{ "id": 468399, "title": "discrepancy betwwen annotation for kidney1 dense and kidney1 voi?", "url": "/competitions/blood-vessel-segmentation/discussion/468399", "author_name": "hengck23", "post_date": "2024-01-16T14:19:53.601000", "votes": 7, "comment_count": 7, "views": 0, "content": "

There may be bug in my code but i find that labeling for kidney1 dense and kidney1 voi are \"different\".
\nTHe annotation process is described in https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/454353 and also in the paper.

\n

my question is:

\n
    \n
  1. is annotation done in the original scan resolution (e.g. 25.14um/voxel for public test) or binned resolution (e.g. to 50.28um/voxel for public test).
    2. \n
\n

This is important becuase the behaviour of 3d flood fill is different for different resolution. espeically for the private test.

\n

Here,if annotation is done at scan 15.77um/voxel , we expect more fine vessel in ground truth.

\n

if annotation is done at scan binned resolution, 63.08um/voxel (bin x4), there will be less small vessel.

\n
\n

reference paper:
\nDeep Learning for Vascular Segmentation and Applications in Phase Contrast Tomography Imaging
\nhttps://arxiv.org/pdf/2311.13319.pdf

\n

\"It should be noted that with the binning, median filter and manual approach, vessels with diameter as small as
\n1-2 pixels could be segmented\"

\n

i think segmentation and thresholding is applied after binning according to the paper ???

", "messages": [ { "id": 2604579, "postDate": "2024-01-16T14:19:53.600Z", "content": "

There may be bug in my code but i find that labeling for kidney1 dense and kidney1 voi are \"different\".
\nTHe annotation process is described in https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/454353 and also in the paper.

\n

my question is:

\n
    \n
  1. is annotation done in the original scan resolution (e.g. 25.14um/voxel for public test) or binned resolution (e.g. to 50.28um/voxel for public test).
    2. \n
\n

This is important becuase the behaviour of 3d flood fill is different for different resolution. espeically for the private test.

\n

Here,if annotation is done at scan 15.77um/voxel , we expect more fine vessel in ground truth.

\n

if annotation is done at scan binned resolution, 63.08um/voxel (bin x4), there will be less small vessel.

\n
\n

reference paper:
\nDeep Learning for Vascular Segmentation and Applications in Phase Contrast Tomography Imaging
\nhttps://arxiv.org/pdf/2311.13319.pdf

\n

\"It should be noted that with the binning, median filter and manual approach, vessels with diameter as small as
\n1-2 pixels could be segmented\"

\n

i think segmentation and thresholding is applied after binning according to the paper ???

", "rawMarkdown": "There may be bug in my code but i find that labeling for kidney1 dense and kidney1 voi are \"different\".\nTHe annotation process is described in https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/454353 and also in the paper.\n\nmy question is:\n1. is annotation done in the original scan resolution (e.g. 25.14um/voxel for public test) or binned resolution (e.g. to 50.28um/voxel for public test).\n\nThis is important becuase the behaviour of 3d flood fill is different for different resolution. espeically for the private test. \n\nHere,if annotation is done at scan 15.77um/voxel , we expect more fine vessel in ground truth. \n\nif annotation is done at scan binned resolution, 63.08um/voxel (bin x4), there will be less small vessel.\n\n---\nreference paper:\nDeep Learning for Vascular Segmentation and Applications in Phase Contrast Tomography Imaging\nhttps://arxiv.org/pdf/2311.13319.pdf\n\n\"It should be noted that with the binning, median filter and manual approach, vessels with diameter as small as\n1-2 pixels could be segmented\"\n\ni think segmentation and thresholding is applied after binning according to the paper ???\n\n\n", "votes": 7 }, { "id": 2609835, "postDate": "2024-01-19T18:11:12.603Z", "content": "

Hi sorry missed this, yes segmentation is done on binned data.

", "rawMarkdown": "Hi sorry missed this, yes segmentation is done on binned data.", "votes": 1 }, { "id": 2605476, "postDate": "2024-01-17T05:06:31.307Z", "content": "

I once fine-tuned kidney_1_dense model on Kidney_1_voi, and got public score 0.

", "rawMarkdown": "I once fine-tuned kidney_1_dense model on Kidney_1_voi, and got public score 0.", "votes": 1, "replies": [ { "id": 2607785, "postDate": "2024-01-18T12:42:42.280Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 2607784, "postDate": "2024-01-18T12:40:54.520Z", "content": "

The spacing for every 10 chapters of kidney_1_voi is 46.8 to 52 um, and it is not exactly the same as kidney_1_dense, which is 50 um. But this is a fine difference and I think it will be almost the same. Do you mean that there is no label for kidney_1_dense that contains the label area of kidney_1_voi?

", "rawMarkdown": "The spacing for every 10 chapters of kidney_1_voi is 46.8 to 52 um, and it is not exactly the same as kidney_1_dense, which is 50 um. But this is a fine difference and I think it will be almost the same. Do you mean that there is no label for kidney_1_dense that contains the label area of kidney_1_voi?", "replies": [ { "id": 2611942, "postDate": "2024-01-21T04:49:55.920Z", "content": "

@ajobseeker - may I ask is there a way to convert or transform kidney_1_voi to the values provided for Public and Private Test? (see below copied from the Data section) especially Private as it is a bit different to what is provided in train. Or does the binning have to be done before tiff saves?
\nWas wondering if

\n
\n

The spacing for every 10 chapters of kidney_1_voi is 46.8 to 52 um, and it is not exactly the same as kidney_1_dense, which is 50 um

\n
\n

would that mean for Private Test every 12 or so chapters of kidney_1_voi?

\n

It is OK if you would rather not comment before competition ends. Thanks in advance!

\n
\n

kidney_1_voi - A high-resolution subset of kidney_1, at 5.2um resolution.

\n

Public Test:
\n Continuous 3D part of a whole human kidney imaged with HiP-CT - Originally scanned at 25.14um/voxel and binned to 50.28um/voxel (bin x2) before segmentation.

\n

Private Test:
\n Continuous 3D part of a whole human kidney imaged with HiP-CT - Originally scanned at 15.77um/voxel binned to 63.08um/voxel (bin x4) before segmentation.

\n
", "rawMarkdown": "@ajobseeker - may I ask is there a way to convert or transform kidney_1_voi to the values provided for Public and Private Test? (see below copied from the Data section) especially Private as it is a bit different to what is provided in train. Or does the binning have to be done before tiff saves?\nWas wondering if \n>The spacing for every 10 chapters of kidney_1_voi is 46.8 to 52 um, and it is not exactly the same as kidney_1_dense, which is 50 um\n\nwould that mean for Private Test every 12 or so chapters of kidney_1_voi?\n\n\nIt is OK if you would rather not comment before competition ends. Thanks in advance!\n\n>kidney_1_voi - A high-resolution subset of kidney_1, at 5.2um resolution.\n\n>Public Test:\nContinuous 3D part of a whole human kidney imaged with HiP-CT - Originally scanned at 25.14um/voxel and binned to 50.28um/voxel (bin x2) before segmentation.\n\n>Private Test:\nContinuous 3D part of a whole human kidney imaged with HiP-CT - Originally scanned at 15.77um/voxel binned to 63.08um/voxel (bin x4) before segmentation.\n" } ] }, { "id": 2605704, "postDate": "2024-01-17T07:31:18.513Z", "content": "

In the Data

\n
\n

We provided the higher resolution 5.2um/voxel VOI, to give some example data of different resolutions as we thought this might be useful.

\n
\n

Had considered if this could be binned to similar for public and private test and then be useful for validation? For both on the Data page it says they are binned before segmentation. Thought the same process was done for Train except kidney_1_voi.

\n

Maybe you want to add a mention for Claire Walsh on your post here to help for a response?

", "rawMarkdown": "In the Data \n>We provided the higher resolution 5.2um/voxel VOI, to give some example data of different resolutions as we thought this might be useful.\n\nHad considered if this could be binned to similar for public and private test and then be useful for validation? For both on the Data page it says they are binned before segmentation. Thought the same process was done for Train except kidney_1_voi.\n\nMaybe you want to add a mention for Claire Walsh on your post here to help for a response?\n" }, { "id": 2604681, "postDate": "2024-01-16T15:47:03.480Z", "content": "

I noticed that you mentioned 3D flood fill. Are you using 3D flood fill in post-processing, or are you using it to complete the annotation of the upper half of the kidney based on the lower half in the kidney_3_dense dataset?

", "rawMarkdown": "I noticed that you mentioned 3D flood fill. Are you using 3D flood fill in post-processing, or are you using it to complete the annotation of the upper half of the kidney based on the lower half in the kidney_3_dense dataset?" } ], "comments": [ { "id": 2609835, "author_name": "Claire Walsh", "author_url": "", "post_date": "2024-01-19T18:11:12.603000", "content": "

Hi sorry missed this, yes segmentation is done on binned data.

", "votes": 1, "replies": [] }, { "id": 2605476, "author_name": "dragon zhang", "author_url": "", "post_date": "2024-01-17T05:06:31.307000", "content": "

I once fine-tuned kidney_1_dense model on Kidney_1_voi, and got public score 0.

", "votes": 1, "replies": [ { "id": 2607785, "author_name": "", "author_url": "", "post_date": "2024-01-18T12:42:42.280000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 2607784, "author_name": "HB", "author_url": "", "post_date": "2024-01-18T12:40:54.520000", "content": "

The spacing for every 10 chapters of kidney_1_voi is 46.8 to 52 um, and it is not exactly the same as kidney_1_dense, which is 50 um. But this is a fine difference and I think it will be almost the same. Do you mean that there is no label for kidney_1_dense that contains the label area of kidney_1_voi?

", "votes": 0, "replies": [ { "id": 2611942, "author_name": "something4kag", "author_url": "", "post_date": "2024-01-21T04:49:55.920000", "content": "

@ajobseeker - may I ask is there a way to convert or transform kidney_1_voi to the values provided for Public and Private Test? (see below copied from the Data section) especially Private as it is a bit different to what is provided in train. Or does the binning have to be done before tiff saves?
\nWas wondering if

\n
\n

The spacing for every 10 chapters of kidney_1_voi is 46.8 to 52 um, and it is not exactly the same as kidney_1_dense, which is 50 um

\n
\n

would that mean for Private Test every 12 or so chapters of kidney_1_voi?

\n

It is OK if you would rather not comment before competition ends. Thanks in advance!

\n
\n

kidney_1_voi - A high-resolution subset of kidney_1, at 5.2um resolution.

\n

Public Test:
\n Continuous 3D part of a whole human kidney imaged with HiP-CT - Originally scanned at 25.14um/voxel and binned to 50.28um/voxel (bin x2) before segmentation.

\n

Private Test:
\n Continuous 3D part of a whole human kidney imaged with HiP-CT - Originally scanned at 15.77um/voxel binned to 63.08um/voxel (bin x4) before segmentation.

\n
", "votes": 0, "replies": [] } ] }, { "id": 2605704, "author_name": "something4kag", "author_url": "", "post_date": "2024-01-17T07:31:18.513000", "content": "

In the Data

\n
\n

We provided the higher resolution 5.2um/voxel VOI, to give some example data of different resolutions as we thought this might be useful.

\n
\n

Had considered if this could be binned to similar for public and private test and then be useful for validation? For both on the Data page it says they are binned before segmentation. Thought the same process was done for Train except kidney_1_voi.

\n

Maybe you want to add a mention for Claire Walsh on your post here to help for a response?

", "votes": 0, "replies": [] }, { "id": 2604681, "author_name": "", "author_url": "", "post_date": "2024-01-16T15:47:03.480000", "content": "

I noticed that you mentioned 3D flood fill. Are you using 3D flood fill in post-processing, or are you using it to complete the annotation of the upper half of the kidney based on the lower half in the kidney_3_dense dataset?

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2604579": "There may be bug in my code but i find that labeling for kidney1 dense and kidney1 voi are \"different\".\nTHe annotation process is described in https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/454353 and also in the paper.\n\nmy question is:\n1. is annotation done in the original scan resolution (e.g. 25.14um/voxel for public test) or binned resolution (e.g. to 50.28um/voxel for public test).\n\nThis is important becuase the behaviour of 3d flood fill is different for different resolution. espeically for the private test. \n\nHere,if annotation is done at scan 15.77um/voxel , we expect more fine vessel in ground truth. \n\nif annotation is done at scan binned resolution, 63.08um/voxel (bin x4), there will be less small vessel.\n\n---\nreference paper:\nDeep Learning for Vascular Segmentation and Applications in Phase Contrast Tomography Imaging\nhttps://arxiv.org/pdf/2311.13319.pdf\n\n\"It should be noted that with the binning, median filter and manual approach, vessels with diameter as small as\n1-2 pixels could be segmented\"\n\ni think segmentation and thresholding is applied after binning according to the paper ???\n\n\n", "2609835": "Hi sorry missed this, yes segmentation is done on binned data.", "2605476": "I once fine-tuned kidney_1_dense model on Kidney_1_voi, and got public score 0.", "2607784": "The spacing for every 10 chapters of kidney_1_voi is 46.8 to 52 um, and it is not exactly the same as kidney_1_dense, which is 50 um. But this is a fine difference and I think it will be almost the same. Do you mean that there is no label for kidney_1_dense that contains the label area of kidney_1_voi?", "2605704": "In the Data \n>We provided the higher resolution 5.2um/voxel VOI, to give some example data of different resolutions as we thought this might be useful.\n\nHad considered if this could be binned to similar for public and private test and then be useful for validation? For both on the Data page it says they are binned before segmentation. Thought the same process was done for Train except kidney_1_voi.\n\nMaybe you want to add a mention for Claire Walsh on your post here to help for a response?\n", "2604681": "I noticed that you mentioned 3D flood fill. Are you using 3D flood fill in post-processing, or are you using it to complete the annotation of the upper half of the kidney based on the lower half in the kidney_3_dense dataset?" } }