There are 2 tasks for this competition.
\nTask 1: brain tumor segmentation, hosted separately: https://www.synapse.org/#!Synapse:syn25829067/wiki/610863
\nTask 2: radiogenomic classification, hosted here on Kaggle.
\nNot sure why 2 tasks for the same competition are hosted on different platforms.
\nAre we allowed to use Task 1 data for Task 2? For example, training a segmentation model on Task 1 data and applying it to Task 2 data? Technically, you have to register and request access to Task 1 data so not sure if this would be allowed under the external data rules.
\nAppreciate clarification on this issue.
\n", "messages": [ { "id": 1390649, "postDate": "2021-07-16T23:11:36.120Z", "content": "There are 2 tasks for this competition.
\nTask 1: brain tumor segmentation, hosted separately: https://www.synapse.org/#!Synapse:syn25829067/wiki/610863
\nTask 2: radiogenomic classification, hosted here on Kaggle.
\nNot sure why 2 tasks for the same competition are hosted on different platforms.
\nAre we allowed to use Task 1 data for Task 2? For example, training a segmentation model on Task 1 data and applying it to Task 2 data? Technically, you have to register and request access to Task 1 data so not sure if this would be allowed under the external data rules.
\nAppreciate clarification on this issue.
\n", "rawMarkdown": "There are 2 tasks for this competition. \n\nTask 1: brain tumor segmentation, hosted separately: https://www.synapse.org/#!Synapse:syn25829067/wiki/610863\n\nTask 2: radiogenomic classification, hosted here on Kaggle.\n\nNot sure why 2 tasks for the same competition are hosted on different platforms.\n\nAre we allowed to use Task 1 data for Task 2? For example, training a segmentation model on Task 1 data and applying it to Task 2 data? Technically, you have to register and request access to Task 1 data so not sure if this would be allowed under the external data rules. \n\nAppreciate clarification on this issue.\n\n@juliaelliott ", "votes": 82 }, { "id": 1410469, "postDate": "2021-08-02T17:52:50.127Z", "content": "@vaillant Participants are free to use the dataset from Task 1 (Segmentation task not hosted on Kaggle) if they wish, but note that the data for segmentation are based on resampled, co-registered, NIFTI files (of the same number of slices across all series and all subjects). Whereas the classification task's dataset is based on non-resampled, non-co-registered, DICOM files (of varying number of slices). So working with the data between these formats to work with the respective tasks' input & submission requirements may require some creativity.
", "rawMarkdown": "@vaillant Participants are free to use the dataset from Task 1 (Segmentation task not hosted on Kaggle) if they wish, but note that the data for segmentation are based on resampled, co-registered, NIFTI files (of the same number of slices across all series and all subjects). Whereas the classification task's dataset is based on non-resampled, non-co-registered, DICOM files (of varying number of slices). So working with the data between these formats to work with the respective tasks' input & submission requirements may require some creativity.", "votes": 10, "replies": [ { "id": 1464610, "postDate": "2021-08-10T16:24:43.573Z", "content": "Hi @juliaelliott ! Thanks for this clarification.
\nWhat exactly does it mean that data are resampled and co-registered? I guess that the resampling refers to up- or down-sampling to some specified number of slices or resolution, but what does the co-registration refer to?
+1
\nI am curious about this procedure/technique as well.
", "rawMarkdown": "+1\n\nI am curious about this procedure/technique as well.", "votes": 3 }, { "id": 1464678, "postDate": "2021-08-10T16:59:11.857Z", "content": "Co-registration is the process of fitting all images into the same space, I think they registered against T1C.
\nhttps://en.wikipedia.org/wiki/Image_registration
", "rawMarkdown": "Co-registration is the process of fitting all images into the same space, I think they registered against T1C. \n\nhttps://en.wikipedia.org/wiki/Image_registration", "votes": 7 }, { "id": 1464703, "postDate": "2021-08-10T17:06:23.823Z", "content": "Thanks for the quick and poignant response.
\nSo to clarify,
\nFLAIR, T1W and T2W would have affine transformations preformed upon them to transform them into the same orientation/size as T1C (T1wCE). At this point they would then be merged in some way?
\nThe Wikipedia link helped a lot (especially the application section), are there other resources (perhaps related to previous BraTS challenges that you might be aware of?
\nThanks again!
", "rawMarkdown": "Thanks for the quick and poignant response. \n\nSo to clarify,\n\nFLAIR, T1W and T2W would have affine transformations preformed upon them to transform them into the same orientation/size as T1C (T1wCE). At this point they would then be merged in some way?\n\nThe Wikipedia link helped a lot (especially the application section), are there other resources (perhaps related to previous BraTS challenges that you might be aware of?\n\nThanks again!", "votes": 1 }, { "id": 1464732, "postDate": "2021-08-10T17:21:36.277Z", "content": "I think is just a rigid registration, to avoid the deformation of the brain (keeps morphological features ok!), check the data paper from task 1 https://arxiv.org/pdf/2107.02314.pdf. If you want to learn more about registration check https://fsl.fmrib.ox.ac.uk/fslcourse/online_materials.html, but knowing what it is, I think is enough for the challenge. They do not perform any data fusion, that's on us. But using several modalities seems to improve harmonization along different datasets, check https://www.mdpi.com/2073-4425/9/8/382/htm
", "rawMarkdown": "I think is just a rigid registration, to avoid the deformation of the brain (keeps morphological features ok!), check the data paper from task 1 https://arxiv.org/pdf/2107.02314.pdf. If you want to learn more about registration check https://fsl.fmrib.ox.ac.uk/fslcourse/online_materials.html, but knowing what it is, I think is enough for the challenge. They do not perform any data fusion, that's on us. But using several modalities seems to improve harmonization along different datasets, check https://www.mdpi.com/2073-4425/9/8/382/htm", "votes": 2 }, { "id": 1464999, "postDate": "2021-08-10T19:48:43.157Z", "content": "Hi @smoschou55,
\nThe BraTS challenge data for Task 1 (segmentation task) is co-registered to the same anatomical template i.e. SRI24. More details are available at -(https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2915788/ )
", "rawMarkdown": "Hi @smoschou55, \n\nThe BraTS challenge data for Task 1 ([segmentation task](https://www.synapse.org/#!Synapse:syn25829067/wiki/610863)) is co-registered to the same anatomical template i.e. SRI24. More details are available at -(https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2915788/ ) \n\n", "votes": 4 }, { "id": 1465998, "postDate": "2021-08-11T09:07:37.483Z", "content": "Thanks very much, @dschettler8845, @ujjwalbaid, @ranafago ! These are some wonderful resources and things are much clearer now!
\nSo co-registration is just a transformation (and rescaling) and alignment into the same coordinate system and this can happen in a number of ways. From the FSL course Material (https://fsl.fmrib.ox.ac.uk/fslcourse/online_materials.html) it seems to me that:
\nHowever, it is still not clear to me whether you can do this process to align images between different modalities e.g. T2w with FLAIR and T1w, there were originally captured to be in different planes (sagittal, coronal, axial). I guess I'm still trying to wrap my head around this. Maybe we don't really care about doing the co-registration across modalities and consistency within each modality group is sufficient. If you have any insights here would be much appreciated.
\nThanks again! That has been tremendously helpful!
", "rawMarkdown": "Thanks very much, @dschettler8845, @ujjwalbaid, @ranafago ! These are some wonderful resources and things are much clearer now!\n\nSo co-registration is just a transformation (and rescaling) and alignment into the same coordinate system and this can happen in a number of ways. From the FSL course Material (https://fsl.fmrib.ox.ac.uk/fslcourse/online_materials.html) it seems to me that:\n1. the Rigid Registration is great for single-subject transformations (single patient) to align different scans of the same subject\n2. Afine + Non-linear Registrations are good for cross-subject transformations (multi-patient studies) to align scans of all subjects\n\nHowever, it is still not clear to me whether you can do this process to align images between different modalities e.g. T2w with FLAIR and T1w, there were originally captured to be in different planes (sagittal, coronal, axial). I guess I'm still trying to wrap my head around this. Maybe we don't really care about doing the co-registration across modalities and consistency within each modality group is sufficient. If you have any insights here would be much appreciated.\n\nThanks again! That has been tremendously helpful!", "votes": 2 }, { "id": 1466486, "postDate": "2021-08-11T13:40:09.270Z", "content": "From what I've seen results registering different modalities which have been captured in different planes are good, but the methods I've used for registration are too slow for the challenge :(. Take into account that you have around 90 secs to predict each case (9 hours/400 cases). Thanks for the clarification @ujjwalbaid
", "rawMarkdown": "From what I've seen results registering different modalities which have been captured in different planes are good, but the methods I've used for registration are too slow for the challenge :(. Take into account that you have around 90 secs to predict each case (9 hours/400 cases). Thanks for the clarification @ujjwalbaid ", "votes": 2 }, { "id": 1467256, "postDate": "2021-08-11T22:14:08.610Z", "content": "MR images are not actually acquired in the planes we see, rather they're acquired as a volumes or 3D arrays of voxels. Then, they are reconstructed into the planes we visualize them in. This allows the technologist to align the images relative to the anatomy instead of to the machine. Theoretically, all the images in a single study are relative to the same coordinate space. The DICOM tags ImagePositionPatient and ImageOrientationPatient are used to relate image to each other. @boojum made a great notebook demonstrating this -> https://www.kaggle.com/boojum/connecting-voxel-spaces
\nRegistration on same-patient is usually done to track pathologies over time, across studies .. I.E. to determine tumor growth, track nodules etc.
\nI haven't seen an example of registration on different patients. It seems that would require a strictly controlled and clean dataset of studies to be useful.
", "rawMarkdown": "MR images are not actually acquired in the planes we see, rather they're acquired as a volumes or 3D arrays of voxels. Then, they are reconstructed into the planes we visualize them in. This allows the technologist to align the images relative to the anatomy instead of to the machine. *Theoretically*, all the images in a single study are relative to the same coordinate space. The DICOM tags ImagePositionPatient and ImageOrientationPatient are used to relate image to each other. @boojum made a great notebook demonstrating this -> https://www.kaggle.com/boojum/connecting-voxel-spaces\n\nRegistration on same-patient is usually done to track pathologies over time, across studies .. I.E. to determine tumor growth, track nodules etc.\n\nI haven't seen an example of registration on different patients. It seems that would require a strictly controlled and clean dataset of studies to be useful.", "votes": 7 }, { "id": 1467962, "postDate": "2021-08-12T07:42:39.307Z", "content": "Thanks very much @davidbroberts ! That really clarifies a lot, and what a wonderful reference to @boojum 's notebook that I had totally missed!
\nIt does make sense that since this is the same study and the data essentially refer to the same 3D volumes there should be a way to align the views or switch into the same planes between different modalities. Okay, so that answers it and it seems like going through the hassle of messing with different registrations is not the way to go for this challenge and the great pipeline that @boojum put together should be more than enough.
\nThanks again for the help and the information! Great to have some domain expertise shared!
", "rawMarkdown": "Thanks very much @davidbroberts ! That really clarifies a lot, and what a wonderful reference to @boojum 's notebook that I had totally missed! \n\nIt does make sense that since this is the same study and the data essentially refer to the same 3D volumes there should be a way to align the views or switch into the same planes between different modalities. Okay, so that answers it and it seems like going through the hassle of messing with different registrations is not the way to go for this challenge and the great pipeline that @boojum put together should be more than enough.\n\nThanks again for the help and the information! Great to have some domain expertise shared! ", "votes": 2 }, { "id": 1493784, "postDate": "2021-08-28T06:48:06.420Z", "content": "@juliaelliott Thanks for your response!
\nwhat labels do we use for the segmentation data?
", "rawMarkdown": "@juliaelliott Thanks for your response!\n\nwhat labels do we use for the segmentation data?" } ] }, { "id": 1399942, "postDate": "2021-07-25T19:54:03.573Z", "content": "I made a small preprocessing pipeline that transforms the .dmcs into niftis and registers to the SRI24 atlas as the data from task one. However, the registration is very slow… if anybody knows any fast tool It would be very helpful for me.
\nhttps://www.kaggle.com/ranafago/preprocessing-dcm-to-nifti
", "rawMarkdown": "I made a small preprocessing pipeline that transforms the .dmcs into niftis and registers to the SRI24 atlas as the data from task one. However, the registration is very slow... if anybody knows any fast tool It would be very helpful for me.\n\nhttps://www.kaggle.com/ranafago/preprocessing-dcm-to-nifti", "votes": 10, "replies": [ { "id": 1483329, "postDate": "2021-08-20T15:22:00.917Z", "content": "you can see here - there are examples of image registration https://simpleitk.readthedocs.io/en/release/link_examples.html
", "rawMarkdown": "you can see here - there are examples of image registration https://simpleitk.readthedocs.io/en/release/link_examples.html", "votes": 1 } ] }, { "id": 1484586, "postDate": "2021-08-21T12:49:52.147Z", "content": "Just commenting here for visibility that I uploaded the task 1 dataset.
\nHere’s the discussion post with relevant links to the dataset and a notebook showing how to access the images (untar and load)
\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/266679
", "rawMarkdown": "Just commenting here for visibility that **I uploaded the task 1 dataset.**\n\nHere’s the discussion post with relevant links to the dataset and a notebook showing how to access the images (untar and load)\n\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/266679", "votes": 6, "replies": [ { "id": 1491282, "postDate": "2021-08-26T09:20:39.637Z", "content": "I have used you dataset, and you unzip notebook, for running the first 300 imgs into UNET, and I have upload the results into this public dataset in order to try segmentation, lets see how figure out the non-resampled or non-co-registered things.
\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/268169
", "rawMarkdown": "I have used you dataset, and you unzip notebook, for running the first 300 imgs into UNET, and I have upload the results into this public dataset in order to try segmentation, lets see how figure out the non-resampled or non-co-registered things. \n\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/268169", "votes": 2 } ] }, { "id": 1393733, "postDate": "2021-07-19T21:12:40.723Z", "content": "There is some data in the dataset for task 2 missing in the dataset for task 1 (Cases 169, 245, 308, 408, 564, 688, 794, 998), and the task 2 testing dataset is not in the task 1 dataset.
", "rawMarkdown": "There is some data in the dataset for task 2 missing in the dataset for task 1 (Cases 169, 245, 308, 408, 564, 688, 794, 998), and the task 2 testing dataset is not in the task 1 dataset. ", "votes": 3, "replies": [ { "id": 1484588, "postDate": "2021-08-21T12:51:20.650Z", "content": "Thanks for this!
", "rawMarkdown": "Thanks for this! " } ] }, { "id": 1493782, "postDate": "2021-08-28T06:45:23.107Z", "content": "here is a Notebook to load BraTS 2021 data into kaggle and the no. of cases missing in the classification dataset is compared at the end.
", "rawMarkdown": "here is a [Notebook](https://www.kaggle.com/snish9/get-brats-2021) to load BraTS 2021 data into kaggle and the no. of cases missing in the classification dataset is compared at the end.", "votes": 1 }, { "id": 1393516, "postDate": "2021-07-19T17:43:11.793Z", "content": "Hi Ian, I am just wondering why you state there are two tasks. It seems to me that it is a possibility to segment and then classify, but you could also classify directly. What am I missing?
\nThanks!
that's what I am thinking. I just started looking at this competition. It seems to be is segmentation followed by classification. Please let me know if I am not thinking in the correct direction.
", "rawMarkdown": "that's what I am thinking. I just started looking at this competition. It seems to be is segmentation followed by classification. Please let me know if I am not thinking in the correct direction. " } ] }, { "id": 1398956, "postDate": "2021-07-24T17:38:41.307Z", "content": "I think if you'll use the Task 1 data during Task 2 training you'll may have problems during inference time if you'll not have the test data in the same Task 1 format
", "rawMarkdown": "I think if you'll use the Task 1 data during Task 2 training you'll may have problems during inference time if you'll not have the test data in the same Task 1 format", "votes": -2 }, { "id": 1446353, "postDate": "2021-08-04T11:58:59.390Z", "content": "Not sure if we can use data from task 1 to task 2. Can anyone clarify this.
", "rawMarkdown": "Not sure if we can use data from task 1 to task 2. Can anyone clarify this." }, { "id": 1393668, "postDate": "2021-07-19T19:34:15.027Z", "rawMarkdown": "", "votes": 10, "isDeleted": true }, { "id": 1390866, "postDate": "2021-07-17T06:13:57.350Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1410469, "author_name": "Julia Elliott", "author_url": "", "post_date": "2021-08-02T17:52:50.127000", "content": "@vaillant Participants are free to use the dataset from Task 1 (Segmentation task not hosted on Kaggle) if they wish, but note that the data for segmentation are based on resampled, co-registered, NIFTI files (of the same number of slices across all series and all subjects). Whereas the classification task's dataset is based on non-resampled, non-co-registered, DICOM files (of varying number of slices). So working with the data between these formats to work with the respective tasks' input & submission requirements may require some creativity.
", "votes": 10, "replies": [ { "id": 1464610, "author_name": "Sofia Moschou", "author_url": "", "post_date": "2021-08-10T16:24:43.573000", "content": "Hi @juliaelliott ! Thanks for this clarification.
\nWhat exactly does it mean that data are resampled and co-registered? I guess that the resampling refers to up- or down-sampling to some specified number of slices or resolution, but what does the co-registration refer to?
+1
\nI am curious about this procedure/technique as well.
", "votes": 3, "replies": [] }, { "id": 1464678, "author_name": "Ranafago", "author_url": "", "post_date": "2021-08-10T16:59:11.857000", "content": "Co-registration is the process of fitting all images into the same space, I think they registered against T1C.
\nhttps://en.wikipedia.org/wiki/Image_registration
", "votes": 7, "replies": [] }, { "id": 1464703, "author_name": "Darien Schettler", "author_url": "", "post_date": "2021-08-10T17:06:23.823000", "content": "Thanks for the quick and poignant response.
\nSo to clarify,
\nFLAIR, T1W and T2W would have affine transformations preformed upon them to transform them into the same orientation/size as T1C (T1wCE). At this point they would then be merged in some way?
\nThe Wikipedia link helped a lot (especially the application section), are there other resources (perhaps related to previous BraTS challenges that you might be aware of?
\nThanks again!
", "votes": 1, "replies": [] }, { "id": 1464732, "author_name": "Ranafago", "author_url": "", "post_date": "2021-08-10T17:21:36.277000", "content": "I think is just a rigid registration, to avoid the deformation of the brain (keeps morphological features ok!), check the data paper from task 1 https://arxiv.org/pdf/2107.02314.pdf. If you want to learn more about registration check https://fsl.fmrib.ox.ac.uk/fslcourse/online_materials.html, but knowing what it is, I think is enough for the challenge. They do not perform any data fusion, that's on us. But using several modalities seems to improve harmonization along different datasets, check https://www.mdpi.com/2073-4425/9/8/382/htm
", "votes": 2, "replies": [] }, { "id": 1464999, "author_name": "Ujjwal", "author_url": "", "post_date": "2021-08-10T19:48:43.157000", "content": "Hi @smoschou55,
\nThe BraTS challenge data for Task 1 (segmentation task) is co-registered to the same anatomical template i.e. SRI24. More details are available at -(https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2915788/ )
", "votes": 4, "replies": [] }, { "id": 1465998, "author_name": "Sofia Moschou", "author_url": "", "post_date": "2021-08-11T09:07:37.483000", "content": "Thanks very much, @dschettler8845, @ujjwalbaid, @ranafago ! These are some wonderful resources and things are much clearer now!
\nSo co-registration is just a transformation (and rescaling) and alignment into the same coordinate system and this can happen in a number of ways. From the FSL course Material (https://fsl.fmrib.ox.ac.uk/fslcourse/online_materials.html) it seems to me that:
\nHowever, it is still not clear to me whether you can do this process to align images between different modalities e.g. T2w with FLAIR and T1w, there were originally captured to be in different planes (sagittal, coronal, axial). I guess I'm still trying to wrap my head around this. Maybe we don't really care about doing the co-registration across modalities and consistency within each modality group is sufficient. If you have any insights here would be much appreciated.
\nThanks again! That has been tremendously helpful!
", "votes": 2, "replies": [] }, { "id": 1466486, "author_name": "Ranafago", "author_url": "", "post_date": "2021-08-11T13:40:09.270000", "content": "From what I've seen results registering different modalities which have been captured in different planes are good, but the methods I've used for registration are too slow for the challenge :(. Take into account that you have around 90 secs to predict each case (9 hours/400 cases). Thanks for the clarification @ujjwalbaid
", "votes": 2, "replies": [] }, { "id": 1467256, "author_name": "David Roberts", "author_url": "", "post_date": "2021-08-11T22:14:08.610000", "content": "MR images are not actually acquired in the planes we see, rather they're acquired as a volumes or 3D arrays of voxels. Then, they are reconstructed into the planes we visualize them in. This allows the technologist to align the images relative to the anatomy instead of to the machine. Theoretically, all the images in a single study are relative to the same coordinate space. The DICOM tags ImagePositionPatient and ImageOrientationPatient are used to relate image to each other. @boojum made a great notebook demonstrating this -> https://www.kaggle.com/boojum/connecting-voxel-spaces
\nRegistration on same-patient is usually done to track pathologies over time, across studies .. I.E. to determine tumor growth, track nodules etc.
\nI haven't seen an example of registration on different patients. It seems that would require a strictly controlled and clean dataset of studies to be useful.
", "votes": 7, "replies": [] }, { "id": 1467962, "author_name": "Sofia Moschou", "author_url": "", "post_date": "2021-08-12T07:42:39.307000", "content": "Thanks very much @davidbroberts ! That really clarifies a lot, and what a wonderful reference to @boojum 's notebook that I had totally missed!
\nIt does make sense that since this is the same study and the data essentially refer to the same 3D volumes there should be a way to align the views or switch into the same planes between different modalities. Okay, so that answers it and it seems like going through the hassle of messing with different registrations is not the way to go for this challenge and the great pipeline that @boojum put together should be more than enough.
\nThanks again for the help and the information! Great to have some domain expertise shared!
", "votes": 2, "replies": [] }, { "id": 1493784, "author_name": "Nishant Sachdeva", "author_url": "", "post_date": "2021-08-28T06:48:06.420000", "content": "@juliaelliott Thanks for your response!
\nwhat labels do we use for the segmentation data?
", "votes": 0, "replies": [] } ] }, { "id": 1399942, "author_name": "Ranafago", "author_url": "", "post_date": "2021-07-25T19:54:03.573000", "content": "I made a small preprocessing pipeline that transforms the .dmcs into niftis and registers to the SRI24 atlas as the data from task one. However, the registration is very slow… if anybody knows any fast tool It would be very helpful for me.
\nhttps://www.kaggle.com/ranafago/preprocessing-dcm-to-nifti
", "votes": 10, "replies": [ { "id": 1483329, "author_name": "Zaakcii Ru", "author_url": "", "post_date": "2021-08-20T15:22:00.917000", "content": "you can see here - there are examples of image registration https://simpleitk.readthedocs.io/en/release/link_examples.html
", "votes": 1, "replies": [] } ] }, { "id": 1484586, "author_name": "Darien Schettler", "author_url": "", "post_date": "2021-08-21T12:49:52.147000", "content": "Just commenting here for visibility that I uploaded the task 1 dataset.
\nHere’s the discussion post with relevant links to the dataset and a notebook showing how to access the images (untar and load)
\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/266679
", "votes": 6, "replies": [ { "id": 1491282, "author_name": "Victor Fernandez Albor", "author_url": "", "post_date": "2021-08-26T09:20:39.637000", "content": "I have used you dataset, and you unzip notebook, for running the first 300 imgs into UNET, and I have upload the results into this public dataset in order to try segmentation, lets see how figure out the non-resampled or non-co-registered things.
\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/268169
", "votes": 2, "replies": [] } ] }, { "id": 1393733, "author_name": "Ranafago", "author_url": "", "post_date": "2021-07-19T21:12:40.723000", "content": "There is some data in the dataset for task 2 missing in the dataset for task 1 (Cases 169, 245, 308, 408, 564, 688, 794, 998), and the task 2 testing dataset is not in the task 1 dataset.
", "votes": 3, "replies": [ { "id": 1484588, "author_name": "Darien Schettler", "author_url": "", "post_date": "2021-08-21T12:51:20.650000", "content": "Thanks for this!
", "votes": 0, "replies": [] } ] }, { "id": 1493782, "author_name": "Nishant Sachdeva", "author_url": "", "post_date": "2021-08-28T06:45:23.107000", "content": "here is a Notebook to load BraTS 2021 data into kaggle and the no. of cases missing in the classification dataset is compared at the end.
", "votes": 1, "replies": [] }, { "id": 1393516, "author_name": "Soren", "author_url": "", "post_date": "2021-07-19T17:43:11.793000", "content": "Hi Ian, I am just wondering why you state there are two tasks. It seems to me that it is a possibility to segment and then classify, but you could also classify directly. What am I missing?
\nThanks!
that's what I am thinking. I just started looking at this competition. It seems to be is segmentation followed by classification. Please let me know if I am not thinking in the correct direction.
", "votes": 0, "replies": [] } ] }, { "id": 1398956, "author_name": "Alfonso", "author_url": "", "post_date": "2021-07-24T17:38:41.307000", "content": "I think if you'll use the Task 1 data during Task 2 training you'll may have problems during inference time if you'll not have the test data in the same Task 1 format
", "votes": -2, "replies": [] }, { "id": 1446353, "author_name": "NisMau", "author_url": "", "post_date": "2021-08-04T11:58:59.390000", "content": "Not sure if we can use data from task 1 to task 2. Can anyone clarify this.
", "votes": 0, "replies": [] }, { "id": 1393668, "author_name": "", "author_url": "", "post_date": "2021-07-19T19:34:15.027000", "content": "", "votes": 10, "replies": [] }, { "id": 1390866, "author_name": "", "author_url": "", "post_date": "2021-07-17T06:13:57.350000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1390649": "There are 2 tasks for this competition. \n\nTask 1: brain tumor segmentation, hosted separately: https://www.synapse.org/#!Synapse:syn25829067/wiki/610863\n\nTask 2: radiogenomic classification, hosted here on Kaggle.\n\nNot sure why 2 tasks for the same competition are hosted on different platforms.\n\nAre we allowed to use Task 1 data for Task 2? For example, training a segmentation model on Task 1 data and applying it to Task 2 data? Technically, you have to register and request access to Task 1 data so not sure if this would be allowed under the external data rules. \n\nAppreciate clarification on this issue.\n\n@juliaelliott ", "1410469": "@vaillant Participants are free to use the dataset from Task 1 (Segmentation task not hosted on Kaggle) if they wish, but note that the data for segmentation are based on resampled, co-registered, NIFTI files (of the same number of slices across all series and all subjects). Whereas the classification task's dataset is based on non-resampled, non-co-registered, DICOM files (of varying number of slices). So working with the data between these formats to work with the respective tasks' input & submission requirements may require some creativity.", "1399942": "I made a small preprocessing pipeline that transforms the .dmcs into niftis and registers to the SRI24 atlas as the data from task one. However, the registration is very slow... if anybody knows any fast tool It would be very helpful for me.\n\nhttps://www.kaggle.com/ranafago/preprocessing-dcm-to-nifti", "1484586": "Just commenting here for visibility that **I uploaded the task 1 dataset.**\n\nHere’s the discussion post with relevant links to the dataset and a notebook showing how to access the images (untar and load)\n\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/266679", "1393733": "There is some data in the dataset for task 2 missing in the dataset for task 1 (Cases 169, 245, 308, 408, 564, 688, 794, 998), and the task 2 testing dataset is not in the task 1 dataset. ", "1493782": "here is a [Notebook](https://www.kaggle.com/snish9/get-brats-2021) to load BraTS 2021 data into kaggle and the no. of cases missing in the classification dataset is compared at the end.", "1393516": "Hi Ian, I am just wondering why you state there are two tasks. It seems to me that it is a possibility to segment and then classify, but you could also classify directly. What am I missing?\nThanks!", "1398956": "I think if you'll use the Task 1 data during Task 2 training you'll may have problems during inference time if you'll not have the test data in the same Task 1 format", "1446353": "Not sure if we can use data from task 1 to task 2. Can anyone clarify this.", "1393668": "", "1390866": "" } }