According to the train.csv it seems we only have labels per patient, so when it comes to the slices (instance numbers) of CT scans, the labels become weak. We can't tell which slice (instance number) is positive or negative, and that's causing some uncertainty.
Now, the situation may get a bit trickier because we have multiple scans (series) per patient. If the labels vary between these scans, it only makes the label even weaker.
\nimage_level_labels.csv. It does provide some helpful info about the labels of series and instance numbers, but all labels can't be extracted from here.
So, I was wondering if we could get some clarification regarding the labels.
", "messages": [ { "id": 2367272, "postDate": "2023-07-31T13:14:11.640Z", "content": "According to the train.csv it seems we only have labels per patient, so when it comes to the slices (instance numbers) of CT scans, the labels become weak. We can't tell which slice (instance number) is positive or negative, and that's causing some uncertainty.
Now, the situation may get a bit trickier because we have multiple scans (series) per patient. If the labels vary between these scans, it only makes the label even weaker.
\nimage_level_labels.csv. It does provide some helpful info about the labels of series and instance numbers, but all labels can't be extracted from here.
So, I was wondering if we could get some clarification regarding the labels.
", "rawMarkdown": "According to the `train.csv` it seems we only have labels per patient, so when it comes to the slices (instance numbers) of CT scans, the labels become weak. We can't tell which slice (instance number) is positive or negative, and that's causing some uncertainty.\n\nNow, the situation may get a bit trickier because we have multiple scans (series) per patient. If the labels vary between these scans, it only makes the label even weaker. \n\n`image_level_labels.csv`. It does provide some helpful info about the labels of series and instance numbers, but all labels can't be extracted from here.\n\nSo, I was wondering if we could get some clarification regarding the labels.", "votes": 12 }, { "id": 2369511, "postDate": "2023-08-01T19:30:41.323Z", "content": "Please see this post for some clarification.
\nhttps://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/428538
I was thinking this as well. It seems the real work in this comp will be identifying which slices contain specific anatomy.
", "rawMarkdown": "I was thinking this as well. It seems the real work in this comp will be identifying which slices contain specific anatomy.", "votes": 1, "replies": [ { "id": 2367350, "postDate": "2023-07-31T14:05:34.700Z", "content": "Do you think we can select slices that are likely to have the trauma like using the middle slice in some applications.
", "rawMarkdown": "Do you think we can select slices that are likely to have the trauma like using the middle slice in some applications.", "replies": [ { "id": 2367364, "postDate": "2023-07-31T14:14:37.097Z", "content": "Each organ will be visualized in many slices. We know the Liver, Kidneys and Spleen are in the upper abdomen, and usually, the liver is the first abdominal organ seen when descending through slices. The liver extends down into the pelvis in some cases, but we're not likely to see kidneys down where the hips are.
\nWe know the liver is on the patient's right (viewer's left), the spleen is on the left .. and there's a kidney on each side. The kidneys are normally in the posterior (bottom) of the axial slices.
\nOn the other hand, the bowels are going to be visualized in slices throughout the abdomen and pelvis. Some slices will have all three organs and some bowel. Some slices will only have bowel.
", "rawMarkdown": "Each organ will be visualized in many slices. We know the Liver, Kidneys and Spleen are in the upper abdomen, and usually, the liver is the first abdominal organ seen when descending through slices. The liver extends down into the pelvis in some cases, but we're not likely to see kidneys down where the hips are.\n\nWe know the liver is on the patient's right (viewer's left), the spleen is on the left .. and there's a kidney on each side. The kidneys are normally in the posterior (bottom) of the axial slices.\n\nOn the other hand, the bowels are going to be visualized in slices throughout the abdomen and pelvis. Some slices will have all three organs and some bowel. Some slices will only have bowel." }, { "id": 2367416, "postDate": "2023-07-31T14:41:56.890Z", "content": "I'm not sure if choosing one slice that has many organ will help because in that slice we don't see all parts of an organ (kidney, liver etc) so even a slice contains many organs it is possible that the part of the organ is not injured at all.
", "rawMarkdown": "I'm not sure if choosing one slice that has many organ will help because in that slice we don't see all parts of an organ (kidney, liver etc) so even a slice contains many organs it is possible that the part of the organ is not injured at all." }, { "id": 2367431, "postDate": "2023-07-31T14:47:35.280Z", "content": "I think it will be critical to identify all slices that contain each organ.
", "rawMarkdown": "I think it will be critical to identify all slices that contain each organ." }, { "id": 2367481, "postDate": "2023-07-31T15:15:00.140Z", "content": "i think segmentation folder will help, it contain masks of organs in each slice. But don't know which label will be suitable for which organ. Everything is vague. But in condition that we can identify all slices that contain each organ, we still don't know this slice have.
\nWe still don't know if that slice has any pathology or not? Because the labels provided are at the patient level, there are no separate labels for each slice. I think the doctor actually looks at all the slices to judge the lesion but maybe some slices show the lesion and others don't.
\nIn this case, 3d model can help but it is complicated and accuracy have not been high so far
Scan level labels is already weak enough and 3D models is the only way but if a label is not visible on one scan then it is a problem.
", "rawMarkdown": "Scan level labels is already weak enough and 3D models is the only way but if a label is not visible on one scan then it is a problem.", "votes": 1, "replies": [ { "id": 2367316, "postDate": "2023-07-31T13:44:19.117Z", "content": "Yes, that is what I'm concerned with. Then we've to consider 4D models or multi inputs models
", "rawMarkdown": "Yes, that is what I'm concerned with. Then we've to consider 4D models or multi inputs models", "votes": 1, "replies": [ { "id": 2425911, "postDate": "2023-09-06T09:12:07.137Z", "content": "4D? Meaning aggregating 2 or more 3D data?
", "rawMarkdown": "4D? Meaning aggregating 2 or more 3D data?" } ] } ] }, { "id": 2381096, "postDate": "2023-08-09T01:52:39.137Z", "content": "I think we have to choose the slices. Medically speaking, certain slices may contain useful information.
", "rawMarkdown": "I think we have to choose the slices. Medically speaking, certain slices may contain useful information. \n" } ], "comments": [ { "id": 2369511, "author_name": "JeffRudie", "author_url": "", "post_date": "2023-08-01T19:30:41.323000", "content": "Please see this post for some clarification.
\nhttps://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/428538
I was thinking this as well. It seems the real work in this comp will be identifying which slices contain specific anatomy.
", "votes": 1, "replies": [ { "id": 2367350, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2023-07-31T14:05:34.700000", "content": "Do you think we can select slices that are likely to have the trauma like using the middle slice in some applications.
", "votes": 0, "replies": [ { "id": 2367364, "author_name": "David Roberts", "author_url": "", "post_date": "2023-07-31T14:14:37.097000", "content": "Each organ will be visualized in many slices. We know the Liver, Kidneys and Spleen are in the upper abdomen, and usually, the liver is the first abdominal organ seen when descending through slices. The liver extends down into the pelvis in some cases, but we're not likely to see kidneys down where the hips are.
\nWe know the liver is on the patient's right (viewer's left), the spleen is on the left .. and there's a kidney on each side. The kidneys are normally in the posterior (bottom) of the axial slices.
\nOn the other hand, the bowels are going to be visualized in slices throughout the abdomen and pelvis. Some slices will have all three organs and some bowel. Some slices will only have bowel.
", "votes": 0, "replies": [] }, { "id": 2367416, "author_name": "Awsaf", "author_url": "", "post_date": "2023-07-31T14:41:56.890000", "content": "I'm not sure if choosing one slice that has many organ will help because in that slice we don't see all parts of an organ (kidney, liver etc) so even a slice contains many organs it is possible that the part of the organ is not injured at all.
", "votes": 0, "replies": [] }, { "id": 2367431, "author_name": "David Roberts", "author_url": "", "post_date": "2023-07-31T14:47:35.280000", "content": "I think it will be critical to identify all slices that contain each organ.
", "votes": 0, "replies": [] }, { "id": 2367481, "author_name": "Vietnamese-NHNAM", "author_url": "", "post_date": "2023-07-31T15:15:00.140000", "content": "i think segmentation folder will help, it contain masks of organs in each slice. But don't know which label will be suitable for which organ. Everything is vague. But in condition that we can identify all slices that contain each organ, we still don't know this slice have.
\nWe still don't know if that slice has any pathology or not? Because the labels provided are at the patient level, there are no separate labels for each slice. I think the doctor actually looks at all the slices to judge the lesion but maybe some slices show the lesion and others don't.
\nIn this case, 3d model can help but it is complicated and accuracy have not been high so far
Scan level labels is already weak enough and 3D models is the only way but if a label is not visible on one scan then it is a problem.
", "votes": 1, "replies": [ { "id": 2367316, "author_name": "Awsaf", "author_url": "", "post_date": "2023-07-31T13:44:19.117000", "content": "Yes, that is what I'm concerned with. Then we've to consider 4D models or multi inputs models
", "votes": 1, "replies": [ { "id": 2425911, "author_name": "Gyula Maloveczky4", "author_url": "", "post_date": "2023-09-06T09:12:07.137000", "content": "4D? Meaning aggregating 2 or more 3D data?
", "votes": 0, "replies": [] } ] } ] }, { "id": 2381096, "author_name": "Nguyen Quang Hieu", "author_url": "", "post_date": "2023-08-09T01:52:39.137000", "content": "I think we have to choose the slices. Medically speaking, certain slices may contain useful information.
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2367272": "According to the `train.csv` it seems we only have labels per patient, so when it comes to the slices (instance numbers) of CT scans, the labels become weak. We can't tell which slice (instance number) is positive or negative, and that's causing some uncertainty.\n\nNow, the situation may get a bit trickier because we have multiple scans (series) per patient. If the labels vary between these scans, it only makes the label even weaker. \n\n`image_level_labels.csv`. It does provide some helpful info about the labels of series and instance numbers, but all labels can't be extracted from here.\n\nSo, I was wondering if we could get some clarification regarding the labels.", "2369511": "Please see this post for some clarification. \nhttps://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/428538", "2367326": "I was thinking this as well. It seems the real work in this comp will be identifying which slices contain specific anatomy.", "2367310": "Scan level labels is already weak enough and 3D models is the only way but if a label is not visible on one scan then it is a problem.", "2381096": "I think we have to choose the slices. Medically speaking, certain slices may contain useful information. \n" } }