{ "id": 534873, "title": "If you think you saw everything, check for flipped axials", "url": "/competitions/rsna-2024-lumbar-spine-degenerative-classification/discussion/534873", "author_name": "SSS", "post_date": "2024-09-19T00:54:52.302000", "votes": 23, "comment_count": 23, "views": 0, "content": "

Hi,

\n

I will just leave this cockroach here.

\n

\"\"

", "messages": [ { "id": 2992733, "postDate": "2024-09-19T00:54:52.303Z", "content": "

Hi,

\n

I will just leave this cockroach here.

\n

\"\"

", "rawMarkdown": "Hi,\n\nI will just leave this cockroach here.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6259210%2Fba6e16d91dbb4616f1e9568110bb715d%2FScreenshot%20from%202024-09-18%2020-54-30.png?generation=1726707280992832&alt=media)", "votes": 23 }, { "id": 2994435, "postDate": "2024-09-20T23:49:42.473Z", "content": "

this competition is \"relative\" easy if not for the annotation errors.
\nannotation errors includes:

\n\n

issue is: \"is the hidden test data the same?\"
\n(ior past RSNA competition, extra care and verification are taken to ensure hidden test data are correct. It is not known for this competition)

\n

my suggestion: one should do a probe to see if there are annotation error in hidden test data
\n(yes this can be done by doing some test on onsistency of the predicted points)

\n

or you can ask the host to clarify

", "rawMarkdown": "this competition is \"relative\" easy if not for the annotation errors.\nannotation errors includes:\n- inconsistency between L5 and S1 labeling within a condition\n- inconsistency level labeling within a study. if you project all labels of axial and sagittal views to common 3d world coord, they are not aliged\n- inconsistency level labeling within a series. left and right point of same level are not aliged\n- wrong point labels (e.g. 2 level point coincide, left/right point coincide, point becomes at corner of image, etc ...)\n- missing point labels\n- a couple of wrong images\n\nissue is: \"is the hidden test data the same?\"\n(ior past RSNA competition, extra care and verification are taken to ensure hidden test data are correct. It is not known for this competition)\n\nmy suggestion: one should do a probe to see if there are annotation error in hidden test data\n(yes this can be done by doing some test on onsistency of the predicted points)\n\nor you can ask the host to clarify\n", "votes": 6, "replies": [ { "id": 2994440, "postDate": "2024-09-21T00:26:59.577Z", "content": "

That's a good summary.

\n

We can add to that:

\n
    \n
  1. Few series have broken translation vector, e.g. broken IOP.
    2. \n
  2. inconsistency between L5 and S1 labeling within a condition - I don't know if that includes some of the levels start not from the trapezoid shaped sacrum but arbitrary higher. I consulted with few researchers and radiologists, showed few train examples, they confirmed that the L5/S1 level did start from the trapezoid sacrum (either host mislabled it or there was 6th vertebra which again mislabeled).
    3. \n
  3. Different sizes / px_spacing of the images within the same axial series
    4. \n
", "rawMarkdown": "That's a good summary.\n\nWe can add to that:\n 1. Few series have broken translation vector, e.g. broken IOP.\n 2. inconsistency between L5 and S1 labeling within a condition - I don't know if that includes some of the levels start not from the trapezoid shaped sacrum but arbitrary higher. I consulted with few researchers and radiologists, showed few train examples, they confirmed that the L5/S1 level did start from the trapezoid sacrum (either host mislabled it or there was 6th vertebra which again mislabeled). \n 3. Different sizes / px_spacing of the images within the same axial series \n \n\n ", "votes": 3 } ] }, { "id": 2993289, "postDate": "2024-09-19T16:08:07.543Z", "content": "

I think handling flip is one of the suggested pre-processing steps when working with 3D dicom images. Usually, we will convert images to the Patient Coordinate System [https://www.mathworks.com/help/medical-imaging/ug/medical-image-coordinate-systems.html].
\nI haven't checked this series. Maybe this case has a wrong ImageOrientationPatient.

", "rawMarkdown": "I think handling flip is one of the suggested pre-processing steps when working with 3D dicom images. Usually, we will convert images to the Patient Coordinate System [https://www.mathworks.com/help/medical-imaging/ug/medical-image-coordinate-systems.html].\nI haven't checked this series. Maybe this case has a wrong ImageOrientationPatient. ", "votes": 1, "replies": [ { "id": 2993336, "postDate": "2024-09-19T17:42:55.703Z", "content": "

Instance 22
\nIOP: [0.99980393341667, -0.0083668796327, 0.01794686740798, -0.0035436501509, 0.81610967299172, 0.57788618619322]
\nIPP [-94.092687872478, -51.3609149255, -52.850467206281]

\n

Instance 3
\nIOP: [0.99882801193367, -0.0472992708347, -0.0102655518674, -0.0374032361361, -0.6197100347129, -0.7839390733996]
\nIPP [-90.00849110667, 115.54427238215, 64.925080217744]

", "rawMarkdown": "Instance 22\nIOP: [0.99980393341667, -0.0083668796327, 0.01794686740798, -0.0035436501509, 0.81610967299172, 0.57788618619322]\nIPP [-94.092687872478, -51.3609149255, -52.850467206281]\n\nInstance 3\nIOP: [0.99882801193367, -0.0472992708347, -0.0102655518674, -0.0374032361361, -0.6197100347129, -0.7839390733996]\nIPP [-90.00849110667, 115.54427238215, 64.925080217744]", "votes": 2, "replies": [ { "id": 2993555, "postDate": "2024-09-19T23:22:06.087Z", "content": "

This series looks like this. There are a lot of non-parallel slices in the axial T2 dataset. It's truly annoying. \"\"

", "rawMarkdown": "This series looks like this. There are a lot of non-parallel slices in the axial T2 dataset. It's truly annoying. ![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1501163%2F9c2949f19de33d2482bb4e2dcd25c7e1%2Fdownload.png?generation=1726787977216306&alt=media)", "votes": 2 }, { "id": 2993556, "postDate": "2024-09-19T23:27:58.297Z", "content": "

Thank you for verifying.

", "rawMarkdown": "Thank you for verifying.", "votes": 3 } ] } ] }, { "id": 2997127, "postDate": "2024-09-24T09:08:35.420Z", "content": "

I am more concerned about the stability of this competition.

", "rawMarkdown": "I am more concerned about the stability of this competition.", "votes": 2 }, { "id": 2994889, "postDate": "2024-09-21T15:54:23.943Z", "content": "

Grate cockroach here. 😝

", "rawMarkdown": "Grate cockroach here. 😝", "votes": 2 }, { "id": 2992878, "postDate": "2024-09-19T07:44:48.027Z", "content": "

It's flipped or rotated? I mean, till now my workflow assumes sides are not reflected in axials.

", "rawMarkdown": "It's flipped or rotated? I mean, till now my workflow assumes sides are not reflected in axials.", "votes": 2, "replies": [ { "id": 2993024, "postDate": "2024-09-19T11:29:35.417Z", "content": "

hah, you might be 50% right.

", "rawMarkdown": "hah, you might be 50% right.", "votes": 2, "replies": [ { "id": 2993074, "postDate": "2024-09-19T11:58:21.817Z", "content": "

If that's true good news is an important improvement on score once fixed O_O

\n

Actually, may be I should check it carefully. But my axial segmentation loss and validations suggest that there is no reflections. I mean, it seems like with respect back and belly left and right are always where they should be. May be a small portion could pass unnoticed, but not many cases.

", "rawMarkdown": "If that's true good news is an important improvement on score once fixed O_O\n\nActually, may be I should check it carefully. But my axial segmentation loss and validations suggest that there is no reflections. I mean, it seems like with respect back and belly left and right are always where they should be. May be a small portion could pass unnoticed, but not many cases.", "votes": 1 }, { "id": 2993128, "postDate": "2024-09-19T12:38:28.160Z", "content": "

I just dropped it from the training dataset.

", "rawMarkdown": "I just dropped it from the training dataset.", "votes": 1 }, { "id": 2993392, "postDate": "2024-09-19T18:47:01.797Z", "content": "

Just to leave it checked. In the 24686 slices that I've used to train axial segmentation, just picked from coor_df and softly augmented not selected, all left x coordinates were bigger than x right coordinates. And y coordinates were strongly correlated. As spected.

\n

EDIT: Actually the rule x_left < x_right for flip detection doesn't work for the case you've shown. I won't visually check 13k images, just belive. But I think the context that model takes for decide should be very close to:

\n
    \n
  1. Where is back and belly.
    2. \n
  2. Ok then L and R are…
    3. \n
", "rawMarkdown": "Just to leave it checked. In the 24686 slices that I've used to train axial segmentation, just picked from coor_df and softly augmented not selected, all left x coordinates were bigger than x right coordinates. And y coordinates were strongly correlated. As spected.\n\nEDIT: Actually the rule x_left < x_right for flip detection doesn't work for the case you've shown. I won't visually check 13k images, just belive. But I think the context that model takes for decide should be very close to:\n\n1. Where is back and belly.\n2. Ok then L and R are...", "votes": 2 }, { "id": 2993405, "postDate": "2024-09-19T18:56:47.450Z", "content": "

That is interesting, I assume you are sampling multiple instances close to the one provided by the host, otherwise there will be roughly 13k slices total

", "rawMarkdown": "That is interesting, I assume you are sampling multiple instances close to the one provided by the host, otherwise there will be roughly 13k slices total", "votes": 1 }, { "id": 2993416, "postDate": "2024-09-19T19:01:52.060Z", "content": "

Exactly. But with caution. There is many discontinuities.

", "rawMarkdown": "Exactly. But with caution. There is many discontinuities." }, { "id": 2993422, "postDate": "2024-09-19T19:10:11.117Z", "content": "

I wanted to run the SAM2 on top of that to seg the central canal, though it is still in my backlog : ), did you try it by chance?

", "rawMarkdown": "I wanted to run the SAM2 on top of that to seg the central canal, though it is still in my backlog : ), did you try it by chance?", "votes": 1 }, { "id": 2993425, "postDate": "2024-09-19T19:13:19.753Z", "content": "

No. Simple UNet's for segmentation. Pretty decent results for axial slices.

", "rawMarkdown": "No. Simple UNet's for segmentation. Pretty decent results for axial slices." } ] } ] }, { "id": 3000461, "postDate": "2024-09-27T16:37:54.523Z", "content": "

hello sir.i wanted to ask if i add random flip to the transforms can i make the model to learn the feature of the illness rather than focus on its position. and how many wrong with the dataset and how can i solve them.thanks you sir!!!!!!!!!

", "rawMarkdown": "hello sir.i wanted to ask if i add random flip to the transforms can i make the model to learn the feature of the illness rather than focus on its position. and how many wrong with the dataset and how can i solve them.thanks you sir!!!!!!!!!" }, { "id": 2992899, "postDate": "2024-09-19T08:22:28.310Z", "content": "

How rare case is it? If it is 1 per 100, I wouldn't care at all. It is almost impossible to build dataset as big as this one without inaccuracies. Btw, did you check whether left condition is always on the right side of image?

", "rawMarkdown": "How rare case is it? If it is 1 per 100, I wouldn't care at all. It is almost impossible to build dataset as big as this one without inaccuracies. Btw, did you check whether left condition is always on the right side of image?" }, { "id": 2992754, "postDate": "2024-09-19T01:57:11.533Z", "rawMarkdown": "", "votes": -1, "isDeleted": true, "replies": [ { "id": 2992755, "postDate": "2024-09-19T02:09:36.177Z", "content": "

There are 79979 axials total. Guess how I found this one :)

", "rawMarkdown": "There are `79979` axials total. Guess how I found this one :)", "votes": 1 } ] }, { "id": 2997251, "postDate": "2024-09-24T11:21:08.933Z", "content": "

very informative!

", "rawMarkdown": "very informative!" }, { "id": 2995870, "postDate": "2024-09-22T17:47:45.843Z", "content": "

very informative! thank you for sharing

", "rawMarkdown": "very informative! thank you for sharing" } ], "comments": [ { "id": 2994435, "author_name": "hengck23", "author_url": "", "post_date": "2024-09-20T23:49:42.473000", "content": "

this competition is \"relative\" easy if not for the annotation errors.
\nannotation errors includes:

\n\n

issue is: \"is the hidden test data the same?\"
\n(ior past RSNA competition, extra care and verification are taken to ensure hidden test data are correct. It is not known for this competition)

\n

my suggestion: one should do a probe to see if there are annotation error in hidden test data
\n(yes this can be done by doing some test on onsistency of the predicted points)

\n

or you can ask the host to clarify

", "votes": 6, "replies": [ { "id": 2994440, "author_name": "SSS", "author_url": "", "post_date": "2024-09-21T00:26:59.577000", "content": "

That's a good summary.

\n

We can add to that:

\n
    \n
  1. Few series have broken translation vector, e.g. broken IOP.
    2. \n
  2. inconsistency between L5 and S1 labeling within a condition - I don't know if that includes some of the levels start not from the trapezoid shaped sacrum but arbitrary higher. I consulted with few researchers and radiologists, showed few train examples, they confirmed that the L5/S1 level did start from the trapezoid sacrum (either host mislabled it or there was 6th vertebra which again mislabeled).
    3. \n
  3. Different sizes / px_spacing of the images within the same axial series
    4. \n
", "votes": 3, "replies": [] } ] }, { "id": 2993289, "author_name": "Gaofeng Huang", "author_url": "", "post_date": "2024-09-19T16:08:07.543000", "content": "

I think handling flip is one of the suggested pre-processing steps when working with 3D dicom images. Usually, we will convert images to the Patient Coordinate System [https://www.mathworks.com/help/medical-imaging/ug/medical-image-coordinate-systems.html].
\nI haven't checked this series. Maybe this case has a wrong ImageOrientationPatient.

", "votes": 1, "replies": [ { "id": 2993336, "author_name": "SSS", "author_url": "", "post_date": "2024-09-19T17:42:55.703000", "content": "

Instance 22
\nIOP: [0.99980393341667, -0.0083668796327, 0.01794686740798, -0.0035436501509, 0.81610967299172, 0.57788618619322]
\nIPP [-94.092687872478, -51.3609149255, -52.850467206281]

\n

Instance 3
\nIOP: [0.99882801193367, -0.0472992708347, -0.0102655518674, -0.0374032361361, -0.6197100347129, -0.7839390733996]
\nIPP [-90.00849110667, 115.54427238215, 64.925080217744]

", "votes": 2, "replies": [ { "id": 2993555, "author_name": "Gaofeng Huang", "author_url": "", "post_date": "2024-09-19T23:22:06.087000", "content": "

This series looks like this. There are a lot of non-parallel slices in the axial T2 dataset. It's truly annoying. \"\"

", "votes": 2, "replies": [] }, { "id": 2993556, "author_name": "SSS", "author_url": "", "post_date": "2024-09-19T23:27:58.297000", "content": "

Thank you for verifying.

", "votes": 3, "replies": [] } ] } ] }, { "id": 2997127, "author_name": "", "author_url": "", "post_date": "2024-09-24T09:08:35.420000", "content": "

I am more concerned about the stability of this competition.

", "votes": 2, "replies": [] }, { "id": 2994889, "author_name": "", "author_url": "", "post_date": "2024-09-21T15:54:23.943000", "content": "

Grate cockroach here. 😝

", "votes": 2, "replies": [] }, { "id": 2992878, "author_name": "Ángel Jacinto Sánchez Ruiz", "author_url": "", "post_date": "2024-09-19T07:44:48.027000", "content": "

It's flipped or rotated? I mean, till now my workflow assumes sides are not reflected in axials.

", "votes": 2, "replies": [ { "id": 2993024, "author_name": "SSS", "author_url": "", "post_date": "2024-09-19T11:29:35.417000", "content": "

hah, you might be 50% right.

", "votes": 2, "replies": [ { "id": 2993074, "author_name": "Ángel Jacinto Sánchez Ruiz", "author_url": "", "post_date": "2024-09-19T11:58:21.817000", "content": "

If that's true good news is an important improvement on score once fixed O_O

\n

Actually, may be I should check it carefully. But my axial segmentation loss and validations suggest that there is no reflections. I mean, it seems like with respect back and belly left and right are always where they should be. May be a small portion could pass unnoticed, but not many cases.

", "votes": 1, "replies": [] }, { "id": 2993128, "author_name": "SSS", "author_url": "", "post_date": "2024-09-19T12:38:28.160000", "content": "

I just dropped it from the training dataset.

", "votes": 1, "replies": [] }, { "id": 2993392, "author_name": "Ángel Jacinto Sánchez Ruiz", "author_url": "", "post_date": "2024-09-19T18:47:01.797000", "content": "

Just to leave it checked. In the 24686 slices that I've used to train axial segmentation, just picked from coor_df and softly augmented not selected, all left x coordinates were bigger than x right coordinates. And y coordinates were strongly correlated. As spected.

\n

EDIT: Actually the rule x_left < x_right for flip detection doesn't work for the case you've shown. I won't visually check 13k images, just belive. But I think the context that model takes for decide should be very close to:

\n
    \n
  1. Where is back and belly.
    2. \n
  2. Ok then L and R are…
    3. \n
", "votes": 2, "replies": [] }, { "id": 2993405, "author_name": "SSS", "author_url": "", "post_date": "2024-09-19T18:56:47.450000", "content": "

That is interesting, I assume you are sampling multiple instances close to the one provided by the host, otherwise there will be roughly 13k slices total

", "votes": 1, "replies": [] }, { "id": 2993416, "author_name": "Ángel Jacinto Sánchez Ruiz", "author_url": "", "post_date": "2024-09-19T19:01:52.060000", "content": "

Exactly. But with caution. There is many discontinuities.

", "votes": 0, "replies": [] }, { "id": 2993422, "author_name": "SSS", "author_url": "", "post_date": "2024-09-19T19:10:11.117000", "content": "

I wanted to run the SAM2 on top of that to seg the central canal, though it is still in my backlog : ), did you try it by chance?

", "votes": 1, "replies": [] }, { "id": 2993425, "author_name": "Ángel Jacinto Sánchez Ruiz", "author_url": "", "post_date": "2024-09-19T19:13:19.753000", "content": "

No. Simple UNet's for segmentation. Pretty decent results for axial slices.

", "votes": 0, "replies": [] } ] } ] }, { "id": 3000461, "author_name": "Otto_lumous", "author_url": "", "post_date": "2024-09-27T16:37:54.523000", "content": "

hello sir.i wanted to ask if i add random flip to the transforms can i make the model to learn the feature of the illness rather than focus on its position. and how many wrong with the dataset and how can i solve them.thanks you sir!!!!!!!!!

", "votes": 0, "replies": [] }, { "id": 2992899, "author_name": "Rafał Pawłowski", "author_url": "", "post_date": "2024-09-19T08:22:28.310000", "content": "

How rare case is it? If it is 1 per 100, I wouldn't care at all. It is almost impossible to build dataset as big as this one without inaccuracies. Btw, did you check whether left condition is always on the right side of image?

", "votes": 0, "replies": [] }, { "id": 2992754, "author_name": "", "author_url": "", "post_date": "2024-09-19T01:57:11.533000", "content": "", "votes": -1, "replies": [ { "id": 2992755, "author_name": "SSS", "author_url": "", "post_date": "2024-09-19T02:09:36.177000", "content": "

There are 79979 axials total. Guess how I found this one :)

", "votes": 1, "replies": [] } ] }, { "id": 2997251, "author_name": "Zafar", "author_url": "", "post_date": "2024-09-24T11:21:08.933000", "content": "

very informative!

", "votes": 0, "replies": [] }, { "id": 2995870, "author_name": "Amann", "author_url": "", "post_date": "2024-09-22T17:47:45.843000", "content": "

very informative! thank you for sharing

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2992733": "Hi,\n\nI will just leave this cockroach here.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6259210%2Fba6e16d91dbb4616f1e9568110bb715d%2FScreenshot%20from%202024-09-18%2020-54-30.png?generation=1726707280992832&alt=media)", "2994435": "this competition is \"relative\" easy if not for the annotation errors.\nannotation errors includes:\n- inconsistency between L5 and S1 labeling within a condition\n- inconsistency level labeling within a study. if you project all labels of axial and sagittal views to common 3d world coord, they are not aliged\n- inconsistency level labeling within a series. left and right point of same level are not aliged\n- wrong point labels (e.g. 2 level point coincide, left/right point coincide, point becomes at corner of image, etc ...)\n- missing point labels\n- a couple of wrong images\n\nissue is: \"is the hidden test data the same?\"\n(ior past RSNA competition, extra care and verification are taken to ensure hidden test data are correct. It is not known for this competition)\n\nmy suggestion: one should do a probe to see if there are annotation error in hidden test data\n(yes this can be done by doing some test on onsistency of the predicted points)\n\nor you can ask the host to clarify\n", "2993289": "I think handling flip is one of the suggested pre-processing steps when working with 3D dicom images. Usually, we will convert images to the Patient Coordinate System [https://www.mathworks.com/help/medical-imaging/ug/medical-image-coordinate-systems.html].\nI haven't checked this series. Maybe this case has a wrong ImageOrientationPatient. ", "2997127": "I am more concerned about the stability of this competition.", "2994889": "Grate cockroach here. 😝", "2992878": "It's flipped or rotated? I mean, till now my workflow assumes sides are not reflected in axials.", "3000461": "hello sir.i wanted to ask if i add random flip to the transforms can i make the model to learn the feature of the illness rather than focus on its position. and how many wrong with the dataset and how can i solve them.thanks you sir!!!!!!!!!", "2992899": "How rare case is it? If it is 1 per 100, I wouldn't care at all. It is almost impossible to build dataset as big as this one without inaccuracies. Btw, did you check whether left condition is always on the right side of image?", "2992754": "", "2997251": "very informative!", "2995870": "very informative! thank you for sharing" } }