{ "id": 417374, "title": "Can you annotate this image accurately?", "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/417374", "author_name": "", "post_date": "2023-06-15T12:47:38.580927700Z", "votes": 23, "comment_count": 8, "views": 0, "content": "

Suddenly, Let's annotate this image.
\n\"8cb7d2ef7d2c.png\"
\nThis image's id is 8cb7d2ef7d2c, which belongs to WSI 2, Dataset 2.
\nA good medical doctor might be able to annotate with evidence. However, I am not, and neither are untrained machine learning models.
\nNow, let's look at the answer, GroundTruth.
\n\"GroundTruth\"
\nHere it is.
\nI cannot distinguish the white area on the left from the white area on the right. I actually let my best model infer and this is what I got.
\nLeft is input, middle is GT, and right is output.
\n\"result
\nMy model seems to be confused. Is there a way to train the model correctly?
\nIn fact, this image is not isolated. I used the tile metadata, tile_meta.csv, and connected the images. The image including the surrounding area looks like this. (The image has been reduced to 0.25x size.)
\n\"connected
\nWouldn't things be a little different if we knew what was going on around the image in problem? (I am not sure yet).
\nThis shows that using the provided tile images as is produces less-than-ideal results.
\nIt is important to give the model a larger area.
\nTherefore, the key to winning this competition will be to review the annotations and utilize metadata, etc…
\nI am not confident we can do that, so at worst I will have to withdraw…

\n

The color of the image is wrong because of the incorrect use of OpenCV. Sorry.

\n

Notebooks I have made that focus on WSI 1 and 2

\n

A similar topic was posted by Chenglu. (I noticed this after I finished writing it)

", "messages": [ { "id": "2303737", "postDate": "06/15/2023 12:47:38", "content": "

Suddenly, Let's annotate this image.
\n\"8cb7d2ef7d2c.png\"
\nThis image's id is 8cb7d2ef7d2c, which belongs to WSI 2, Dataset 2.
\nA good medical doctor might be able to annotate with evidence. However, I am not, and neither are untrained machine learning models.
\nNow, let's look at the answer, GroundTruth.
\n\"GroundTruth\"
\nHere it is.
\nI cannot distinguish the white area on the left from the white area on the right. I actually let my best model infer and this is what I got.
\nLeft is input, middle is GT, and right is output.
\n\"result
\nMy model seems to be confused. Is there a way to train the model correctly?
\nIn fact, this image is not isolated. I used the tile metadata, tile_meta.csv, and connected the images. The image including the surrounding area looks like this. (The image has been reduced to 0.25x size.)
\n\"connected
\nWouldn't things be a little different if we knew what was going on around the image in problem? (I am not sure yet).
\nThis shows that using the provided tile images as is produces less-than-ideal results.
\nIt is important to give the model a larger area.
\nTherefore, the key to winning this competition will be to review the annotations and utilize metadata, etc…
\nI am not confident we can do that, so at worst I will have to withdraw…

\n

The color of the image is wrong because of the incorrect use of OpenCV. Sorry.

\n

Notebooks I have made that focus on WSI 1 and 2

\n

A similar topic was posted by Chenglu. (I noticed this after I finished writing it)

", "rawMarkdown": "Suddenly, Let's annotate this image.\n![8cb7d2ef7d2c.png](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2294613%2F0a5c8fd66370e21b34c4c4537bab724f%2F8cb7d2ef7d2c.png?generation=1686829800681235&alt=media)\nThis image's id is 8cb7d2ef7d2c, which belongs to WSI 2, Dataset 2.\nA good medical doctor might be able to annotate with evidence. However, I am not, and neither are untrained machine learning models.\nNow, let's look at the answer, GroundTruth.\n![GroundTruth](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2294613%2Fd7a977c782867979092d40db12aa813b%2F8cb7d2ef7d2c_mask.png?generation=1686830566915274&alt=media)\nHere it is.\nI cannot distinguish the white area on the left from the white area on the right. I actually let my best model infer and this is what I got.\nLeft is input, middle is GT, and right is output.\n![result of inference](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2294613%2F4519f684fb29efbc97bc075e1bef8ca2%2F9.png?generation=1686830802312243&alt=media)\nMy model seems to be confused. Is there a way to train the model correctly?\nIn fact, this image is not isolated. I used the tile metadata, tile_meta.csv, and connected the images. The image including the surrounding area looks like this. (The image has been reduced to 0.25x size.)\n![connected image](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2294613%2F5fdc4e49e9db6309bb8de40b77085fdc%2F__results___21_0.png?generation=1686831428666902&alt=media)\nWouldn't things be a little different if we knew what was going on around the image in problem? (I am not sure yet).\nThis shows that using the provided tile images as is produces less-than-ideal results.\nIt is important to give the model a larger area.\nTherefore, the key to winning this competition will be to review the annotations and utilize metadata, etc...\nI am not confident we can do that, so at worst I will have to withdraw...\n\n*The color of the image is wrong because of the incorrect use of OpenCV. Sorry.*\n\n[Notebooks I have made that focus on WSI 1 and 2](https://www.kaggle.com/code/itsuki9180/investigate-tiles-on-wsi-1-and-2)\n\n[A similar topic](https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/417314) was posted by [Chenglu](https://www.kaggle.com/snaker). (I noticed this after I finished writing it)", "votes": null }, { "id": "2303764", "postDate": "06/15/2023 13:12:36", "content": "

I believe that utilizing the entire WSI to provide additional context, particularly for the edges of images, is an impressive .
\nWe can at least use train_metadata to generate additional image slices as data augementation…

", "rawMarkdown": "I believe that utilizing the entire WSI to provide additional context, particularly for the edges of images, is an impressive ~~but challenging task without test_metadata~~. \nWe can at least use train_metadata to generate additional image slices as data augementation...", "votes": null }, { "id": "2303783", "postDate": "06/15/2023 13:24:05", "content": "

thanks for sharing, it's good to know that we can achieve 0.5 without context information ; )

", "rawMarkdown": "thanks for sharing, it's good to know that we can achieve 0.5 without context information ; )", "votes": null }, { "id": "2303932", "postDate": "06/15/2023 15:01:30", "content": "

https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/414669#2303907

\n

we can use tile_meta.csv in test set!!

", "rawMarkdown": "https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/414669#2303907\n\nwe can use `tile_meta.csv` in test set!!", "votes": null }, { "id": "2304311", "postDate": "06/15/2023 23:13:46", "content": "

@katherinegustilo

\n

Are these large vessels arterioles or venules?

\n

This competition focuses on microvasculature in the kidney, including arterioles, capillaries and venules. Larger vessels structures, i.e. arteries and veins, were not segmented in this dataset.\n
\nIt would be greatly appreciated if you could add to the biological overview the approximate sizes of arterioles and venules and examples of large vessels excluded!

", "rawMarkdown": "katherinegustilo \n\nAre these large vessels arterioles or venules?\n\n```This competition focuses on microvasculature in the kidney, including arterioles, capillaries and venules. Larger vessels structures, i.e. arteries and veins, were not segmented in this dataset.\n```\nIt would be greatly appreciated if you could add to the biological overview the approximate sizes of arterioles and venules and examples of large vessels excluded!", "votes": null }, { "id": "2312380", "postDate": "06/21/2023 21:15:39", "content": "

Hi Patriot,

\n

Approximate sizes of arterioles and venules are a matter of some debate as these structures are throughout the body and have a dynamic nature, i.e. a venule can be dilated or not. For this specific example from dataset 2, this structure is likely either a dilated venule or a small vein. This dataset was not expert validated. If it were, this structure would have been excluded as being too large for relevance. Such structures are not present in dataset 1.

\n

Best,
\nKate Gustilo, Anatomist and Research Analyst

", "rawMarkdown": "Hi Patriot,\n\nApproximate sizes of arterioles and venules are a matter of some debate as these structures are throughout the body and have a dynamic nature, i.e. a venule can be dilated or not. For this specific example from dataset 2, this structure is likely either a dilated venule or a small vein. This dataset was not expert validated. If it were, this structure would have been excluded as being too large for relevance. Such structures are not present in dataset 1. \n\nBest,\nKate Gustilo, Anatomist and Research Analyst", "votes": null }, { "id": "2313100", "postDate": "06/22/2023 12:14:18", "content": "

This may not have much to do with the competition, but I would like to know the length of 1 pixel in these images. Also, is the scale the same for all WSI?

\n

By \"dynamic nature\", do you mean that the blood vessels contract and dilate?

\n

Also, this is pure curiosity, but does it make sense to annotate only the small vessels? Are there situations where you want to focus only on microvessels for renal tissue analysis?

\n

Sorry for asking so many questions. I'm sorry if this is a question that has been mentioned somewhere else.
\nIf you could answer them in your spare time I would be very happy.

", "rawMarkdown": "This may not have much to do with the competition, but I would like to know the length of 1 pixel in these images. Also, is the scale the same for all WSI?\n\nBy \"dynamic nature\", do you mean that the blood vessels contract and dilate?\n\nAlso, this is pure curiosity, but does it make sense to annotate only the small vessels? Are there situations where you want to focus only on microvessels for renal tissue analysis?\n\nSorry for asking so many questions. I'm sorry if this is a question that has been mentioned somewhere else.\nIf you could answer them in your spare time I would be very happy.", "votes": null }, { "id": "2318811", "postDate": "06/26/2023 15:48:13", "content": "

Hi YYama,

\n

Each pixel is 0.5 micrometers in length. This scale is the same across all WSIs in the dataset.

\n

Yes, blood vessels contract and dilate (vasodilation/vasoconstriction), so their size would have some dependancy on the state the vessels were in during tissue sectioning.

\n

We are focusing on microvasculature in order to support the Vasculature Common Coordinate Framework (VCCF). Please see the competition overview for more information and links to papers describing this effort.

\n

Thank you for your questions and curiosity.

\n

Best,
\nKate Gustilo

", "rawMarkdown": "Hi YYama,\n\nEach pixel is 0.5 micrometers in length. This scale is the same across all WSIs in the dataset. \n\nYes, blood vessels contract and dilate (vasodilation/vasoconstriction), so their size would have some dependancy on the state the vessels were in during tissue sectioning.\n\nWe are focusing on microvasculature in order to support the Vasculature Common Coordinate Framework (VCCF). Please see the competition overview for more information and links to papers describing this effort. \n\nThank you for your questions and curiosity.\n\nBest,\nKate Gustilo", "votes": null }, { "id": "2319349", "postDate": "06/27/2023 04:57:21", "content": "

Hi Kate,

\n

Then the scales are unified, which is quite good news for me!

\n

I read about the VCCF initiative. I see, it is interesting. I am doing medical imaging research and I started to think that we could make use of the VCCF's findings…

\n

Thank you for all your replies!

", "rawMarkdown": "Hi Kate,\n\nThen the scales are unified, which is quite good news for me!\n\nI read about the VCCF initiative. I see, it is interesting. I am doing medical imaging research and I started to think that we could make use of the VCCF's findings...\n\nThank you for all your replies!", "votes": null } ], "comments": [ { "id": 2303764, "author_name": "zzy990106", "author_url": "", "post_date": "06/15/2023 13:12:36", "content": "

I believe that utilizing the entire WSI to provide additional context, particularly for the edges of images, is an impressive .
\nWe can at least use train_metadata to generate additional image slices as data augementation…

", "votes": null, "replies": [] }, { "id": 2303783, "author_name": "snaker", "author_url": "", "post_date": "06/15/2023 13:24:05", "content": "

thanks for sharing, it's good to know that we can achieve 0.5 without context information ; )

", "votes": null, "replies": [] }, { "id": 2303932, "author_name": "snaker", "author_url": "", "post_date": "06/15/2023 15:01:30", "content": "

https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/414669#2303907

\n

we can use tile_meta.csv in test set!!

", "votes": null, "replies": [] }, { "id": 2304311, "author_name": "abebe9849", "author_url": "", "post_date": "06/15/2023 23:13:46", "content": "

@katherinegustilo

\n

Are these large vessels arterioles or venules?

\n

This competition focuses on microvasculature in the kidney, including arterioles, capillaries and venules. Larger vessels structures, i.e. arteries and veins, were not segmented in this dataset.\n
\nIt would be greatly appreciated if you could add to the biological overview the approximate sizes of arterioles and venules and examples of large vessels excluded!

", "votes": null, "replies": [ { "id": 2312380, "author_name": "katherinegustilo", "author_url": "", "post_date": "06/21/2023 21:15:39", "content": "

Hi Patriot,

\n

Approximate sizes of arterioles and venules are a matter of some debate as these structures are throughout the body and have a dynamic nature, i.e. a venule can be dilated or not. For this specific example from dataset 2, this structure is likely either a dilated venule or a small vein. This dataset was not expert validated. If it were, this structure would have been excluded as being too large for relevance. Such structures are not present in dataset 1.

\n

Best,
\nKate Gustilo, Anatomist and Research Analyst

", "votes": null, "replies": [ { "id": 2313100, "author_name": "yosukeyama", "author_url": "", "post_date": "06/22/2023 12:14:18", "content": "

This may not have much to do with the competition, but I would like to know the length of 1 pixel in these images. Also, is the scale the same for all WSI?

\n

By \"dynamic nature\", do you mean that the blood vessels contract and dilate?

\n

Also, this is pure curiosity, but does it make sense to annotate only the small vessels? Are there situations where you want to focus only on microvessels for renal tissue analysis?

\n

Sorry for asking so many questions. I'm sorry if this is a question that has been mentioned somewhere else.
\nIf you could answer them in your spare time I would be very happy.

", "votes": null, "replies": [ { "id": 2318811, "author_name": "katherinegustilo", "author_url": "", "post_date": "06/26/2023 15:48:13", "content": "

Hi YYama,

\n

Each pixel is 0.5 micrometers in length. This scale is the same across all WSIs in the dataset.

\n

Yes, blood vessels contract and dilate (vasodilation/vasoconstriction), so their size would have some dependancy on the state the vessels were in during tissue sectioning.

\n

We are focusing on microvasculature in order to support the Vasculature Common Coordinate Framework (VCCF). Please see the competition overview for more information and links to papers describing this effort.

\n

Thank you for your questions and curiosity.

\n

Best,
\nKate Gustilo

", "votes": null, "replies": [ { "id": 2319349, "author_name": "yosukeyama", "author_url": "", "post_date": "06/27/2023 04:57:21", "content": "

Hi Kate,

\n

Then the scales are unified, which is quite good news for me!

\n

I read about the VCCF initiative. I see, it is interesting. I am doing medical imaging research and I started to think that we could make use of the VCCF's findings…

\n

Thank you for all your replies!

", "votes": null, "replies": [] } ] } ] } ] } ] } ], "raw_markdown_by_id": { "2303737": "Suddenly, Let's annotate this image.\n![8cb7d2ef7d2c.png](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2294613%2F0a5c8fd66370e21b34c4c4537bab724f%2F8cb7d2ef7d2c.png?generation=1686829800681235&alt=media)\nThis image's id is 8cb7d2ef7d2c, which belongs to WSI 2, Dataset 2.\nA good medical doctor might be able to annotate with evidence. However, I am not, and neither are untrained machine learning models.\nNow, let's look at the answer, GroundTruth.\n![GroundTruth](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2294613%2Fd7a977c782867979092d40db12aa813b%2F8cb7d2ef7d2c_mask.png?generation=1686830566915274&alt=media)\nHere it is.\nI cannot distinguish the white area on the left from the white area on the right. I actually let my best model infer and this is what I got.\nLeft is input, middle is GT, and right is output.\n![result of inference](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2294613%2F4519f684fb29efbc97bc075e1bef8ca2%2F9.png?generation=1686830802312243&alt=media)\nMy model seems to be confused. Is there a way to train the model correctly?\nIn fact, this image is not isolated. I used the tile metadata, tile_meta.csv, and connected the images. The image including the surrounding area looks like this. (The image has been reduced to 0.25x size.)\n![connected image](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2294613%2F5fdc4e49e9db6309bb8de40b77085fdc%2F__results___21_0.png?generation=1686831428666902&alt=media)\nWouldn't things be a little different if we knew what was going on around the image in problem? (I am not sure yet).\nThis shows that using the provided tile images as is produces less-than-ideal results.\nIt is important to give the model a larger area.\nTherefore, the key to winning this competition will be to review the annotations and utilize metadata, etc...\nI am not confident we can do that, so at worst I will have to withdraw...\n\n*The color of the image is wrong because of the incorrect use of OpenCV. Sorry.*\n\n[Notebooks I have made that focus on WSI 1 and 2](https://www.kaggle.com/code/itsuki9180/investigate-tiles-on-wsi-1-and-2)\n\n[A similar topic](https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/417314) was posted by [Chenglu](https://www.kaggle.com/snaker). (I noticed this after I finished writing it)", "2303764": "I believe that utilizing the entire WSI to provide additional context, particularly for the edges of images, is an impressive ~~but challenging task without test_metadata~~. \nWe can at least use train_metadata to generate additional image slices as data augementation...", "2303783": "thanks for sharing, it's good to know that we can achieve 0.5 without context information ; )", "2303932": "https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/414669#2303907\n\nwe can use `tile_meta.csv` in test set!!", "2304311": "katherinegustilo \n\nAre these large vessels arterioles or venules?\n\n```This competition focuses on microvasculature in the kidney, including arterioles, capillaries and venules. Larger vessels structures, i.e. arteries and veins, were not segmented in this dataset.\n```\nIt would be greatly appreciated if you could add to the biological overview the approximate sizes of arterioles and venules and examples of large vessels excluded!", "2312380": "Hi Patriot,\n\nApproximate sizes of arterioles and venules are a matter of some debate as these structures are throughout the body and have a dynamic nature, i.e. a venule can be dilated or not. For this specific example from dataset 2, this structure is likely either a dilated venule or a small vein. This dataset was not expert validated. If it were, this structure would have been excluded as being too large for relevance. Such structures are not present in dataset 1. \n\nBest,\nKate Gustilo, Anatomist and Research Analyst", "2313100": "This may not have much to do with the competition, but I would like to know the length of 1 pixel in these images. Also, is the scale the same for all WSI?\n\nBy \"dynamic nature\", do you mean that the blood vessels contract and dilate?\n\nAlso, this is pure curiosity, but does it make sense to annotate only the small vessels? Are there situations where you want to focus only on microvessels for renal tissue analysis?\n\nSorry for asking so many questions. I'm sorry if this is a question that has been mentioned somewhere else.\nIf you could answer them in your spare time I would be very happy.", "2318811": "Hi YYama,\n\nEach pixel is 0.5 micrometers in length. This scale is the same across all WSIs in the dataset. \n\nYes, blood vessels contract and dilate (vasodilation/vasoconstriction), so their size would have some dependancy on the state the vessels were in during tissue sectioning.\n\nWe are focusing on microvasculature in order to support the Vasculature Common Coordinate Framework (VCCF). Please see the competition overview for more information and links to papers describing this effort. \n\nThank you for your questions and curiosity.\n\nBest,\nKate Gustilo", "2319349": "Hi Kate,\n\nThen the scales are unified, which is quite good news for me!\n\nI read about the VCCF initiative. I see, it is interesting. I am doing medical imaging research and I started to think that we could make use of the VCCF's findings...\n\nThank you for all your replies!" }, "source": "meta" }