{
  "id": 215986,
  "title": "Insights from the previous HPA competition (video)",
  "url": "/competitions/hpa-single-cell-image-classification/discussion/215986",
  "author_name": "Darek Kłeczek",
  "post_date": "2021-02-01T06:25:14.713000",
  "votes": 18,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I found this Kaggle Meetup video from the previous HPA competition very helpful: <a href=\"https://youtu.be/bAgjqdQ0VE8\" target=\"_blank\">https://youtu.be/bAgjqdQ0VE8</a></p>\n<p>My notes: </p>\n<ul>\n<li>Images are produced by green fluorescent protein (discovered in jellyfish) - we can take the gene for this, copy it, and then put it in another cell, at the end of another gene for protein. Whenever that protein is produced in a cell, this dye will be produced with it. We can then visualize if that protein is being produced, and its location. If we know where the protein is located in the cell, we can deduce its function as well</li>\n<li>Images in the competition use four of these dyes: the red/green/blue/yellow channels don’t represent a spectrum of colors, but rather a part of a cell that is stained with the relevant dye. Red stands for microtubules, blue for nucleoplasm and yellow for endoplasmic reticulum. These are used as reference points vs. green which marks the protein of interest. </li>\n<li>In the previous HPA competition, the majority of the top competitors dropped the yellow “channel” and used RGB instead of RGBY, without affecting their scores. Theory: microtubules and ER are typically in the same location, so it didn’t add much information and was safe to skip. </li>\n<li>Lots of external data can be used. </li>\n<li>In the previous HPA competition, most participants used Focal Loss, which should help with class imbalance. </li>\n</ul>",
  "messages": [
    {
      "id": 1180182,
      "postDate": "2021-02-01T06:25:14.713Z",
      "content": "<p>I found this Kaggle Meetup video from the previous HPA competition very helpful: <a href=\"https://youtu.be/bAgjqdQ0VE8\" target=\"_blank\">https://youtu.be/bAgjqdQ0VE8</a></p>\n<p>My notes: </p>\n<ul>\n<li>Images are produced by green fluorescent protein (discovered in jellyfish) - we can take the gene for this, copy it, and then put it in another cell, at the end of another gene for protein. Whenever that protein is produced in a cell, this dye will be produced with it. We can then visualize if that protein is being produced, and its location. If we know where the protein is located in the cell, we can deduce its function as well</li>\n<li>Images in the competition use four of these dyes: the red/green/blue/yellow channels don’t represent a spectrum of colors, but rather a part of a cell that is stained with the relevant dye. Red stands for microtubules, blue for nucleoplasm and yellow for endoplasmic reticulum. These are used as reference points vs. green which marks the protein of interest. </li>\n<li>In the previous HPA competition, the majority of the top competitors dropped the yellow “channel” and used RGB instead of RGBY, without affecting their scores. Theory: microtubules and ER are typically in the same location, so it didn’t add much information and was safe to skip. </li>\n<li>Lots of external data can be used. </li>\n<li>In the previous HPA competition, most participants used Focal Loss, which should help with class imbalance. </li>\n</ul>",
      "rawMarkdown": "I found this Kaggle Meetup video from the previous HPA competition very helpful: https://youtu.be/bAgjqdQ0VE8\n\nMy notes: \n- Images are produced by green fluorescent protein (discovered in jellyfish) - we can take the gene for this, copy it, and then put it in another cell, at the end of another gene for protein. Whenever that protein is produced in a cell, this dye will be produced with it. We can then visualize if that protein is being produced, and its location. If we know where the protein is located in the cell, we can deduce its function as well\n- Images in the competition use four of these dyes: the red/green/blue/yellow channels don’t represent a spectrum of colors, but rather a part of a cell that is stained with the relevant dye. Red stands for microtubules, blue for nucleoplasm and yellow for endoplasmic reticulum. These are used as reference points vs. green which marks the protein of interest. \n- In the previous HPA competition, the majority of the top competitors dropped the yellow “channel” and used RGB instead of RGBY, without affecting their scores. Theory: microtubules and ER are typically in the same location, so it didn’t add much information and was safe to skip. \n- Lots of external data can be used. \n- In the previous HPA competition, most participants used Focal Loss, which should help with class imbalance. ",
      "votes": 18
    },
    {
      "id": 1264155,
      "postDate": "2021-04-05T23:57:20.937Z",
      "content": "<p>Thanks for this, I have been dropping the yellow channel because it looks like the red one in the images I checked, but it's nice to get some more confirmation on this.</p>",
      "rawMarkdown": "Thanks for this, I have been dropping the yellow channel because it looks like the red one in the images I checked, but it's nice to get some more confirmation on this."
    }
  ],
  "comments": [
    {
      "id": 1264155,
      "author_name": "Jason O. Jensen",
      "author_url": "",
      "post_date": "2021-04-05T23:57:20.937000",
      "content": "<p>Thanks for this, I have been dropping the yellow channel because it looks like the red one in the images I checked, but it's nice to get some more confirmation on this.</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1180182": "I found this Kaggle Meetup video from the previous HPA competition very helpful: https://youtu.be/bAgjqdQ0VE8\n\nMy notes: \n- Images are produced by green fluorescent protein (discovered in jellyfish) - we can take the gene for this, copy it, and then put it in another cell, at the end of another gene for protein. Whenever that protein is produced in a cell, this dye will be produced with it. We can then visualize if that protein is being produced, and its location. If we know where the protein is located in the cell, we can deduce its function as well\n- Images in the competition use four of these dyes: the red/green/blue/yellow channels don’t represent a spectrum of colors, but rather a part of a cell that is stained with the relevant dye. Red stands for microtubules, blue for nucleoplasm and yellow for endoplasmic reticulum. These are used as reference points vs. green which marks the protein of interest. \n- In the previous HPA competition, the majority of the top competitors dropped the yellow “channel” and used RGB instead of RGBY, without affecting their scores. Theory: microtubules and ER are typically in the same location, so it didn’t add much information and was safe to skip. \n- Lots of external data can be used. \n- In the previous HPA competition, most participants used Focal Loss, which should help with class imbalance. ",
    "1264155": "Thanks for this, I have been dropping the yellow channel because it looks like the red one in the images I checked, but it's nice to get some more confirmation on this."
  }
}