{
  "id": 68983,
  "title": "some questions about dataset instrumentation, image properties",
  "url": "/competitions/human-protein-atlas-image-classification/discussion/68983",
  "author_name": "",
  "post_date": "2018-10-19T07:38:53.281297100Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Human Protein (Cell) Atlas questions:</p>\n\n<h2>1 Immunofluorescence:</h2>\n\n<h1>1.1 (target protein selective) fluorescent antibodies:</h1>\n\n<p><strong>Question A</strong>: intrinsically fluorescent? or fluorescent after modification? (out of curiosity)</p>\n\n<p><strong>Question B</strong>: fluorescent antibodies introduced internally (modified genome to express these antibodies) ? or fluorescent antibodies introduced externally (fluorescent antibody stain applied from outside the cell culture, then perhaps washed/rinsed)? (complications: what if a target protein is in fact present in a dark structure, but the fluorescent antibody is unable to reach the protein because it fails to pass relevant membrane pores? what if the antibody is slowly diffusing into a relevant vacuole with the target protein, but destroyed or exported at a greater rate than it is imported?)</p>\n\n<h1>1.2 target proteins selectivity</h1>\n\n<p><strong>Question C</strong>: how do we make sure that the fluorescent antibody does not display affinity towards the other proteins? or is this excluded on the basis of information theoretic arguments reducing the probability to a negligible level? (curiosity)</p>\n\n<h2>2 Confocal Microscopy</h2>\n\n<h1>2.1 Imaging geometry</h1>\n\n<p>With \"pure\" I mean all images of the dataset being of same type, as opposed to mixed image types.</p>\n\n<p><strong>Question D</strong>: are the images: 1) purely constant z slices/sections of a 3D z-stack 2) purely merged 2D images 3) mixed image types (both slices and merged images occur in the dataset) ? (crucial for the more classical image processing I am attempting)</p>\n\n<p><strong>Question E</strong>: If (some or all) images are merged, how are they merged: 1) maximum pixel value? 2) standard deviation of pixel value? 3) sum of pixel values for the same column of 3D stack 4) other</p>\n\n<p><strong>Question F</strong>: Is the magnification: 1) constant with depth (isometric projection) or 2) roughly inversely proportional to depth (perspective projection)?</p>\n\n<p><strong>Question G</strong>: For classification and or object detection, I am convinced a lot of information is lost in the 3D to 2D conversion, it seems almost trivial to me to group pixels in a 3D stack of slices by using the 3D equivalent of the <em>flood fill algorithm</em> such that individual structures are \"selectable\" and their indivudual statistics useful for classification. Why are we given collapsed 2D images for an intrinsically 3D imaging instrument? This is making classification much harder than necessary.</p>\n\n<h1>2.2 Pixel value properties:</h1>\n\n<p><strong>Question H</strong>: assuming either D=1 (purely constant z-slices) or E=3 (sum of pixel values), are the 8-bit grayscale values 1) in the linear domain proportional to the number of photons (i.e. a pixel value of 2X has received twice the amount of photons compared to a pixel value of X)? or 2) in the gammma compressed domain (the previous property no longer holds)?</p>\n\n<p><strong>Question I</strong>: if gamma compressed, what gamma value was used?</p>\n\n<p><strong>Question J</strong>: if not gamma compressed, were the values compressed before displaying for the expert annotators? or did their display system assume the non-compressed values were compressed (de-emphasizing dark pixels, and emphasizing bright pixels)?</p>\n\n<p>Thanks for any clarifications :D</p>",
  "messages": [
    {
      "id": "406407",
      "postDate": "10/19/2018 07:38:53",
      "content": "<p>Human Protein (Cell) Atlas questions:</p>\n\n<h2>1 Immunofluorescence:</h2>\n\n<h1>1.1 (target protein selective) fluorescent antibodies:</h1>\n\n<p><strong>Question A</strong>: intrinsically fluorescent? or fluorescent after modification? (out of curiosity)</p>\n\n<p><strong>Question B</strong>: fluorescent antibodies introduced internally (modified genome to express these antibodies) ? or fluorescent antibodies introduced externally (fluorescent antibody stain applied from outside the cell culture, then perhaps washed/rinsed)? (complications: what if a target protein is in fact present in a dark structure, but the fluorescent antibody is unable to reach the protein because it fails to pass relevant membrane pores? what if the antibody is slowly diffusing into a relevant vacuole with the target protein, but destroyed or exported at a greater rate than it is imported?)</p>\n\n<h1>1.2 target proteins selectivity</h1>\n\n<p><strong>Question C</strong>: how do we make sure that the fluorescent antibody does not display affinity towards the other proteins? or is this excluded on the basis of information theoretic arguments reducing the probability to a negligible level? (curiosity)</p>\n\n<h2>2 Confocal Microscopy</h2>\n\n<h1>2.1 Imaging geometry</h1>\n\n<p>With \"pure\" I mean all images of the dataset being of same type, as opposed to mixed image types.</p>\n\n<p><strong>Question D</strong>: are the images: 1) purely constant z slices/sections of a 3D z-stack 2) purely merged 2D images 3) mixed image types (both slices and merged images occur in the dataset) ? (crucial for the more classical image processing I am attempting)</p>\n\n<p><strong>Question E</strong>: If (some or all) images are merged, how are they merged: 1) maximum pixel value? 2) standard deviation of pixel value? 3) sum of pixel values for the same column of 3D stack 4) other</p>\n\n<p><strong>Question F</strong>: Is the magnification: 1) constant with depth (isometric projection) or 2) roughly inversely proportional to depth (perspective projection)?</p>\n\n<p><strong>Question G</strong>: For classification and or object detection, I am convinced a lot of information is lost in the 3D to 2D conversion, it seems almost trivial to me to group pixels in a 3D stack of slices by using the 3D equivalent of the <em>flood fill algorithm</em> such that individual structures are \"selectable\" and their indivudual statistics useful for classification. Why are we given collapsed 2D images for an intrinsically 3D imaging instrument? This is making classification much harder than necessary.</p>\n\n<h1>2.2 Pixel value properties:</h1>\n\n<p><strong>Question H</strong>: assuming either D=1 (purely constant z-slices) or E=3 (sum of pixel values), are the 8-bit grayscale values 1) in the linear domain proportional to the number of photons (i.e. a pixel value of 2X has received twice the amount of photons compared to a pixel value of X)? or 2) in the gammma compressed domain (the previous property no longer holds)?</p>\n\n<p><strong>Question I</strong>: if gamma compressed, what gamma value was used?</p>\n\n<p><strong>Question J</strong>: if not gamma compressed, were the values compressed before displaying for the expert annotators? or did their display system assume the non-compressed values were compressed (de-emphasizing dark pixels, and emphasizing bright pixels)?</p>\n\n<p>Thanks for any clarifications :D</p>",
      "rawMarkdown": "Human Protein (Cell) Atlas questions:\n\n1 Immunofluorescence:\n---------------------\n\n1.1 (target protein selective) fluorescent antibodies:\n======================================================\n\n**Question A**: intrinsically fluorescent? or fluorescent after modification? (out of curiosity)\n\n**Question B**: fluorescent antibodies introduced internally (modified genome to express these antibodies) ? or fluorescent antibodies introduced externally (fluorescent antibody stain applied from outside the cell culture, then perhaps washed/rinsed)? (complications: what if a target protein is in fact present in a dark structure, but the fluorescent antibody is unable to reach the protein because it fails to pass relevant membrane pores? what if the antibody is slowly diffusing into a relevant vacuole with the target protein, but destroyed or exported at a greater rate than it is imported?)\n\n1.2 target proteins selectivity\n===============================\n\n**Question C**: how do we make sure that the fluorescent antibody does not display affinity towards the other proteins? or is this excluded on the basis of information theoretic arguments reducing the probability to a negligible level? (curiosity)\n\n2 Confocal Microscopy\n---------------------\n\n2.1 Imaging geometry\n====================\n\nWith \"pure\" I mean all images of the dataset being of same type, as opposed to mixed image types.\n\n**Question D**: are the images: 1) purely constant z slices/sections of a 3D z-stack 2) purely merged 2D images 3) mixed image types (both slices and merged images occur in the dataset) ? (crucial for the more classical image processing I am attempting)\n\n**Question E**: If (some or all) images are merged, how are they merged: 1) maximum pixel value? 2) standard deviation of pixel value? 3) sum of pixel values for the same column of 3D stack 4) other\n\n**Question F**: Is the magnification: 1) constant with depth (isometric projection) or 2) roughly inversely proportional to depth (perspective projection)?\n\n**Question G**: For classification and or object detection, I am convinced a lot of information is lost in the 3D to 2D conversion, it seems almost trivial to me to group pixels in a 3D stack of slices by using the 3D equivalent of the *flood fill algorithm* such that individual structures are \"selectable\" and their indivudual statistics useful for classification. Why are we given collapsed 2D images for an intrinsically 3D imaging instrument? This is making classification much harder than necessary.\n\n2.2 Pixel value properties:\n===========================\n\n**Question H**: assuming either D=1 (purely constant z-slices) or E=3 (sum of pixel values), are the 8-bit grayscale values 1) in the linear domain proportional to the number of photons (i.e. a pixel value of 2X has received twice the amount of photons compared to a pixel value of X)? or 2) in the gammma compressed domain (the previous property no longer holds)?\n\n**Question I**: if gamma compressed, what gamma value was used?\n\n**Question J**: if not gamma compressed, were the values compressed before displaying for the expert annotators? or did their display system assume the non-compressed values were compressed (de-emphasizing dark pixels, and emphasizing bright pixels)?\n\nThanks for any clarifications :D",
      "votes": null
    },
    {
      "id": "406600",
      "postDate": "10/19/2018 13:58:01",
      "content": "<p>Hello, and thank you for your questions.</p>\n\n<p>A,B: I hope that the described method on our webpage (<a href=\"https://www.proteinatlas.org/learn/method/immunocytochemistry\">https://www.proteinatlas.org/learn/method/immunocytochemistry</a>) will help. If it still unclear, please do not hesitate to ask.</p>\n\n<p>C: Validation that the antibodies are targeting the correct structure is indeed an important problem within our field. At the HPA we are actively working to make sure that we have antibodies that are targeting the correct proteins. You can read more about our work on that at <a href=\"https://www.proteinatlas.org/about/antibody+validation\">https://www.proteinatlas.org/about/antibody+validation</a>.</p>\n\n<p>D (,E,F,G,H,I): We provide single slice images, non-merged. No gamma correction is made. You can read more about our acquisition and annotation process at: <a href=\"https://www.proteinatlas.org/about/assays+annotation#if\">https://www.proteinatlas.org/about/assays+annotation#if</a></p>\n\n<p>J:  images were converted to jpg made before showing them to the experts. No other compression was made.</p>",
      "rawMarkdown": "Hello, and thank you for your questions.\n\nA,B: I hope that the described method on our webpage (https://www.proteinatlas.org/learn/method/immunocytochemistry) will help. If it still unclear, please do not hesitate to ask.\n\nC: Validation that the antibodies are targeting the correct structure is indeed an important problem within our field. At the HPA we are actively working to make sure that we have antibodies that are targeting the correct proteins. You can read more about our work on that at https://www.proteinatlas.org/about/antibody+validation.\n\nD (,E,F,G,H,I): We provide single slice images, non-merged. No gamma correction is made. You can read more about our acquisition and annotation process at: https://www.proteinatlas.org/about/assays+annotation#if\n\nJ:  images were converted to jpg made before showing them to the experts. No other compression was made.",
      "votes": null
    },
    {
      "id": "407295",
      "postDate": "10/20/2018 20:37:01",
      "content": "<p>thanks for answering my questions!</p>",
      "rawMarkdown": "thanks for answering my questions!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 406600,
      "author_name": "cwinsnes",
      "author_url": "",
      "post_date": "10/19/2018 13:58:01",
      "content": "<p>Hello, and thank you for your questions.</p>\n\n<p>A,B: I hope that the described method on our webpage (<a href=\"https://www.proteinatlas.org/learn/method/immunocytochemistry\">https://www.proteinatlas.org/learn/method/immunocytochemistry</a>) will help. If it still unclear, please do not hesitate to ask.</p>\n\n<p>C: Validation that the antibodies are targeting the correct structure is indeed an important problem within our field. At the HPA we are actively working to make sure that we have antibodies that are targeting the correct proteins. You can read more about our work on that at <a href=\"https://www.proteinatlas.org/about/antibody+validation\">https://www.proteinatlas.org/about/antibody+validation</a>.</p>\n\n<p>D (,E,F,G,H,I): We provide single slice images, non-merged. No gamma correction is made. You can read more about our acquisition and annotation process at: <a href=\"https://www.proteinatlas.org/about/assays+annotation#if\">https://www.proteinatlas.org/about/assays+annotation#if</a></p>\n\n<p>J:  images were converted to jpg made before showing them to the experts. No other compression was made.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 407295,
      "author_name": "ludwigmaes",
      "author_url": "",
      "post_date": "10/20/2018 20:37:01",
      "content": "<p>thanks for answering my questions!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "406407": "Human Protein (Cell) Atlas questions:\n\n1 Immunofluorescence:\n---------------------\n\n1.1 (target protein selective) fluorescent antibodies:\n======================================================\n\n**Question A**: intrinsically fluorescent? or fluorescent after modification? (out of curiosity)\n\n**Question B**: fluorescent antibodies introduced internally (modified genome to express these antibodies) ? or fluorescent antibodies introduced externally (fluorescent antibody stain applied from outside the cell culture, then perhaps washed/rinsed)? (complications: what if a target protein is in fact present in a dark structure, but the fluorescent antibody is unable to reach the protein because it fails to pass relevant membrane pores? what if the antibody is slowly diffusing into a relevant vacuole with the target protein, but destroyed or exported at a greater rate than it is imported?)\n\n1.2 target proteins selectivity\n===============================\n\n**Question C**: how do we make sure that the fluorescent antibody does not display affinity towards the other proteins? or is this excluded on the basis of information theoretic arguments reducing the probability to a negligible level? (curiosity)\n\n2 Confocal Microscopy\n---------------------\n\n2.1 Imaging geometry\n====================\n\nWith \"pure\" I mean all images of the dataset being of same type, as opposed to mixed image types.\n\n**Question D**: are the images: 1) purely constant z slices/sections of a 3D z-stack 2) purely merged 2D images 3) mixed image types (both slices and merged images occur in the dataset) ? (crucial for the more classical image processing I am attempting)\n\n**Question E**: If (some or all) images are merged, how are they merged: 1) maximum pixel value? 2) standard deviation of pixel value? 3) sum of pixel values for the same column of 3D stack 4) other\n\n**Question F**: Is the magnification: 1) constant with depth (isometric projection) or 2) roughly inversely proportional to depth (perspective projection)?\n\n**Question G**: For classification and or object detection, I am convinced a lot of information is lost in the 3D to 2D conversion, it seems almost trivial to me to group pixels in a 3D stack of slices by using the 3D equivalent of the *flood fill algorithm* such that individual structures are \"selectable\" and their indivudual statistics useful for classification. Why are we given collapsed 2D images for an intrinsically 3D imaging instrument? This is making classification much harder than necessary.\n\n2.2 Pixel value properties:\n===========================\n\n**Question H**: assuming either D=1 (purely constant z-slices) or E=3 (sum of pixel values), are the 8-bit grayscale values 1) in the linear domain proportional to the number of photons (i.e. a pixel value of 2X has received twice the amount of photons compared to a pixel value of X)? or 2) in the gammma compressed domain (the previous property no longer holds)?\n\n**Question I**: if gamma compressed, what gamma value was used?\n\n**Question J**: if not gamma compressed, were the values compressed before displaying for the expert annotators? or did their display system assume the non-compressed values were compressed (de-emphasizing dark pixels, and emphasizing bright pixels)?\n\nThanks for any clarifications :D",
    "406600": "Hello, and thank you for your questions.\n\nA,B: I hope that the described method on our webpage (https://www.proteinatlas.org/learn/method/immunocytochemistry) will help. If it still unclear, please do not hesitate to ask.\n\nC: Validation that the antibodies are targeting the correct structure is indeed an important problem within our field. At the HPA we are actively working to make sure that we have antibodies that are targeting the correct proteins. You can read more about our work on that at https://www.proteinatlas.org/about/antibody+validation.\n\nD (,E,F,G,H,I): We provide single slice images, non-merged. No gamma correction is made. You can read more about our acquisition and annotation process at: https://www.proteinatlas.org/about/assays+annotation#if\n\nJ:  images were converted to jpg made before showing them to the experts. No other compression was made.",
    "407295": "thanks for answering my questions!"
  },
  "source": "meta"
}