.
\nThis explanation may not be enough for you to understand, so please read the kernel if you are interested.
However, unlike audio data, this time we are dealing with image data, so the dimension of the data will increase by one. Therefore, it is unclear whether a method like SED is feasible in terms of computational resources.
\nIn any case, I think that one of the most creative and meaningful moments in Kaggle is when we are thinking about how to apply existing and useful methods to new data and problems. This competition has very little data, which makes me a little nervous about the LB shake, but I hope we can both enjoy it until the end.
\nHappy Kaggling!
", "votes": 10, "replies": [] }, { "id": 1392008, "author_name": "DeepUnderstanding", "author_url": "", "post_date": "2021-07-18T09:35:59.833000", "content": "Thanks, @maxwell110 , those are really valuable insights
", "votes": 2, "replies": [] }, { "id": 1392331, "author_name": "", "author_url": "", "post_date": "2021-07-18T14:58:29.240000", "content": "", "votes": 0, "replies": [] }, { "id": 1393041, "author_name": "Pavel Ponenkov", "author_url": "", "post_date": "2021-07-19T10:39:22.113000", "content": "@maxwell110 Judging by other discussions, there are tumor segmentation models. In theory, we can filter tumor-free images
", "votes": 1, "replies": [] }, { "id": 1511152, "author_name": "Ines Riahi", "author_url": "", "post_date": "2021-09-13T06:57:32.960000", "content": "@pavelponenkov Would you please send some of those discussions?
", "votes": 0, "replies": [] } ] }, { "id": 1394377, "author_name": "RD", "author_url": "", "post_date": "2021-07-20T10:22:24.540000", "content": "I think you are overthinking it. There is no video, and no variable number of images or missing labels per image. We need to classify each subject. Each subject has exactly four 3D images, one for each imaging modality, i.e, T1, T1ce, T2, and FLAIR. These 3D images have different resolutions, so different number of slices through the head were recorded. In dicoms, you see one image per recorded slice, so a variable number of images, but it is all just parts of the same 3D image.
", "votes": 2, "replies": [ { "id": 1394431, "author_name": "Alvaro Francesc Budria Fernández", "author_url": "", "post_date": "2021-07-20T10:56:35.580000", "content": "So you are suggesting to look at say, the FLAIR of a patient as a 3D image. Then you would have one FLAIR per patient, with each FLAIR having its own resolution (due to each FLAIR having its own number of images).
\nBut what type of model can accommodate for the varying number of slices of the 3D image without forcing all 3D images (e.g. FLAIR) to have the same resolution (e.g. by interpolation)?
Good question, I don't know. I plan to start by realigning and resampling all 3d images to a common space, segment the tumor using some available tools, for which the images have to be in a same space anyway, and I don't know what I will do such data next :D
", "votes": 0, "replies": [] } ] }, { "id": 1464586, "author_name": "quadcore/Richard Epstein", "author_url": "", "post_date": "2021-08-10T16:16:31.863000", "content": "From the rules:
\nSubmissions may not use or incorporate information from hand labeling or human prediction of the validation dataset or test data records.
\nMy reading of this is that you can annotate the train dataset (say to develop an algorithm that finds which images have masses). But you cannot hand label the public test data. Also, you don't have access to the private test data, so you cannot hand label that.
", "votes": 0, "replies": [ { "id": 1464792, "author_name": "Zaakcii Ru", "author_url": "", "post_date": "2021-08-10T17:55:54.897000", "content": "I can't find where it is written
", "votes": 0, "replies": [] } ] }, { "id": 1394120, "author_name": "LucaMTB", "author_url": "", "post_date": "2021-07-20T07:27:50.920000", "content": "My first try was to to take a mean of all pictures of a patient for every folder (FLAIR, T1w, T2, T1Gd) an train a CNN on that mean, but it did not work
", "votes": 0, "replies": [] }, { "id": 1388395, "author_name": "ElenaEB", "author_url": "", "post_date": "2021-07-14T22:37:00.093000", "content": "Will you also be intending on \" using only T2WI.\"? Thanks for sharing your approach thoughts. New to this type of competition, so it's nice to learn from everyone, bit by bit.
", "votes": 0, "replies": [] }, { "id": 1388065, "author_name": "DeepUnderstanding", "author_url": "", "post_date": "2021-07-14T16:14:16.483000", "content": "For segmentation don't we require masks for training? how will we get that
", "votes": 0, "replies": [ { "id": 1388136, "author_name": "Reuben Schmidt", "author_url": "", "post_date": "2021-07-14T17:11:13.313000", "content": "Yeah, if I get time I will make the masks and share them! But if you have a look at a collection of what these cancers look like (here for example), they're pretty easy to spot even for the layperson
", "votes": 0, "replies": [] }, { "id": 1388155, "author_name": "DeepUnderstanding", "author_url": "", "post_date": "2021-07-14T17:21:45.703000", "content": "Thanks for that link, seems cool.
\nCan you tell me how would you make masks of the images
This is my favourite tool for the job: https://www.medseg.ai/
", "votes": 2, "replies": [] }, { "id": 1388789, "author_name": "Ranafago", "author_url": "", "post_date": "2021-07-15T08:22:57.157000", "content": "Check https://fets-ai.github.io/Challenge/ I think the organization uploaded the tool they used to obtain the train tumor mask.
", "votes": 1, "replies": [] }, { "id": 1462561, "author_name": "Zaakcii Ru", "author_url": "", "post_date": "2021-08-09T22:40:57.827000", "content": "is manual annotation allowed in this contest?
", "votes": 0, "replies": [] }, { "id": 1466405, "author_name": "Haw Keat", "author_url": "", "post_date": "2021-08-11T12:55:08.300000", "content": "that will my question as well, I thought manual annotation is not allowed?
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1387381": "Having a look at the data, I can see we have 672 subjects (585 train + 87 test) which presumably all feature a brain tumour. Our task is to classify these tumours for the presence or absence of a genetic biomarker - providing a non-invasive screening test to direct treatment for glioblastoma multiforme (GBM, which also goes by the name 'glioblastoma').\n\nMy proposed pipeline is to perform per-voxel classification on the tumour, then produce an output based on a majority voting over the voxel-wise classes. This is similar to the approach described in [this paper](http://www.ajnr.org/content/42/5/845). Interested to hear what other approaches Kagglers are thinking about.", "1391565": "Right now we are playing around with an end2end approach, just feeding a 4-channel video to a 3D CNN and see what happens. So far we know that 3D convolution takes quite a long time, even after reducing the dynamic range of the images (and thus reducing the dataset size to ~8GB). We haven't even managed to train a net yet l0l.\nAnother idea is to try a seq2seq approach, feeding a CNN the successive frames of a video, then sending the embedding to an LSTM and keeping the classification probability at the last time step. We'll see.", "1394377": "I think you are overthinking it. There is no video, and no variable number of images or missing labels per image. We need to classify each subject. Each subject has exactly four 3D images, one for each imaging modality, i.e, T1, T1ce, T2, and FLAIR. These 3D images have different resolutions, so different number of slices through the head were recorded. In dicoms, you see one image per recorded slice, so a variable number of images, but it is all just parts of the same 3D image. ", "1464586": "From the rules:\n\nSubmissions may not use or incorporate information from hand labeling or human prediction of the validation dataset or test data records.\n\nMy reading of this is that you can annotate the train dataset (say to develop an algorithm that finds which images have masses). But you cannot hand label the public test data. Also, you don't have access to the private test data, so you cannot hand label that.", "1394120": "My first try was to to take a mean of all pictures of a patient for every folder (FLAIR, T1w, T2, T1Gd) an train a CNN on that mean, but it did not work", "1388395": "Will you also be intending on \" using only T2WI.\"? Thanks for sharing your approach thoughts. New to this type of competition, so it's nice to learn from everyone, bit by bit.", "1388065": "For segmentation don't we require masks for training? how will we get that" } }