{
  "id": 336472,
  "title": "My first impressions",
  "url": "/competitions/hubmap-organ-segmentation/discussion/336472",
  "author_name": "",
  "post_date": "2022-07-11T10:30:59.609339600Z",
  "votes": 33,
  "comment_count": 10,
  "views": 0,
  "content": "<ul>\n<li>There are 351 images and 4909 annotations. Dataset is relatively smaller compared to similar competitions but there are still fair amount of annotations.</li>\n<li>Annotations are given as polygons and rle masks. In some of the discussions, organizers mentioned that polygons are more raw so I guess rle masks are derived from them. I checked polygons and masks side by side and they appear to be consistent. This allows us to try different approaches such as instance segmentation, semantic segmentation and even domain specific algorithms such as cellpose.</li>\n<li>I checked some of the images and annotations. Annotation quality is not pixel perfect but it looks okay so far. Some of the images may have incorrect annotations like the ones below.</li>\n</ul>\n<p>Red lines are polygons and yellow areas are masks.</p>\n<p>Small line in the middle is probably an incorrect annotation.<br>\n<img src=\"https://i.ibb.co/hB31Fms/4301-annotations.png\" alt=\"4301\"></p>\n<p>Annotation at the bottom looks pretty bad, especially its bottom edge.<br>\n<img src=\"https://i.ibb.co/kcp1DDm/6021-annotations.png\" alt=\"6021\"></p>\n<ul>\n<li><p>Since there is only one class and evaluation metric is dice coefficient, you don't have to worry about overlaps.</p></li>\n<li><p>Some of the images might require extra cleaning steps. For instance:<br>\n<img src=\"https://i.ibb.co/z2Xg616/31800-annotations.png\" alt=\"31800\"></p></li>\n<li><p>I think there will be a middle scale shake up at the end since  all of the hidden test set images are taken from another source (hubmap). This is the reason why I really liked this competition. Teams that can simulate different pixel sizes and tissue thicknesses will probably win this competition.</p></li>\n</ul>\n<p>Those are my initial observations. Good luck to everyone!</p>",
  "messages": [
    {
      "id": "1851532",
      "postDate": "07/11/2022 10:30:59",
      "content": "<ul>\n<li>There are 351 images and 4909 annotations. Dataset is relatively smaller compared to similar competitions but there are still fair amount of annotations.</li>\n<li>Annotations are given as polygons and rle masks. In some of the discussions, organizers mentioned that polygons are more raw so I guess rle masks are derived from them. I checked polygons and masks side by side and they appear to be consistent. This allows us to try different approaches such as instance segmentation, semantic segmentation and even domain specific algorithms such as cellpose.</li>\n<li>I checked some of the images and annotations. Annotation quality is not pixel perfect but it looks okay so far. Some of the images may have incorrect annotations like the ones below.</li>\n</ul>\n<p>Red lines are polygons and yellow areas are masks.</p>\n<p>Small line in the middle is probably an incorrect annotation.<br>\n<img src=\"https://i.ibb.co/hB31Fms/4301-annotations.png\" alt=\"4301\"></p>\n<p>Annotation at the bottom looks pretty bad, especially its bottom edge.<br>\n<img src=\"https://i.ibb.co/kcp1DDm/6021-annotations.png\" alt=\"6021\"></p>\n<ul>\n<li><p>Since there is only one class and evaluation metric is dice coefficient, you don't have to worry about overlaps.</p></li>\n<li><p>Some of the images might require extra cleaning steps. For instance:<br>\n<img src=\"https://i.ibb.co/z2Xg616/31800-annotations.png\" alt=\"31800\"></p></li>\n<li><p>I think there will be a middle scale shake up at the end since  all of the hidden test set images are taken from another source (hubmap). This is the reason why I really liked this competition. Teams that can simulate different pixel sizes and tissue thicknesses will probably win this competition.</p></li>\n</ul>\n<p>Those are my initial observations. Good luck to everyone!</p>",
      "rawMarkdown": "* There are 351 images and 4909 annotations. Dataset is relatively smaller compared to similar competitions but there are still fair amount of annotations.\n* Annotations are given as polygons and rle masks. In some of the discussions, organizers mentioned that polygons are more raw so I guess rle masks are derived from them. I checked polygons and masks side by side and they appear to be consistent. This allows us to try different approaches such as instance segmentation, semantic segmentation and even domain specific algorithms such as cellpose.\n* I checked some of the images and annotations. Annotation quality is not pixel perfect but it looks okay so far. Some of the images may have incorrect annotations like the ones below.\n\nRed lines are polygons and yellow areas are masks.\n\nSmall line in the middle is probably an incorrect annotation.\n![4301](https://i.ibb.co/hB31Fms/4301-annotations.png)\n\nAnnotation at the bottom looks pretty bad, especially its bottom edge.\n![6021](https://i.ibb.co/kcp1DDm/6021-annotations.png)\n\n* Since there is only one class and evaluation metric is dice coefficient, you don't have to worry about overlaps.\n\n* Some of the images might require extra cleaning steps. For instance:\n![31800](https://i.ibb.co/z2Xg616/31800-annotations.png)\n\n* I think there will be a middle scale shake up at the end since ~~some~~ all of the hidden test set images are taken from another source (hubmap). This is the reason why I really liked this competition. Teams that can simulate different pixel sizes and tissue thicknesses will probably win this competition.\n\nThose are my initial observations. Good luck to everyone!",
      "votes": null
    },
    {
      "id": "1851797",
      "postDate": "07/11/2022 14:45:46",
      "content": "<p>little correction: <code>some of the hidden test set images are taken from another source (hubmap)</code> actually <code>75% of public test images and all of the hidden test images in private leaderboard are taken from hubmap</code> </p>",
      "rawMarkdown": "little correction: `some of the hidden test set images are taken from another source (hubmap)` actually `75% of public test images and all of the hidden test images in private leaderboard are taken from hubmap`",
      "votes": null
    },
    {
      "id": "1852566",
      "postDate": "07/12/2022 06:48:01",
      "content": "<p>how can we simulate different tissue thickness and pixel sizes? can you please give a few pointers on this?</p>",
      "rawMarkdown": "how can we simulate different tissue thickness and pixel sizes? can you please give a few pointers on this?",
      "votes": null
    },
    {
      "id": "1852597",
      "postDate": "07/12/2022 07:16:50",
      "content": "<p>I don't know much about tissue thickness but different pixel sizes can be done with scale jitter.</p>",
      "rawMarkdown": "I don't know much about tissue thickness but different pixel sizes can be done with scale jitter.",
      "votes": null
    },
    {
      "id": "1853920",
      "postDate": "07/13/2022 09:47:20",
      "content": "<p><a href=\"https://www.kaggle.com/gunesevitan\" target=\"_blank\">@gunesevitan</a> Thank you for writing this post. </p>\n<p>As for <code>simulating tissue thickness</code> part of your post I have glimpsed over this paper: <a href=\"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3018007/\" target=\"_blank\">Next-generation acceleration and code optimization for light transport in turbid media using GPUs</a>. In the abstract they stated that <code>In biomedical optics, the MC method is the gold standard approach for simulating light transport in biological tissue, both due to its accuracy and its flexibility in modelling realistic, heterogeneous tissue geometry in 3-D.</code> which could <strong>maybe</strong> be used for simulating tissue thickness. As I have not read it completely I can not say for sure now. They have also provided software they used both on <a href=\"https://github.com/ominux/gpumcml\" target=\"_blank\">github</a> and <a href=\"https://code.google.com/archive/p/gpumcml/\" target=\"_blank\">google code archive</a>.</p>\n<p>By the way I have ran across the said paper in the following paper: <a href=\"https://opg.optica.org/boe/fulltext.cfm?uri=boe-4-7-1176&amp;id=258001\" target=\"_blank\">Estimating soft tissue thickness from light-tissue interactions––a simulation study</a>. In the <em>Methods</em> -&gt; <em>Monte Carlo simulation and tissue model</em> section.</p>\n<p>Have fun!</p>",
      "rawMarkdown": "gunesevitan Thank you for writing this post. \n\nAs for `simulating tissue thickness` part of your post I have glimpsed over this paper: [Next-generation acceleration and code optimization for light transport in turbid media using GPUs](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3018007/). In the abstract they stated that `In biomedical optics, the MC method is the gold standard approach for simulating light transport in biological tissue, both due to its accuracy and its flexibility in modelling realistic, heterogeneous tissue geometry in 3-D.` which could **maybe** be used for simulating tissue thickness. As I have not read it completely I can not say for sure now. They have also provided software they used both on [github](https://github.com/ominux/gpumcml) and [google code archive](https://code.google.com/archive/p/gpumcml/).\n\nBy the way I have ran across the said paper in the following paper: [Estimating soft tissue thickness from light-tissue interactions––a simulation study](https://opg.optica.org/boe/fulltext.cfm?uri=boe-4-7-1176&id=258001). In the *Methods* -> *Monte Carlo simulation and tissue model* section.\n\nHave fun!",
      "votes": null
    },
    {
      "id": "1871053",
      "postDate": "07/26/2022 03:45:30",
      "content": "<p>Regarding the \"extra cleaning image\", you might not need to clean that image because it could resemble a shifted or cropped image.</p>",
      "rawMarkdown": "Regarding the \"extra cleaning image\", you might not need to clean that image because it could resemble a shifted or cropped image.",
      "votes": null
    },
    {
      "id": "1871242",
      "postDate": "07/26/2022 07:01:24",
      "content": "<p><a href=\"https://www.kaggle.com/nishantbhansali\" target=\"_blank\">@nishantbhansali</a>  concerning the tissue thickness there is a notebook displaying impact of tissue size on staining: <a href=\"https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/333083\" target=\"_blank\">https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/333083</a></p>\n<p>and as pictures are taken with a light microscope, it makes sense that the thicker the tissue, the less light passes through, making it darker (except the non tissue parts of course).</p>\n<p><a href=\"https://www.kaggle.com/gunesevitan\" target=\"_blank\">@gunesevitan</a> what do you mean with necessary cleaning steps in the 3rd image?</p>",
      "rawMarkdown": "nishantbhansali  concerning the tissue thickness there is a notebook displaying impact of tissue size on staining: https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/333083\n\nand as pictures are taken with a light microscope, it makes sense that the thicker the tissue, the less light passes through, making it darker (except the non tissue parts of course).\n\n@gunesevitan what do you mean with necessary cleaning steps in the 3rd image?",
      "votes": null
    },
    {
      "id": "1879628",
      "postDate": "08/01/2022 07:10:19",
      "content": "<p>You are right. It is kind of an implicit augmentation.</p>",
      "rawMarkdown": "You are right. It is kind of an implicit augmentation.",
      "votes": null
    },
    {
      "id": "1888700",
      "postDate": "08/07/2022 18:20:33",
      "content": "<p>I shared a notebook about simulating different pixel sizes and tissue thicknesses.</p>\n<p><a href=\"https://www.kaggle.com/code/gunesevitan/pixel-size-and-tissue-thickness-domain-adaptation\" target=\"_blank\">https://www.kaggle.com/code/gunesevitan/pixel-size-and-tissue-thickness-domain-adaptation</a></p>",
      "rawMarkdown": "I shared a notebook about simulating different pixel sizes and tissue thicknesses.\n\nhttps://www.kaggle.com/code/gunesevitan/pixel-size-and-tissue-thickness-domain-adaptation",
      "votes": null
    },
    {
      "id": "1934033",
      "postDate": "09/11/2022 05:18:22",
      "content": "<p>Hello, have you tried cellpose？</p>",
      "rawMarkdown": "Hello, have you tried cellpose？",
      "votes": null
    },
    {
      "id": "1934056",
      "postDate": "09/11/2022 05:39:54",
      "content": "<p>Nope, I haven't tried it yet.</p>",
      "rawMarkdown": "Nope, I haven't tried it yet.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1851797,
      "author_name": "muhammad4hmed",
      "author_url": "",
      "post_date": "07/11/2022 14:45:46",
      "content": "<p>little correction: <code>some of the hidden test set images are taken from another source (hubmap)</code> actually <code>75% of public test images and all of the hidden test images in private leaderboard are taken from hubmap</code> </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1852566,
      "author_name": "nishantbhansali",
      "author_url": "",
      "post_date": "07/12/2022 06:48:01",
      "content": "<p>how can we simulate different tissue thickness and pixel sizes? can you please give a few pointers on this?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1852597,
          "author_name": "gunesevitan",
          "author_url": "",
          "post_date": "07/12/2022 07:16:50",
          "content": "<p>I don't know much about tissue thickness but different pixel sizes can be done with scale jitter.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1853920,
      "author_name": "temuujinerdene",
      "author_url": "",
      "post_date": "07/13/2022 09:47:20",
      "content": "<p><a href=\"https://www.kaggle.com/gunesevitan\" target=\"_blank\">@gunesevitan</a> Thank you for writing this post. </p>\n<p>As for <code>simulating tissue thickness</code> part of your post I have glimpsed over this paper: <a href=\"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3018007/\" target=\"_blank\">Next-generation acceleration and code optimization for light transport in turbid media using GPUs</a>. In the abstract they stated that <code>In biomedical optics, the MC method is the gold standard approach for simulating light transport in biological tissue, both due to its accuracy and its flexibility in modelling realistic, heterogeneous tissue geometry in 3-D.</code> which could <strong>maybe</strong> be used for simulating tissue thickness. As I have not read it completely I can not say for sure now. They have also provided software they used both on <a href=\"https://github.com/ominux/gpumcml\" target=\"_blank\">github</a> and <a href=\"https://code.google.com/archive/p/gpumcml/\" target=\"_blank\">google code archive</a>.</p>\n<p>By the way I have ran across the said paper in the following paper: <a href=\"https://opg.optica.org/boe/fulltext.cfm?uri=boe-4-7-1176&amp;id=258001\" target=\"_blank\">Estimating soft tissue thickness from light-tissue interactions––a simulation study</a>. In the <em>Methods</em> -&gt; <em>Monte Carlo simulation and tissue model</em> section.</p>\n<p>Have fun!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1871053,
      "author_name": "datadote",
      "author_url": "",
      "post_date": "07/26/2022 03:45:30",
      "content": "<p>Regarding the \"extra cleaning image\", you might not need to clean that image because it could resemble a shifted or cropped image.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1879628,
          "author_name": "gunesevitan",
          "author_url": "",
          "post_date": "08/01/2022 07:10:19",
          "content": "<p>You are right. It is kind of an implicit augmentation.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1871242,
      "author_name": "bridgeoverwater",
      "author_url": "",
      "post_date": "07/26/2022 07:01:24",
      "content": "<p><a href=\"https://www.kaggle.com/nishantbhansali\" target=\"_blank\">@nishantbhansali</a>  concerning the tissue thickness there is a notebook displaying impact of tissue size on staining: <a href=\"https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/333083\" target=\"_blank\">https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/333083</a></p>\n<p>and as pictures are taken with a light microscope, it makes sense that the thicker the tissue, the less light passes through, making it darker (except the non tissue parts of course).</p>\n<p><a href=\"https://www.kaggle.com/gunesevitan\" target=\"_blank\">@gunesevitan</a> what do you mean with necessary cleaning steps in the 3rd image?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1888700,
      "author_name": "gunesevitan",
      "author_url": "",
      "post_date": "08/07/2022 18:20:33",
      "content": "<p>I shared a notebook about simulating different pixel sizes and tissue thicknesses.</p>\n<p><a href=\"https://www.kaggle.com/code/gunesevitan/pixel-size-and-tissue-thickness-domain-adaptation\" target=\"_blank\">https://www.kaggle.com/code/gunesevitan/pixel-size-and-tissue-thickness-domain-adaptation</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1934033,
      "author_name": "wangqihanginthesky",
      "author_url": "",
      "post_date": "09/11/2022 05:18:22",
      "content": "<p>Hello, have you tried cellpose？</p>",
      "votes": null,
      "replies": [
        {
          "id": 1934056,
          "author_name": "gunesevitan",
          "author_url": "",
          "post_date": "09/11/2022 05:39:54",
          "content": "<p>Nope, I haven't tried it yet.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1851532": "* There are 351 images and 4909 annotations. Dataset is relatively smaller compared to similar competitions but there are still fair amount of annotations.\n* Annotations are given as polygons and rle masks. In some of the discussions, organizers mentioned that polygons are more raw so I guess rle masks are derived from them. I checked polygons and masks side by side and they appear to be consistent. This allows us to try different approaches such as instance segmentation, semantic segmentation and even domain specific algorithms such as cellpose.\n* I checked some of the images and annotations. Annotation quality is not pixel perfect but it looks okay so far. Some of the images may have incorrect annotations like the ones below.\n\nRed lines are polygons and yellow areas are masks.\n\nSmall line in the middle is probably an incorrect annotation.\n![4301](https://i.ibb.co/hB31Fms/4301-annotations.png)\n\nAnnotation at the bottom looks pretty bad, especially its bottom edge.\n![6021](https://i.ibb.co/kcp1DDm/6021-annotations.png)\n\n* Since there is only one class and evaluation metric is dice coefficient, you don't have to worry about overlaps.\n\n* Some of the images might require extra cleaning steps. For instance:\n![31800](https://i.ibb.co/z2Xg616/31800-annotations.png)\n\n* I think there will be a middle scale shake up at the end since ~~some~~ all of the hidden test set images are taken from another source (hubmap). This is the reason why I really liked this competition. Teams that can simulate different pixel sizes and tissue thicknesses will probably win this competition.\n\nThose are my initial observations. Good luck to everyone!",
    "1851797": "little correction: `some of the hidden test set images are taken from another source (hubmap)` actually `75% of public test images and all of the hidden test images in private leaderboard are taken from hubmap`",
    "1852566": "how can we simulate different tissue thickness and pixel sizes? can you please give a few pointers on this?",
    "1852597": "I don't know much about tissue thickness but different pixel sizes can be done with scale jitter.",
    "1853920": "gunesevitan Thank you for writing this post. \n\nAs for `simulating tissue thickness` part of your post I have glimpsed over this paper: [Next-generation acceleration and code optimization for light transport in turbid media using GPUs](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3018007/). In the abstract they stated that `In biomedical optics, the MC method is the gold standard approach for simulating light transport in biological tissue, both due to its accuracy and its flexibility in modelling realistic, heterogeneous tissue geometry in 3-D.` which could **maybe** be used for simulating tissue thickness. As I have not read it completely I can not say for sure now. They have also provided software they used both on [github](https://github.com/ominux/gpumcml) and [google code archive](https://code.google.com/archive/p/gpumcml/).\n\nBy the way I have ran across the said paper in the following paper: [Estimating soft tissue thickness from light-tissue interactions––a simulation study](https://opg.optica.org/boe/fulltext.cfm?uri=boe-4-7-1176&id=258001). In the *Methods* -> *Monte Carlo simulation and tissue model* section.\n\nHave fun!",
    "1871053": "Regarding the \"extra cleaning image\", you might not need to clean that image because it could resemble a shifted or cropped image.",
    "1871242": "nishantbhansali  concerning the tissue thickness there is a notebook displaying impact of tissue size on staining: https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/333083\n\nand as pictures are taken with a light microscope, it makes sense that the thicker the tissue, the less light passes through, making it darker (except the non tissue parts of course).\n\n@gunesevitan what do you mean with necessary cleaning steps in the 3rd image?",
    "1879628": "You are right. It is kind of an implicit augmentation.",
    "1888700": "I shared a notebook about simulating different pixel sizes and tissue thicknesses.\n\nhttps://www.kaggle.com/code/gunesevitan/pixel-size-and-tissue-thickness-domain-adaptation",
    "1934033": "Hello, have you tried cellpose？",
    "1934056": "Nope, I haven't tried it yet."
  },
  "source": "meta"
}