\n
![\"pprojection\"](\"https://i.postimg.cc/mhWZ4C69/projection.jpg\")
One thing that i observed in this competition (image from this notebook):
\n
If you take a look at the upper image and the CVC with purple and red - both are denoted as normal. On the second image brown ( = purple on the previous) and purple ( = red on the previous) are denoted as borderline, however they remained on the exact same position… there is only one thing that changed (ant thus changed the perception of the radiologist) is the projection!
\nOn the left side the circles show the area above the clavicle… to put it simple: the \"larger\" this areal gets (by the same patient), the more deviates the projection from the perpendicular plane relative to chest and the more it changes the shape / correlation of lines with the organ shapes.
\nTo visualise it, blue is the first projection, red is the second and the green arrow show the \"repositioning of the x-ray generator\".
\n![\"projection](\"https://i.postimg.cc/bNHfdsCq/projectionchange.jpg\")
Image modified, from here
\nThis should be definitely exploited to assess the normal / borderline and borderline / abnormal CVC positions… however this phenomenon is the same of the ETTs to ( = in one projection normal, in the other borderline).
", "messages": [ { "id": "1220725", "postDate": "02/28/2021 09:46:46", "content": "One thing that i observed in this competition (image from this notebook):
\n
If you take a look at the upper image and the CVC with purple and red - both are denoted as normal. On the second image brown ( = purple on the previous) and purple ( = red on the previous) are denoted as borderline, however they remained on the exact same position… there is only one thing that changed (ant thus changed the perception of the radiologist) is the projection!
\nOn the left side the circles show the area above the clavicle… to put it simple: the \"larger\" this areal gets (by the same patient), the more deviates the projection from the perpendicular plane relative to chest and the more it changes the shape / correlation of lines with the organ shapes.
\nTo visualise it, blue is the first projection, red is the second and the green arrow show the \"repositioning of the x-ray generator\".
\n
Image modified, from here
\nThis should be definitely exploited to assess the normal / borderline and borderline / abnormal CVC positions… however this phenomenon is the same of the ETTs to ( = in one projection normal, in the other borderline).
", "rawMarkdown": "One thing that i observed in this competition (image from [this notebook](https://www.kaggle.com/tomohiroh/ranzcr-samepatient-images)):\n\n\nIf you take a look at the upper image and the CVC with purple and red - both are denoted as normal. On the second image brown ( = purple on the previous) and purple ( = red on the previous) are denoted as borderline, however they remained on the exact same position... there is only one thing that changed (ant thus changed the perception of the radiologist) is the **projection**! \n\nOn the left side the circles show the area above the clavicle... to put it simple: the \"larger\" this areal gets (by the same patient), the more deviates the projection from the perpendicular plane relative to chest and the more it changes the shape / correlation of lines with the organ shapes.\n\nTo visualise it, blue is the first projection, red is the second and the green arrow show the \"repositioning of the x-ray generator\".\n\n\nImage modified, from [here](https://www.researchgate.net/publication/340586359_Planning_and_coordination_of_the_radiological_response_to_the_coronavirus_disease_2019_COVID-19_pandemic_the_Singapore_experience/figures?lo=1)\n\nThis ~~could be might~~ should be definitely exploited to assess the normal / borderline and borderline / abnormal CVC positions... however this phenomenon is the same of the ETTs to ( = in one projection normal, in the other borderline).", "votes": null }, { "id": "1220739", "postDate": "02/28/2021 10:05:51", "content": "thank you, i would like to ask one important question as you are exploring and scrutinizing the dataset, \"is there any label noise? how much mislabeled data do we have in this dataset according to your estimation? asking this question because it is easy to make mistakes while doing medical image annotation, so i would like to know how much wrong annotated data we've got?\"
", "rawMarkdown": "thank you, i would like to ask one important question as you are exploring and scrutinizing the dataset, \"is there any label noise? how much mislabeled data do we have in this dataset according to your estimation? asking this question because it is easy to make mistakes while doing medical image annotation, so i would like to know how much wrong annotated data we've got?\"", "votes": null }, { "id": "1220795", "postDate": "02/28/2021 11:06:10", "content": "\n\n\nis there any label noise? how much mislabeled data do we have in this dataset according to your estimation?
\n
Mislabelled or inter- / intra-observer error? I remember having seen some threads about outliers that revealed some mislabeled lines like ETT --> CVC but that should be rather easy to solve…
\nMuch larger problem is the intra- inter-observer error that is clearly visible from the example above, and that is something that is very disturbing and could be only overcome if the labels are reviewed… but it would not help you to score better (since the test set is probably affected from the same intra- inter observer error.
", "rawMarkdown": "mobassir,\n> is there any label noise? how much mislabeled data do we have in this dataset according to your estimation?\n\nMislabelled or inter- / intra-observer error? I remember having seen some threads about outliers that revealed some mislabeled lines like ETT --> CVC but that should be rather easy to solve...\n\nMuch larger problem is the intra- inter-observer error that is clearly visible from the example above, and that is something that is very disturbing and could be only overcome if the labels are reviewed... but it would not help you to score better (since the test set is probably affected from the same intra- inter observer error.", "votes": null }, { "id": "1228060", "postDate": "03/06/2021 04:50:19", "content": "@sandorkonya - considering this and some other examples and would like to ask if you could comment on the difference (if any) when using Portable equipment or not. Also if the patient is supine or semi-erect if known. e.g. is it more likely to have projection differences for portable or not supine?
\nWould imagine there might be differences in image quality in original images but that may have changed in converting to jpg for the competition
\nThanks in advance for your views.
@something4kag ,
\nyou are completely right, the patient's position (semi-erect or supine) affects the projection of the silouettes of the organs to (in semi-erect the air in the stomach is usually visible and in supine not).
In my experience it is always better to use png or tif and i avoid using jpg if i can.
", "rawMarkdown": "something4kag ,\nyou are completely right, the patient's position (semi-erect or supine) affects the projection of the silouettes of the organs to (in semi-erect the air in the stomach is usually visible and in supine not).\n\nIn my experience it is always better to use png or tif and i avoid using jpg if i can.", "votes": null }, { "id": "1228654", "postDate": "03/06/2021 16:12:07", "content": "Interesting point, some of the patients have a lot of X-rays. It looks like in this case it isn't actually multiple different X-rays but rather two X-rays at the same time from different angles. Do you have any hypothesis about why some specific patient ID's have 100+ X-rays in this dataset?
", "rawMarkdown": "Interesting point, some of the patients have a lot of X-rays. It looks like in this case it isn't actually multiple different X-rays but rather two X-rays at the same time from different angles. Do you have any hypothesis about why some specific patient ID's have 100+ X-rays in this dataset?", "votes": null }, { "id": "1228657", "postDate": "03/06/2021 16:12:54", "content": "Secondary question, do you think that if a patient has had issues with catheter placement in the past they are more likely to have problematic placements in the future?
", "rawMarkdown": "Secondary question, do you think that if a patient has had issues with catheter placement in the past they are more likely to have problematic placements in the future?", "votes": null }, { "id": "1228705", "postDate": "03/06/2021 16:59:32", "content": "@ryches ,
\nNo, they are indeed multiple x-rays at different time points (probably consecutive days) with some on the same day (after cvc replacement or ett reposition) but not at the same time.
\nRe second question: by difficult anatomy it may be possible that a CVC \"goes\" always in the false direction, so yes, it may be possible but this is something you can relatively easily test from the labels - or?
Based on the imagehash same patient in ChestX, the second image is duplicate of 00001836_119.png so follow up# 119 for patient 1836 that has over 130 images in that dataset. Have not found the duplicate of the first image, but does show how follow up# could be useful. Even age because the patient images start at 2 years earlier I and if aging means more likely difficult placements. It would seem likely abnormal would be followed by normal if corrected.
\nIf the test sets have same patients in train, then this seems problematic for this competition.
i have one question for the expert. i divide the images by patient. i find that e.g. ETT can become from normal to abnormal? is it possible that the Catheter and Line Position can be shifted by their own????
\nor are these just label noise???
", "rawMarkdown": "i have one question for the expert. i divide the images by patient. i find that e.g. ETT can become from normal to abnormal? is it possible that the Catheter and Line Position can be shifted by their own????\n\nor are these just label noise???", "votes": null }, { "id": "1238326", "postDate": "03/14/2021 22:06:38", "content": "The patients get washed, turned, get physiotherapy. The localisation of ETT & CVC may change during these activities.
\nIt can be inter- (or intra) observer disagreement/error to.
For ETT specifically, even head positioning can alter placement. Suppose an ETT is in the mid-lower trachea and the head is tilted up, if the patients head then tilts all the way down, the tube could move as far as normal to abnormal.
", "rawMarkdown": "For ETT specifically, even head positioning can alter placement. Suppose an ETT is in the mid-lower trachea and the head is tilted up, if the patients head then tilts all the way down, the tube could move as far as normal to abnormal.", "votes": null } ], "comments": [ { "id": 1220739, "author_name": "mobassir", "author_url": "", "post_date": "02/28/2021 10:05:51", "content": "thank you, i would like to ask one important question as you are exploring and scrutinizing the dataset, \"is there any label noise? how much mislabeled data do we have in this dataset according to your estimation? asking this question because it is easy to make mistakes while doing medical image annotation, so i would like to know how much wrong annotated data we've got?\"
", "votes": null, "replies": [ { "id": 1220795, "author_name": "sandorkonya", "author_url": "", "post_date": "02/28/2021 11:06:10", "content": "\n\n\nis there any label noise? how much mislabeled data do we have in this dataset according to your estimation?
\n
Mislabelled or inter- / intra-observer error? I remember having seen some threads about outliers that revealed some mislabeled lines like ETT --> CVC but that should be rather easy to solve…
\nMuch larger problem is the intra- inter-observer error that is clearly visible from the example above, and that is something that is very disturbing and could be only overcome if the labels are reviewed… but it would not help you to score better (since the test set is probably affected from the same intra- inter observer error.
", "votes": null, "replies": [] } ] }, { "id": 1228060, "author_name": "something4kag", "author_url": "", "post_date": "03/06/2021 04:50:19", "content": "@sandorkonya - considering this and some other examples and would like to ask if you could comment on the difference (if any) when using Portable equipment or not. Also if the patient is supine or semi-erect if known. e.g. is it more likely to have projection differences for portable or not supine?
\nWould imagine there might be differences in image quality in original images but that may have changed in converting to jpg for the competition
\nThanks in advance for your views.
@something4kag ,
\nyou are completely right, the patient's position (semi-erect or supine) affects the projection of the silouettes of the organs to (in semi-erect the air in the stomach is usually visible and in supine not).
In my experience it is always better to use png or tif and i avoid using jpg if i can.
", "votes": null, "replies": [] } ] }, { "id": 1228654, "author_name": "ryches", "author_url": "", "post_date": "03/06/2021 16:12:07", "content": "Interesting point, some of the patients have a lot of X-rays. It looks like in this case it isn't actually multiple different X-rays but rather two X-rays at the same time from different angles. Do you have any hypothesis about why some specific patient ID's have 100+ X-rays in this dataset?
", "votes": null, "replies": [ { "id": 1228657, "author_name": "ryches", "author_url": "", "post_date": "03/06/2021 16:12:54", "content": "Secondary question, do you think that if a patient has had issues with catheter placement in the past they are more likely to have problematic placements in the future?
", "votes": null, "replies": [] }, { "id": 1228705, "author_name": "sandorkonya", "author_url": "", "post_date": "03/06/2021 16:59:32", "content": "@ryches ,
\nNo, they are indeed multiple x-rays at different time points (probably consecutive days) with some on the same day (after cvc replacement or ett reposition) but not at the same time.
\nRe second question: by difficult anatomy it may be possible that a CVC \"goes\" always in the false direction, so yes, it may be possible but this is something you can relatively easily test from the labels - or?
Based on the imagehash same patient in ChestX, the second image is duplicate of 00001836_119.png so follow up# 119 for patient 1836 that has over 130 images in that dataset. Have not found the duplicate of the first image, but does show how follow up# could be useful. Even age because the patient images start at 2 years earlier I and if aging means more likely difficult placements. It would seem likely abnormal would be followed by normal if corrected.
\nIf the test sets have same patients in train, then this seems problematic for this competition.
i have one question for the expert. i divide the images by patient. i find that e.g. ETT can become from normal to abnormal? is it possible that the Catheter and Line Position can be shifted by their own????
\nor are these just label noise???
", "votes": null, "replies": [ { "id": 1238326, "author_name": "sandorkonya", "author_url": "", "post_date": "03/14/2021 22:06:38", "content": "The patients get washed, turned, get physiotherapy. The localisation of ETT & CVC may change during these activities.
\nIt can be inter- (or intra) observer disagreement/error to.
For ETT specifically, even head positioning can alter placement. Suppose an ETT is in the mid-lower trachea and the head is tilted up, if the patients head then tilts all the way down, the tube could move as far as normal to abnormal.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1220725": "One thing that i observed in this competition (image from [this notebook](https://www.kaggle.com/tomohiroh/ranzcr-samepatient-images)):\n\n\nIf you take a look at the upper image and the CVC with purple and red - both are denoted as normal. On the second image brown ( = purple on the previous) and purple ( = red on the previous) are denoted as borderline, however they remained on the exact same position... there is only one thing that changed (ant thus changed the perception of the radiologist) is the **projection**! \n\nOn the left side the circles show the area above the clavicle... to put it simple: the \"larger\" this areal gets (by the same patient), the more deviates the projection from the perpendicular plane relative to chest and the more it changes the shape / correlation of lines with the organ shapes.\n\nTo visualise it, blue is the first projection, red is the second and the green arrow show the \"repositioning of the x-ray generator\".\n\n\nImage modified, from [here](https://www.researchgate.net/publication/340586359_Planning_and_coordination_of_the_radiological_response_to_the_coronavirus_disease_2019_COVID-19_pandemic_the_Singapore_experience/figures?lo=1)\n\nThis ~~could be might~~ should be definitely exploited to assess the normal / borderline and borderline / abnormal CVC positions... however this phenomenon is the same of the ETTs to ( = in one projection normal, in the other borderline).", "1220739": "thank you, i would like to ask one important question as you are exploring and scrutinizing the dataset, \"is there any label noise? how much mislabeled data do we have in this dataset according to your estimation? asking this question because it is easy to make mistakes while doing medical image annotation, so i would like to know how much wrong annotated data we've got?\"", "1220795": "mobassir,\n> is there any label noise? how much mislabeled data do we have in this dataset according to your estimation?\n\nMislabelled or inter- / intra-observer error? I remember having seen some threads about outliers that revealed some mislabeled lines like ETT --> CVC but that should be rather easy to solve...\n\nMuch larger problem is the intra- inter-observer error that is clearly visible from the example above, and that is something that is very disturbing and could be only overcome if the labels are reviewed... but it would not help you to score better (since the test set is probably affected from the same intra- inter observer error.", "1228060": "sandorkonya - considering this and some other examples and would like to ask if you could comment on the difference (if any) when using Portable equipment or not. Also if the patient is supine or semi-erect if known. e.g. is it more likely to have projection differences for portable or not supine? \nWould imagine there might be differences in image quality in original images but that may have changed in converting to jpg for the competition\nThanks in advance for your views.", "1228456": "something4kag ,\nyou are completely right, the patient's position (semi-erect or supine) affects the projection of the silouettes of the organs to (in semi-erect the air in the stomach is usually visible and in supine not).\n\nIn my experience it is always better to use png or tif and i avoid using jpg if i can.", "1228654": "Interesting point, some of the patients have a lot of X-rays. It looks like in this case it isn't actually multiple different X-rays but rather two X-rays at the same time from different angles. Do you have any hypothesis about why some specific patient ID's have 100+ X-rays in this dataset?", "1228657": "Secondary question, do you think that if a patient has had issues with catheter placement in the past they are more likely to have problematic placements in the future?", "1228705": "ryches ,\nNo, they are indeed multiple x-rays at different time points (probably consecutive days) with some on the same day (after cvc replacement or ett reposition) but not at the same time.\nRe second question: by difficult anatomy it may be possible that a CVC \"goes\" always in the false direction, so yes, it may be possible but this is something you can relatively easily test from the labels - or?", "1229144": "Based on the imagehash same patient in ChestX, the second image is duplicate of 00001836_119.png so follow up# 119 for patient 1836 that has over 130 images in that dataset. Have not found the duplicate of the first image, but does show how follow up# could be useful. Even age because the patient images start at 2 years earlier I and if aging means more likely difficult placements. It would seem likely abnormal would be followed by normal if corrected. \nIf the test sets have same patients in train, then this seems problematic for this competition.", "1238252": "i have one question for the expert. i divide the images by patient. i find that e.g. ETT can become from normal to abnormal? is it possible that the Catheter and Line Position can be shifted by their own????\n\nor are these just label noise???", "1238326": "The patients get washed, turned, get physiotherapy. The localisation of ETT & CVC may change during these activities. \nIt can be inter- (or intra) observer disagreement/error to.", "1238430": "For ETT specifically, even head positioning can alter placement. Suppose an ETT is in the mid-lower trachea and the head is tilted up, if the patients head then tilts all the way down, the tube could move as far as normal to abnormal." }, "source": "meta" }