Based on some of the discussion posts we wanted to clarify the three types of labels provided in the challenge and give more reasoning and clinical context.
\nStudy level labels:
\na. There are study level labels for bowel, extravasation, liver, spleen, and kidneys.
\nb. For bowel and extravasation, the study level label is either injury or no injury.
\nc. For liver, spleen and kidneys, there are three classes: no injury, low-grade injury or high-grade injury. Low-grade injuries refer to AAST injury grades I, II and III and high-grade injuries refer to grades IV and V. The AAST injury scales are different for different organs will have different appearances- https://radiopaedia.org/articles/aast-injury-scoring-scales. The reasoning for this separation is that typically grade IV and V injuries require surgical intervention while grades I, II and III do not.
\nd. There is also a study level label for “any_injury” which is positive if any of the organs are injured.
\ne. The study level labels are what we are asking for predictions on the test cases.
Image level labels:
\na. These are provided only for extravasation and bowel injury. This is because these types of injures can have a more variable appearance and anatomic location. Extravasation refers to active bleeding, which can be seen anywhere in the body. including within organs and in different soft tissues. Bowel injury includes injuries to the mesentery, which is what attaches the intestines to the abdominal wall.
Segmentations:
\na. In order to provide more anatomical context to where the injuries will will be present, we have provided voxelwise segmentations on a subset of 206 training cases that are enriched for the presence of significant injuries.
\nb. These segmentations were manually revised by expert radiologists after being generated from a model trained with 3d_fullres nnU-Net on the total segmentator dataset focusing on the organs being evaluated (liver, spleen, left kidney, right kidney and bowel)
\nc. We modified the publicly available total segmentator segmentation labels to follow our schema just using labels for liver, spleen, left kidney, right kidney, and bowel (which combined esophagus, stomach, duodenum, small bowel and colon)
\nd. The total segmentator paper is here: https://pubs.rsna.org/doi/10.1148/ryai.230024
\ne. The total segmentator dataset is here: https://zenodo.org/record/6802614
\nf. The modified total segmentator segmentations used to train the initial model (in *.nii.gz format) can be downloaded here: https://drive.google.com/file/d/1zoSFVlls-j6IUFWuiASFrb1OqVZO37yR/view?usp=sharing
\ng. The segmentation labels are as follows:
\ni. 0 = background
\nii. 1 = liver
\niii. 2 = spleen
\niv. 3 = left kidney
\nv. 4 = right kidney
\nvi. 5 = bowel
\nh. While different labels are provided for left and right kidney, we are asking for the prediction of injury in either kidney. If the left kidney has a low-grade injury while the right kidney has a high-grade injury, the correct prediction would be high-grade.
\ni. Please be aware that the NIFTI files and DICOM files are not in the same orientation. Use the NIFTI header information along with DICOM metadata to determine the appropriate orientation such that the DICOM images and segmentation match.
Jeff Rudie, MD PhD, on behalf of the organizing committee
", "messages": [ { "id": 2369506, "postDate": "2023-08-01T19:28:30.203Z", "content": "Based on some of the discussion posts we wanted to clarify the three types of labels provided in the challenge and give more reasoning and clinical context.
\nStudy level labels:
\na. There are study level labels for bowel, extravasation, liver, spleen, and kidneys.
\nb. For bowel and extravasation, the study level label is either injury or no injury.
\nc. For liver, spleen and kidneys, there are three classes: no injury, low-grade injury or high-grade injury. Low-grade injuries refer to AAST injury grades I, II and III and high-grade injuries refer to grades IV and V. The AAST injury scales are different for different organs will have different appearances- https://radiopaedia.org/articles/aast-injury-scoring-scales. The reasoning for this separation is that typically grade IV and V injuries require surgical intervention while grades I, II and III do not.
\nd. There is also a study level label for “any_injury” which is positive if any of the organs are injured.
\ne. The study level labels are what we are asking for predictions on the test cases.
Image level labels:
\na. These are provided only for extravasation and bowel injury. This is because these types of injures can have a more variable appearance and anatomic location. Extravasation refers to active bleeding, which can be seen anywhere in the body. including within organs and in different soft tissues. Bowel injury includes injuries to the mesentery, which is what attaches the intestines to the abdominal wall.
Segmentations:
\na. In order to provide more anatomical context to where the injuries will will be present, we have provided voxelwise segmentations on a subset of 206 training cases that are enriched for the presence of significant injuries.
\nb. These segmentations were manually revised by expert radiologists after being generated from a model trained with 3d_fullres nnU-Net on the total segmentator dataset focusing on the organs being evaluated (liver, spleen, left kidney, right kidney and bowel)
\nc. We modified the publicly available total segmentator segmentation labels to follow our schema just using labels for liver, spleen, left kidney, right kidney, and bowel (which combined esophagus, stomach, duodenum, small bowel and colon)
\nd. The total segmentator paper is here: https://pubs.rsna.org/doi/10.1148/ryai.230024
\ne. The total segmentator dataset is here: https://zenodo.org/record/6802614
\nf. The modified total segmentator segmentations used to train the initial model (in *.nii.gz format) can be downloaded here: https://drive.google.com/file/d/1zoSFVlls-j6IUFWuiASFrb1OqVZO37yR/view?usp=sharing
\ng. The segmentation labels are as follows:
\ni. 0 = background
\nii. 1 = liver
\niii. 2 = spleen
\niv. 3 = left kidney
\nv. 4 = right kidney
\nvi. 5 = bowel
\nh. While different labels are provided for left and right kidney, we are asking for the prediction of injury in either kidney. If the left kidney has a low-grade injury while the right kidney has a high-grade injury, the correct prediction would be high-grade.
\ni. Please be aware that the NIFTI files and DICOM files are not in the same orientation. Use the NIFTI header information along with DICOM metadata to determine the appropriate orientation such that the DICOM images and segmentation match.
Jeff Rudie, MD PhD, on behalf of the organizing committee
", "rawMarkdown": "Based on some of the discussion posts we wanted to clarify the three types of labels provided in the challenge and give more reasoning and clinical context.\n\n1.\tStudy level labels:\na.\tThere are study level labels for bowel, extravasation, liver, spleen, and kidneys. \nb.\tFor bowel and extravasation, the study level label is either injury or no injury. \nc.\tFor liver, spleen and kidneys, there are three classes: no injury, low-grade injury or high-grade injury. Low-grade injuries refer to AAST injury grades I, II and III and high-grade injuries refer to grades IV and V. The AAST injury scales are different for different organs will have different appearances- https://radiopaedia.org/articles/aast-injury-scoring-scales. The reasoning for this separation is that typically grade IV and V injuries require surgical intervention while grades I, II and III do not. \nd.\tThere is also a study level label for “any_injury” which is positive if any of the organs are injured.\ne.\tThe study level labels are what we are asking for predictions on the test cases.\n\n2.\tImage level labels:\na.\tThese are provided only for extravasation and bowel injury. This is because these types of injures can have a more variable appearance and anatomic location. Extravasation refers to active bleeding, which can be seen anywhere in the body. including within organs and in different soft tissues. Bowel injury includes injuries to the mesentery, which is what attaches the intestines to the abdominal wall. \n\n3.\tSegmentations: \na.\tIn order to provide more anatomical context to where the injuries will will be present, we have provided voxelwise segmentations on a subset of 206 training cases that are enriched for the presence of significant injuries. \nb.\tThese segmentations were manually revised by expert radiologists after being generated from a model trained with 3d_fullres nnU-Net on the total segmentator dataset focusing on the organs being evaluated (liver, spleen, left kidney, right kidney and bowel)\nc.\tWe modified the publicly available total segmentator segmentation labels to follow our schema just using labels for liver, spleen, left kidney, right kidney, and bowel (which combined esophagus, stomach, duodenum, small bowel and colon)\nd.\tThe total segmentator paper is here: https://pubs.rsna.org/doi/10.1148/ryai.230024\ne.\tThe total segmentator dataset is here: https://zenodo.org/record/6802614\nf.\tThe modified total segmentator segmentations used to train the initial model (in *.nii.gz format) can be downloaded here: https://drive.google.com/file/d/1zoSFVlls-j6IUFWuiASFrb1OqVZO37yR/view?usp=sharing\ng.\tThe segmentation labels are as follows:\ni.\t0 = background\nii.\t1 = liver\niii.\t2 = spleen\niv.\t3 = left kidney\nv.\t4 = right kidney\nvi.\t5 = bowel \nh.\tWhile different labels are provided for left and right kidney, we are asking for the prediction of injury in either kidney. If the left kidney has a low-grade injury while the right kidney has a high-grade injury, the correct prediction would be high-grade. \ni.\tPlease be aware that the NIFTI files and DICOM files are not in the same orientation. Use the NIFTI header information along with DICOM metadata to determine the appropriate orientation such that the DICOM images and segmentation match.\n\nJeff Rudie, MD PhD, on behalf of the organizing committee\n", "votes": 78 }, { "id": 2404893, "postDate": "2023-08-23T15:06:02.070Z", "content": "@jeffrudie,
\n> Please be aware that the NIFTI files and DICOM files are not in the same orientation. Use the NIFTI header information along with DICOM metadata to determine the appropriate orientation such that the DICOM images and segmentation match.
In extravasation injuries are you considering PSEUDOANEURYSM or UBO (Unidentified Bright Objects) as EXTRAVASATION?
", "rawMarkdown": "In extravasation injuries are you considering PSEUDOANEURYSM or UBO (Unidentified Bright Objects) as EXTRAVASATION?\n", "votes": 4, "replies": [ { "id": 2394187, "postDate": "2023-08-16T18:41:53.550Z", "content": "both pseudoaneurysms and unidentified bright objects could potentially involve extravasation in some cases
", "rawMarkdown": " both pseudoaneurysms and unidentified bright objects could potentially involve extravasation in some cases", "votes": 2 }, { "id": 2394197, "postDate": "2023-08-16T18:44:17.800Z", "content": "Pseudoaneurysm: This is a type of vascular injury that can occur due to trauma, medical procedures, or other causes. It involves a small hole in the wall of an artery or vein, leading to the formation of a pulsating sac filled with blood. Pseudoaneurysms can result in extravasation of blood into surrounding tissues, leading to swelling, pain, and potential complications. It is considered a form of vascular injury and can indeed be associated with extravasation.
\nUnidentified Bright Objects (UBO): This term is not widely recognized in the medical community , However, in medical imaging, \"bright objects\" typically refer to abnormal areas that appear brighter than the surrounding tissue. These can be indicative of various conditions, such as cysts, tumors, or other abnormalities. Whether UBOs are considered a form of extravasation would depend on the specific context in which the term is being used and the underlying cause of the bright appearance on the imaging.
", "rawMarkdown": "Pseudoaneurysm: This is a type of vascular injury that can occur due to trauma, medical procedures, or other causes. It involves a small hole in the wall of an artery or vein, leading to the formation of a pulsating sac filled with blood. Pseudoaneurysms can result in extravasation of blood into surrounding tissues, leading to swelling, pain, and potential complications. It is considered a form of vascular injury and can indeed be associated with extravasation.\n\nUnidentified Bright Objects (UBO): This term is not widely recognized in the medical community , However, in medical imaging, \"bright objects\" typically refer to abnormal areas that appear brighter than the surrounding tissue. These can be indicative of various conditions, such as cysts, tumors, or other abnormalities. Whether UBOs are considered a form of extravasation would depend on the specific context in which the term is being used and the underlying cause of the bright appearance on the imaging.", "votes": 4 } ] }, { "id": 2392293, "postDate": "2023-08-15T15:23:19.277Z", "content": "@jeffrudie
\nHello, jeffrudie. Thank you for your clarification.
\nI have a question about image level label. I took a look into the image level data and found that not all of the series_id of a certain patient are labeled.
\nFor example, patiend_id 43 has a extravasation injury, the correspond series_id are 24055 and 36714. In the image level data, only instance_id of 24055 are given, without the information about 36714. Does this mean no extravasation injury is found in 36714. or else is this simply means lack of annotation?
And what about the annotation in bowel?
\nHope for your reply
Extravasation Injuries are often confirmed on Sequential Scans. So, in this case, you will not only need to consider Series A but A and B in unison to figure out if Extravasation is present. So probably the Extravasation is seen on instance_id of 24055. That is how we radiologists report extravasation.
", "rawMarkdown": "Extravasation Injuries are often confirmed on Sequential Scans. So, in this case, you will not only need to consider Series A but A and B in unison to figure out if Extravasation is present. So probably the Extravasation is seen on instance_id of 24055. That is how we radiologists report extravasation. ", "votes": 4, "replies": [ { "id": 2392457, "postDate": "2023-08-15T17:27:34.190Z", "content": "Hi, thanks for your reply!
\nI don't have the domain knowledge and it is really helpful.
\nIn the example above, given Series A(24055) and Series B(36714), to figure out extravasation, radiologists look into both of the Series. In this case, extravasation is only seen on series_id 24055, does this mean extravasation is not seen in 36714?
And what about the bowel injury, do radiologist confirm bowel injury on Sequential Scans?
", "rawMarkdown": "Hi, thanks for your reply!\nI don't have the domain knowledge and it is really helpful.\nIn the example above, given Series A(24055) and Series B(36714), to figure out extravasation, radiologists look into both of the Series. In this case, extravasation is only seen on series_id 24055, does this mean extravasation is not seen in 36714?\n\nAnd what about the bowel injury, do radiologist confirm bowel injury on Sequential Scans?", "votes": 4 }, { "id": 2394364, "postDate": "2023-08-16T21:20:48.990Z", "content": "Extravasations need sequential scans. Bowel injury doesn't.
", "rawMarkdown": "Extravasations need sequential scans. Bowel injury doesn't.", "votes": 4 }, { "id": 2427277, "postDate": "2023-09-07T06:25:28.173Z", "content": "Oh wow, very helpful!
\nAre sequential scans used otherwise? Or only yo check extravasation injuries?
In competition rules, there is one rule which is no access internet. So can we use totalsegmantator model for segmantation ?
", "rawMarkdown": "In competition rules, there is one rule which is no access internet. So can we use totalsegmantator model for segmantation ?", "votes": 1, "replies": [ { "id": 2378524, "postDate": "2023-08-07T17:11:21.313Z", "content": "You can train your own segmentation model with the total segmentator dataset and/or the data we have provided with segmentations.
", "rawMarkdown": "You can train your own segmentation model with the total segmentator dataset and/or the data we have provided with segmentations.", "votes": 3, "replies": [ { "id": 2392434, "postDate": "2023-08-15T17:18:33.950Z", "content": "We can use external datasets to train as well?
", "rawMarkdown": "We can use external datasets to train as well?" }, { "id": 2408964, "postDate": "2023-08-26T00:35:44.473Z", "content": "I see you did not get answer from the host. Kaggle general rule on training for this type of competition is that external are ok if they are freely available to all - if you can find it on the internet and download without fee or subscription than its ok.
", "rawMarkdown": "I see you did not get answer from the host. Kaggle general rule on training for this type of competition is that external are ok if they are freely available to all - if you can find it on the internet and download without fee or subscription than its ok. ", "votes": 1 } ] } ] }, { "id": 2369718, "postDate": "2023-08-02T01:36:24.053Z", "content": "Thanks. You posted very useful information. But can you talk about train.csv. Labels represent only the patient level, each individual image has no label for it.
", "rawMarkdown": "Thanks. You posted very useful information. But can you talk about train.csv. Labels represent only the patient level, each individual image has no label for it.", "votes": 1, "replies": [ { "id": 2370027, "postDate": "2023-08-02T06:45:05.983Z", "content": "Yes labels in the train.csv are only provided at the patient/study level. Some patients may have two scans/series, which are different contrast phases (think of them as timepoints/4th dimension). Injuries may be present in one or both of the different series/contrast phases.
\nImage/series level labels were only created for active extravasation and bowel injuries given the variable appearance/location of these types of injuries. Image level labels were not generated for other organ injuries (liver, spleen, kidneys) given the time/effort to create the image level labels and since an organ segmentation models should provide information to localize these organs and help narrow focus to detecting/classify injuries.
Is a segmentation model required for submission as mentioned in the comment above? Is it not simply about building a model and providing predictions for the three DCM files in the provided test images folder, rather than predictions for these three DCM files (2D images)?
", "rawMarkdown": "@jeffrudie \n\nIs a segmentation model required for submission as mentioned in the comment above? Is it not simply about building a model and providing predictions for the three DCM files in the provided test images folder, rather than predictions for these three DCM files (2D images)?" }, { "id": 2436688, "postDate": "2023-09-13T17:53:04.787Z", "content": "A segmentation model is not required. Only a csv file with predictions for study level labels are required. However, it's likely that incorporating a segmentation model will help improve performance.
", "rawMarkdown": "A segmentation model is not required. Only a csv file with predictions for study level labels are required. However, it's likely that incorporating a segmentation model will help improve performance. " }, { "id": 2440366, "postDate": "2023-09-15T13:07:52.593Z", "content": "\nI believe there could be errors in constructing a segmentation model because the entire DICOM dataset is labeled by patient_id. The following situations may result in discrepancies:
\nTo address these issues, it seems inevitable that we must assess the degree of predicted damage per DICOM and input labels accordingly. However, we cannot directly use the labels of 'healthy', 'low', and 'high' assigned at the patient level, nor can we label them ourselves because we are not medical professionals to make such judgments.
", "rawMarkdown": "@jeffrudie \n\nI believe there could be errors in constructing a segmentation model because the entire DICOM dataset is labeled by patient_id. The following situations may result in discrepancies:\n\n1. Organ damage might be classified as a high injury in certain DICOM slices, but as we move away from the wound area, it could become challenging to distinguish between low injury or healthy conditions.\n2. Since a particular DICOM slice represents a single 2D image, predictions based on 3D images might not operate correctly for this slice.\n\nTo address these issues, it seems inevitable that we must assess the degree of predicted damage per DICOM and input labels accordingly. However, we cannot directly use the labels of 'healthy', 'low', and 'high' assigned at the patient level, nor can we label them ourselves because we are not medical professionals to make such judgments.\n" }, { "id": 2440713, "postDate": "2023-09-15T17:15:33.120Z", "content": "The classification for organ damage is a single value for the entire patient. If a single image shows high grade injury for that organ then the label is high grade regardless of whether other slices are low grade or normal. A segmentation model could help narrow where your search to determine if a particular organ is injured.
", "rawMarkdown": "The classification for organ damage is a single value for the entire patient. If a single image shows high grade injury for that organ then the label is high grade regardless of whether other slices are low grade or normal. A segmentation model could help narrow where your search to determine if a particular organ is injured. " }, { "id": 2442341, "postDate": "2023-09-16T22:57:07.973Z", "content": "but during test time , we dosent have semgnation data and involving semgnation to tested images will probabaly couse to exceed the time , so how its can be done ?
", "rawMarkdown": "but during test time , we dosent have semgnation data and involving semgnation to tested images will probabaly couse to exceed the time , so how its can be done ? ", "votes": 1 } ] } ] }, { "id": 2374299, "postDate": "2023-08-04T20:37:20.383Z", "content": "Could we access the segmentator paper as it seems to be behind a paywall.
\nhttps://pubs.rsna.org/doi/full/10.1148/ryai.230024
https://arxiv.org/abs/2208.05868
", "rawMarkdown": "https://arxiv.org/abs/2208.05868", "votes": 9, "replies": [ { "id": 2374327, "postDate": "2023-08-04T21:27:11.797Z", "content": "Thanks that's great!
", "rawMarkdown": "Thanks that's great!" }, { "id": 2374864, "postDate": "2023-08-05T09:00:23.327Z", "content": "Thank you.
", "rawMarkdown": "Thank you.", "votes": 1 } ] } ] }, { "id": 2378586, "postDate": "2023-08-07T17:37:01.927Z", "content": "can i use jupiter notebook for this competition?
", "rawMarkdown": "can i use jupiter notebook for this competition?", "votes": -4, "replies": [ { "id": 2393292, "postDate": "2023-08-16T08:18:07.157Z", "content": "yes you can use Jupyter notebook as well as Google Collaboratory. And you can also use Pycharm maybe.
", "rawMarkdown": "yes you can use Jupyter notebook as well as Google Collaboratory. And you can also use Pycharm maybe." } ] }, { "id": 2440758, "postDate": "2023-09-15T17:55:44.217Z", "content": "@jeffrudie as far as I know, model code runs in no internet environment. So, can't we use any library for inference? and Is it ok to include segmentation model?(for extracting each organ in test)
", "rawMarkdown": "@jeffrudie as far as I know, model code runs in no internet environment. So, can't we use any library for inference? and Is it ok to include segmentation model?(for extracting each organ in test)", "replies": [ { "id": 2440766, "postDate": "2023-09-15T18:01:48.987Z", "content": "I believe you can upload your own segmentation model rather than a model from online
", "rawMarkdown": "I believe you can upload your own segmentation model rather than a model from online" } ] }, { "id": 2412967, "postDate": "2023-08-28T15:53:10.263Z", "content": "\nHi Jeff,
\nI am joining a bit late and it is difficult to catch up on everything posted so far. Can you tell me what the minimum number of slices in a series is in the test data set. I can see that in the train set there are three series with only one slice and the rest have at least forty slices. Also, are all series in the test set available at once in a particular directory or do they get available one at a time for inference?
", "rawMarkdown": "@jeffrudie \n\nHi Jeff,\n\nI am joining a bit late and it is difficult to catch up on everything posted so far. Can you tell me what the minimum number of slices in a series is in the test data set. I can see that in the train set there are three series with only one slice and the rest have at least forty slices. Also, are all series in the test set available at once in a particular directory or do they get available one at a time for inference?", "replies": [ { "id": 2413203, "postDate": "2023-08-28T18:00:35.483Z", "content": "The full testing set is hidden and not shared per Kaggle rules but solutions are applied to full testing set when submitted for inference. The actual testing set is similar to the provided training data as far as number of series/slices.
", "rawMarkdown": "The full testing set is hidden and not shared per Kaggle rules but solutions are applied to full testing set when submitted for inference. The actual testing set is similar to the provided training data as far as number of series/slices." } ] }, { "id": 2412910, "postDate": "2023-08-28T15:21:21.497Z", "content": "The link provided for AAST injury scales is not working. I think https://www.aast.org/resources-detail/injury-scoring-scale is worth checking if you are interested in these scales.
", "rawMarkdown": "The link provided for AAST injury scales is not working. I think [https://www.aast.org/resources-detail/injury-scoring-scale] (url) is worth checking if you are interested in these scales." }, { "id": 2394321, "postDate": "2023-08-16T20:32:38.233Z", "content": "What is meant by extravasation healthy (no injury)? That extravasation is seen but not due to injury? I haven't come across any healthy extravasation. Would it be possible to clarify this?
", "rawMarkdown": "What is meant by extravasation healthy (no injury)? That extravasation is seen but not due to injury? I haven't come across any healthy extravasation. Would it be possible to clarify this?", "replies": [ { "id": 2394336, "postDate": "2023-08-16T20:40:39.070Z", "content": "It means there is no extravasation.
", "rawMarkdown": "It means there is no extravasation." } ] }, { "id": 2382111, "postDate": "2023-08-09T15:19:22.163Z", "content": "the train.csv dataset has binary variables. what do 0 and 1 mean? is 0 = No and 1 = yes?
", "rawMarkdown": "the train.csv dataset has binary variables. what do 0 and 1 mean? is 0 = No and 1 = yes?", "replies": [ { "id": 2382163, "postDate": "2023-08-09T15:54:09.393Z", "rawMarkdown": "", "isDeleted": true }, { "id": 2387646, "postDate": "2023-08-12T20:52:19.920Z", "content": "Yes, this is called one hot encoding.
", "rawMarkdown": "Yes, this is called one hot encoding.", "votes": 1 } ] }, { "id": 2378374, "postDate": "2023-08-07T15:55:30.997Z", "content": "This an inspiring and a great task, Based on some of the discussion posts thank you for your to clarification on the the three types of labels provided in the challenge the problem am having the data set are to large is any ideas?. can i work on just the train data set
", "rawMarkdown": "This an inspiring and a great task, Based on some of the discussion posts thank you for your to clarification on the the three types of labels provided in the challenge the problem am having the data set are to large is any ideas?. can i work on just the train data set " }, { "id": 2382552, "postDate": "2023-08-09T20:23:41.457Z", "rawMarkdown": "", "votes": 1, "isDeleted": true }, { "id": 2374478, "postDate": "2023-08-05T02:58:39.477Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 2374489, "postDate": "2023-08-05T03:13:07.383Z", "content": "yes, segmentation folder is created by TotalSegmentator tool. You can use this tool to create your own mask, but the competition requires no internet, i'm trying to install it on kaggle without internet
", "rawMarkdown": "yes, segmentation folder is created by TotalSegmentator tool. You can use this tool to create your own mask, but the competition requires no internet, i'm trying to install it on kaggle without internet", "votes": -1, "replies": [ { "id": 2374542, "postDate": "2023-08-05T04:47:13.970Z", "rawMarkdown": "", "isDeleted": true } ] } ] }, { "id": 2413127, "postDate": "2023-08-28T17:21:27.007Z", "content": "Thank you for the detailed description
", "rawMarkdown": "Thank you for the detailed description" }, { "id": 2409336, "postDate": "2023-08-26T07:33:50.657Z", "content": "Thanks for clarifying
", "rawMarkdown": "Thanks for clarifying" }, { "id": 2390566, "postDate": "2023-08-14T16:19:57.843Z", "content": "Thank you for the detailed description
", "rawMarkdown": "Thank you for the detailed description" }, { "id": 2383549, "postDate": "2023-08-10T12:40:33.463Z", "content": "thanks for clarifying
", "rawMarkdown": "thanks for clarifying\n" }, { "id": 2400438, "postDate": "2023-08-21T04:35:22.243Z", "content": "Thank you so much.
", "rawMarkdown": "Thank you so much.", "isDeleted": true } ], "comments": [ { "id": 2404893, "author_name": "Yuri Kreinin", "author_url": "", "post_date": "2023-08-23T15:06:02.070000", "content": "@jeffrudie,
\n> Please be aware that the NIFTI files and DICOM files are not in the same orientation. Use the NIFTI header information along with DICOM metadata to determine the appropriate orientation such that the DICOM images and segmentation match.
In extravasation injuries are you considering PSEUDOANEURYSM or UBO (Unidentified Bright Objects) as EXTRAVASATION?
", "votes": 4, "replies": [ { "id": 2394187, "author_name": "Dipak Deshmukh", "author_url": "", "post_date": "2023-08-16T18:41:53.550000", "content": "both pseudoaneurysms and unidentified bright objects could potentially involve extravasation in some cases
", "votes": 2, "replies": [] }, { "id": 2394197, "author_name": "Dipak Deshmukh", "author_url": "", "post_date": "2023-08-16T18:44:17.800000", "content": "Pseudoaneurysm: This is a type of vascular injury that can occur due to trauma, medical procedures, or other causes. It involves a small hole in the wall of an artery or vein, leading to the formation of a pulsating sac filled with blood. Pseudoaneurysms can result in extravasation of blood into surrounding tissues, leading to swelling, pain, and potential complications. It is considered a form of vascular injury and can indeed be associated with extravasation.
\nUnidentified Bright Objects (UBO): This term is not widely recognized in the medical community , However, in medical imaging, \"bright objects\" typically refer to abnormal areas that appear brighter than the surrounding tissue. These can be indicative of various conditions, such as cysts, tumors, or other abnormalities. Whether UBOs are considered a form of extravasation would depend on the specific context in which the term is being used and the underlying cause of the bright appearance on the imaging.
", "votes": 4, "replies": [] } ] }, { "id": 2392293, "author_name": "RihanPiggy", "author_url": "", "post_date": "2023-08-15T15:23:19.277000", "content": "@jeffrudie
\nHello, jeffrudie. Thank you for your clarification.
\nI have a question about image level label. I took a look into the image level data and found that not all of the series_id of a certain patient are labeled.
\nFor example, patiend_id 43 has a extravasation injury, the correspond series_id are 24055 and 36714. In the image level data, only instance_id of 24055 are given, without the information about 36714. Does this mean no extravasation injury is found in 36714. or else is this simply means lack of annotation?
And what about the annotation in bowel?
\nHope for your reply
Extravasation Injuries are often confirmed on Sequential Scans. So, in this case, you will not only need to consider Series A but A and B in unison to figure out if Extravasation is present. So probably the Extravasation is seen on instance_id of 24055. That is how we radiologists report extravasation.
", "votes": 4, "replies": [ { "id": 2392457, "author_name": "RihanPiggy", "author_url": "", "post_date": "2023-08-15T17:27:34.190000", "content": "Hi, thanks for your reply!
\nI don't have the domain knowledge and it is really helpful.
\nIn the example above, given Series A(24055) and Series B(36714), to figure out extravasation, radiologists look into both of the Series. In this case, extravasation is only seen on series_id 24055, does this mean extravasation is not seen in 36714?
And what about the bowel injury, do radiologist confirm bowel injury on Sequential Scans?
", "votes": 4, "replies": [] }, { "id": 2394364, "author_name": "Dr. Datta (AIIMS Delhi)", "author_url": "", "post_date": "2023-08-16T21:20:48.990000", "content": "Extravasations need sequential scans. Bowel injury doesn't.
", "votes": 4, "replies": [] }, { "id": 2427277, "author_name": "Gyula Maloveczky4", "author_url": "", "post_date": "2023-09-07T06:25:28.173000", "content": "Oh wow, very helpful!
\nAre sequential scans used otherwise? Or only yo check extravasation injuries?
In competition rules, there is one rule which is no access internet. So can we use totalsegmantator model for segmantation ?
", "votes": 1, "replies": [ { "id": 2378524, "author_name": "JeffRudie", "author_url": "", "post_date": "2023-08-07T17:11:21.313000", "content": "You can train your own segmentation model with the total segmentator dataset and/or the data we have provided with segmentations.
", "votes": 3, "replies": [ { "id": 2392434, "author_name": "Dr. Datta (AIIMS Delhi)", "author_url": "", "post_date": "2023-08-15T17:18:33.950000", "content": "We can use external datasets to train as well?
", "votes": 0, "replies": [] }, { "id": 2408964, "author_name": "PC Jimmmy", "author_url": "", "post_date": "2023-08-26T00:35:44.473000", "content": "I see you did not get answer from the host. Kaggle general rule on training for this type of competition is that external are ok if they are freely available to all - if you can find it on the internet and download without fee or subscription than its ok.
", "votes": 1, "replies": [] } ] } ] }, { "id": 2369718, "author_name": "Vietnamese-NHNAM", "author_url": "", "post_date": "2023-08-02T01:36:24.053000", "content": "Thanks. You posted very useful information. But can you talk about train.csv. Labels represent only the patient level, each individual image has no label for it.
", "votes": 1, "replies": [ { "id": 2370027, "author_name": "JeffRudie", "author_url": "", "post_date": "2023-08-02T06:45:05.983000", "content": "Yes labels in the train.csv are only provided at the patient/study level. Some patients may have two scans/series, which are different contrast phases (think of them as timepoints/4th dimension). Injuries may be present in one or both of the different series/contrast phases.
\nImage/series level labels were only created for active extravasation and bowel injuries given the variable appearance/location of these types of injuries. Image level labels were not generated for other organ injuries (liver, spleen, kidneys) given the time/effort to create the image level labels and since an organ segmentation models should provide information to localize these organs and help narrow focus to detecting/classify injuries.
Is a segmentation model required for submission as mentioned in the comment above? Is it not simply about building a model and providing predictions for the three DCM files in the provided test images folder, rather than predictions for these three DCM files (2D images)?
", "votes": 0, "replies": [] }, { "id": 2436688, "author_name": "JeffRudie", "author_url": "", "post_date": "2023-09-13T17:53:04.787000", "content": "A segmentation model is not required. Only a csv file with predictions for study level labels are required. However, it's likely that incorporating a segmentation model will help improve performance.
", "votes": 0, "replies": [] }, { "id": 2440366, "author_name": "Min-soo, Kim", "author_url": "", "post_date": "2023-09-15T13:07:52.593000", "content": "\nI believe there could be errors in constructing a segmentation model because the entire DICOM dataset is labeled by patient_id. The following situations may result in discrepancies:
\nTo address these issues, it seems inevitable that we must assess the degree of predicted damage per DICOM and input labels accordingly. However, we cannot directly use the labels of 'healthy', 'low', and 'high' assigned at the patient level, nor can we label them ourselves because we are not medical professionals to make such judgments.
", "votes": 0, "replies": [] }, { "id": 2440713, "author_name": "JeffRudie", "author_url": "", "post_date": "2023-09-15T17:15:33.120000", "content": "The classification for organ damage is a single value for the entire patient. If a single image shows high grade injury for that organ then the label is high grade regardless of whether other slices are low grade or normal. A segmentation model could help narrow where your search to determine if a particular organ is injured.
", "votes": 0, "replies": [] }, { "id": 2442341, "author_name": "SneakyWave12", "author_url": "", "post_date": "2023-09-16T22:57:07.973000", "content": "but during test time , we dosent have semgnation data and involving semgnation to tested images will probabaly couse to exceed the time , so how its can be done ?
", "votes": 1, "replies": [] } ] } ] }, { "id": 2374299, "author_name": "Ada Wong", "author_url": "", "post_date": "2023-08-04T20:37:20.383000", "content": "Could we access the segmentator paper as it seems to be behind a paywall.
\nhttps://pubs.rsna.org/doi/full/10.1148/ryai.230024
https://arxiv.org/abs/2208.05868
", "votes": 9, "replies": [ { "id": 2374327, "author_name": "Ada Wong", "author_url": "", "post_date": "2023-08-04T21:27:11.797000", "content": "Thanks that's great!
", "votes": 0, "replies": [] }, { "id": 2374864, "author_name": "Antonio Félix", "author_url": "", "post_date": "2023-08-05T09:00:23.327000", "content": "Thank you.
", "votes": 1, "replies": [] } ] } ] }, { "id": 2378586, "author_name": "John Mola", "author_url": "", "post_date": "2023-08-07T17:37:01.927000", "content": "can i use jupiter notebook for this competition?
", "votes": -4, "replies": [ { "id": 2393292, "author_name": "ZARAKoo", "author_url": "", "post_date": "2023-08-16T08:18:07.157000", "content": "yes you can use Jupyter notebook as well as Google Collaboratory. And you can also use Pycharm maybe.
", "votes": 0, "replies": [] } ] }, { "id": 2440758, "author_name": "Seo Jong Heon", "author_url": "", "post_date": "2023-09-15T17:55:44.217000", "content": "@jeffrudie as far as I know, model code runs in no internet environment. So, can't we use any library for inference? and Is it ok to include segmentation model?(for extracting each organ in test)
", "votes": 0, "replies": [ { "id": 2440766, "author_name": "JeffRudie", "author_url": "", "post_date": "2023-09-15T18:01:48.987000", "content": "I believe you can upload your own segmentation model rather than a model from online
", "votes": 0, "replies": [] } ] }, { "id": 2412967, "author_name": "KRKirov", "author_url": "", "post_date": "2023-08-28T15:53:10.263000", "content": "\nHi Jeff,
\nI am joining a bit late and it is difficult to catch up on everything posted so far. Can you tell me what the minimum number of slices in a series is in the test data set. I can see that in the train set there are three series with only one slice and the rest have at least forty slices. Also, are all series in the test set available at once in a particular directory or do they get available one at a time for inference?
", "votes": 0, "replies": [ { "id": 2413203, "author_name": "JeffRudie", "author_url": "", "post_date": "2023-08-28T18:00:35.483000", "content": "The full testing set is hidden and not shared per Kaggle rules but solutions are applied to full testing set when submitted for inference. The actual testing set is similar to the provided training data as far as number of series/slices.
", "votes": 0, "replies": [] } ] }, { "id": 2412910, "author_name": "KRKirov", "author_url": "", "post_date": "2023-08-28T15:21:21.497000", "content": "The link provided for AAST injury scales is not working. I think https://www.aast.org/resources-detail/injury-scoring-scale is worth checking if you are interested in these scales.
", "votes": 0, "replies": [] }, { "id": 2394321, "author_name": "Jean Jacques Rousseau", "author_url": "", "post_date": "2023-08-16T20:32:38.233000", "content": "What is meant by extravasation healthy (no injury)? That extravasation is seen but not due to injury? I haven't come across any healthy extravasation. Would it be possible to clarify this?
", "votes": 0, "replies": [ { "id": 2394336, "author_name": "Sohier Dane", "author_url": "", "post_date": "2023-08-16T20:40:39.070000", "content": "It means there is no extravasation.
", "votes": 0, "replies": [] } ] }, { "id": 2382111, "author_name": "Muhsin", "author_url": "", "post_date": "2023-08-09T15:19:22.163000", "content": "the train.csv dataset has binary variables. what do 0 and 1 mean? is 0 = No and 1 = yes?
", "votes": 0, "replies": [ { "id": 2382163, "author_name": "", "author_url": "", "post_date": "2023-08-09T15:54:09.393000", "content": "", "votes": 0, "replies": [] }, { "id": 2387646, "author_name": "Eugene", "author_url": "", "post_date": "2023-08-12T20:52:19.920000", "content": "Yes, this is called one hot encoding.
", "votes": 1, "replies": [] } ] }, { "id": 2378374, "author_name": "Hegxy007", "author_url": "", "post_date": "2023-08-07T15:55:30.997000", "content": "This an inspiring and a great task, Based on some of the discussion posts thank you for your to clarification on the the three types of labels provided in the challenge the problem am having the data set are to large is any ideas?. can i work on just the train data set
", "votes": 0, "replies": [] }, { "id": 2382552, "author_name": "", "author_url": "", "post_date": "2023-08-09T20:23:41.457000", "content": "", "votes": 1, "replies": [] }, { "id": 2374478, "author_name": "", "author_url": "", "post_date": "2023-08-05T02:58:39.477000", "content": "", "votes": 0, "replies": [ { "id": 2374489, "author_name": "Vietnamese-NHNAM", "author_url": "", "post_date": "2023-08-05T03:13:07.383000", "content": "yes, segmentation folder is created by TotalSegmentator tool. You can use this tool to create your own mask, but the competition requires no internet, i'm trying to install it on kaggle without internet
", "votes": -1, "replies": [ { "id": 2374542, "author_name": "", "author_url": "", "post_date": "2023-08-05T04:47:13.970000", "content": "", "votes": 0, "replies": [] } ] } ] }, { "id": 2413127, "author_name": "kakshi0anon", "author_url": "", "post_date": "2023-08-28T17:21:27.007000", "content": "Thank you for the detailed description
", "votes": 0, "replies": [] }, { "id": 2409336, "author_name": "Flof", "author_url": "", "post_date": "2023-08-26T07:33:50.657000", "content": "Thanks for clarifying
", "votes": 0, "replies": [] }, { "id": 2390566, "author_name": "Daryl Fung", "author_url": "", "post_date": "2023-08-14T16:19:57.843000", "content": "Thank you for the detailed description
", "votes": 0, "replies": [] }, { "id": 2383549, "author_name": "martial taga", "author_url": "", "post_date": "2023-08-10T12:40:33.463000", "content": "thanks for clarifying
", "votes": 0, "replies": [] }, { "id": 2400438, "author_name": "", "author_url": "", "post_date": "2023-08-21T04:35:22.243000", "content": "Thank you so much.
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2369506": "Based on some of the discussion posts we wanted to clarify the three types of labels provided in the challenge and give more reasoning and clinical context.\n\n1.\tStudy level labels:\na.\tThere are study level labels for bowel, extravasation, liver, spleen, and kidneys. \nb.\tFor bowel and extravasation, the study level label is either injury or no injury. \nc.\tFor liver, spleen and kidneys, there are three classes: no injury, low-grade injury or high-grade injury. Low-grade injuries refer to AAST injury grades I, II and III and high-grade injuries refer to grades IV and V. The AAST injury scales are different for different organs will have different appearances- https://radiopaedia.org/articles/aast-injury-scoring-scales. The reasoning for this separation is that typically grade IV and V injuries require surgical intervention while grades I, II and III do not. \nd.\tThere is also a study level label for “any_injury” which is positive if any of the organs are injured.\ne.\tThe study level labels are what we are asking for predictions on the test cases.\n\n2.\tImage level labels:\na.\tThese are provided only for extravasation and bowel injury. This is because these types of injures can have a more variable appearance and anatomic location. Extravasation refers to active bleeding, which can be seen anywhere in the body. including within organs and in different soft tissues. Bowel injury includes injuries to the mesentery, which is what attaches the intestines to the abdominal wall. \n\n3.\tSegmentations: \na.\tIn order to provide more anatomical context to where the injuries will will be present, we have provided voxelwise segmentations on a subset of 206 training cases that are enriched for the presence of significant injuries. \nb.\tThese segmentations were manually revised by expert radiologists after being generated from a model trained with 3d_fullres nnU-Net on the total segmentator dataset focusing on the organs being evaluated (liver, spleen, left kidney, right kidney and bowel)\nc.\tWe modified the publicly available total segmentator segmentation labels to follow our schema just using labels for liver, spleen, left kidney, right kidney, and bowel (which combined esophagus, stomach, duodenum, small bowel and colon)\nd.\tThe total segmentator paper is here: https://pubs.rsna.org/doi/10.1148/ryai.230024\ne.\tThe total segmentator dataset is here: https://zenodo.org/record/6802614\nf.\tThe modified total segmentator segmentations used to train the initial model (in *.nii.gz format) can be downloaded here: https://drive.google.com/file/d/1zoSFVlls-j6IUFWuiASFrb1OqVZO37yR/view?usp=sharing\ng.\tThe segmentation labels are as follows:\ni.\t0 = background\nii.\t1 = liver\niii.\t2 = spleen\niv.\t3 = left kidney\nv.\t4 = right kidney\nvi.\t5 = bowel \nh.\tWhile different labels are provided for left and right kidney, we are asking for the prediction of injury in either kidney. If the left kidney has a low-grade injury while the right kidney has a high-grade injury, the correct prediction would be high-grade. \ni.\tPlease be aware that the NIFTI files and DICOM files are not in the same orientation. Use the NIFTI header information along with DICOM metadata to determine the appropriate orientation such that the DICOM images and segmentation match.\n\nJeff Rudie, MD PhD, on behalf of the organizing committee\n", "2404893": "@jeffrudie, \n*> Please be aware that the NIFTI files and DICOM files are not in the same orientation. Use the NIFTI header information along with DICOM metadata to determine the appropriate orientation such that the DICOM images and segmentation match.*\n1. Can you, please, elaborate what NIFTI and DICOM metadata should be taking in consideration in order to align properly the segmentation with CT scan?\n2. Because, the alignment of segmentation is not a goal of this competition, I think that it would be nice and efficient if properly aligned segmentation (or, at least, code example verified by organizers) was available in order to allow competitors to focus on main challenges of the trauma detection", "2392427": "In extravasation injuries are you considering PSEUDOANEURYSM or UBO (Unidentified Bright Objects) as EXTRAVASATION?\n", "2392293": "@jeffrudie \nHello, jeffrudie. Thank you for your clarification.\nI have a question about image level label. I took a look into the image level data and found that not all of the series_id of a certain patient are labeled. \nFor example, patiend_id 43 has a extravasation injury, the correspond series_id are 24055 and 36714. In the image level data, only instance_id of 24055 are given, without the information about 36714. Does this mean no extravasation injury is found in 36714. or else is this simply means lack of annotation?\n\nAnd what about the annotation in bowel? \nHope for your reply", "2377049": "In competition rules, there is one rule which is no access internet. So can we use totalsegmantator model for segmantation ?", "2369718": "Thanks. You posted very useful information. But can you talk about train.csv. Labels represent only the patient level, each individual image has no label for it.", "2374299": "Could we access the segmentator paper as it seems to be behind a paywall.\nhttps://pubs.rsna.org/doi/full/10.1148/ryai.230024", "2378586": "can i use jupiter notebook for this competition?", "2440758": "@jeffrudie as far as I know, model code runs in no internet environment. So, can't we use any library for inference? and Is it ok to include segmentation model?(for extracting each organ in test)", "2412967": "@jeffrudie \n\nHi Jeff,\n\nI am joining a bit late and it is difficult to catch up on everything posted so far. Can you tell me what the minimum number of slices in a series is in the test data set. I can see that in the train set there are three series with only one slice and the rest have at least forty slices. Also, are all series in the test set available at once in a particular directory or do they get available one at a time for inference?", "2412910": "The link provided for AAST injury scales is not working. I think [https://www.aast.org/resources-detail/injury-scoring-scale] (url) is worth checking if you are interested in these scales.", "2394321": "What is meant by extravasation healthy (no injury)? That extravasation is seen but not due to injury? I haven't come across any healthy extravasation. Would it be possible to clarify this?", "2382111": "the train.csv dataset has binary variables. what do 0 and 1 mean? is 0 = No and 1 = yes?", "2378374": "This an inspiring and a great task, Based on some of the discussion posts thank you for your to clarification on the the three types of labels provided in the challenge the problem am having the data set are to large is any ideas?. can i work on just the train data set ", "2382552": "", "2374478": "", "2413127": "Thank you for the detailed description", "2409336": "Thanks for clarifying", "2390566": "Thank you for the detailed description", "2383549": "thanks for clarifying\n", "2400438": "Thank you so much." } }