{
  "id": 226369,
  "title": "Sub-challenge: Locate the centrosome using the microtubules",
  "url": "/competitions/hpa-single-cell-image-classification/discussion/226369",
  "author_name": "Philip Hucklesby",
  "post_date": "2021-03-16T09:34:40.319000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi All,<br>\nI would like to throw out this little sub-challenge I have been thinking about.</p>\n<p>Much of the HPA competition seems to be tractable to non-machine-learning computer vision techniques, so that some extra effort in hand-coded feature detection could potentially solve, say 80% of the problem with 20% of the CPU/GPU power (see discussion here <a href=\"https://www.kaggle.com/lukemerrick/who-needs-a-gpu-shallow-learning-in-julia)\" target=\"_blank\">https://www.kaggle.com/lukemerrick/who-needs-a-gpu-shallow-learning-in-julia)</a>.</p>\n<p>So what would we do with the other 80% CPU/GPU capacity ?</p>\n<p>How about this one ?</p>\n<p><strong>Given the red microtubules layer of a cell, identify the probable position of the centrosome.</strong></p>\n<p>Looking at the images I find that my limited human vision capacities don't do a good job of this. But we have a training set, since we have images where we know from the labels that the centrosome is visible on the green layer and we have the microtubules on the red layer, and we have further examples where the Golgi apparatus is labelled, that also lies near the centrosome. </p>\n<p>Our biologists tell us that the centrosome is at the “origin” of the microtubules.  It looks like there must be a pattern, but I can’t spot the solution.</p>\n<p>It seems to me that this would be a valuable problem to solve because the cells look otherwise pretty much like unoriented round shapes, but since they only have one centrosome, there is a potential unique orientation.</p>\n<p>( I also asked a question on this under <a href=\"https://www.kaggle.com/lnhtrang/single-cell-patterns\" target=\"_blank\">https://www.kaggle.com/lnhtrang/single-cell-patterns</a>)</p>",
  "messages": [
    {
      "id": 1240207,
      "postDate": "2021-03-16T09:34:40.320Z",
      "content": "<p>Hi All,<br>\nI would like to throw out this little sub-challenge I have been thinking about.</p>\n<p>Much of the HPA competition seems to be tractable to non-machine-learning computer vision techniques, so that some extra effort in hand-coded feature detection could potentially solve, say 80% of the problem with 20% of the CPU/GPU power (see discussion here <a href=\"https://www.kaggle.com/lukemerrick/who-needs-a-gpu-shallow-learning-in-julia)\" target=\"_blank\">https://www.kaggle.com/lukemerrick/who-needs-a-gpu-shallow-learning-in-julia)</a>.</p>\n<p>So what would we do with the other 80% CPU/GPU capacity ?</p>\n<p>How about this one ?</p>\n<p><strong>Given the red microtubules layer of a cell, identify the probable position of the centrosome.</strong></p>\n<p>Looking at the images I find that my limited human vision capacities don't do a good job of this. But we have a training set, since we have images where we know from the labels that the centrosome is visible on the green layer and we have the microtubules on the red layer, and we have further examples where the Golgi apparatus is labelled, that also lies near the centrosome. </p>\n<p>Our biologists tell us that the centrosome is at the “origin” of the microtubules.  It looks like there must be a pattern, but I can’t spot the solution.</p>\n<p>It seems to me that this would be a valuable problem to solve because the cells look otherwise pretty much like unoriented round shapes, but since they only have one centrosome, there is a potential unique orientation.</p>\n<p>( I also asked a question on this under <a href=\"https://www.kaggle.com/lnhtrang/single-cell-patterns\" target=\"_blank\">https://www.kaggle.com/lnhtrang/single-cell-patterns</a>)</p>",
      "rawMarkdown": "Hi All,\nI would like to throw out this little sub-challenge I have been thinking about.\n\nMuch of the HPA competition seems to be tractable to non-machine-learning computer vision techniques, so that some extra effort in hand-coded feature detection could potentially solve, say 80% of the problem with 20% of the CPU/GPU power (see discussion here https://www.kaggle.com/lukemerrick/who-needs-a-gpu-shallow-learning-in-julia).\n\nSo what would we do with the other 80% CPU/GPU capacity ?\n\nHow about this one ?\n\n**Given the red microtubules layer of a cell, identify the probable position of the centrosome.**\n\nLooking at the images I find that my limited human vision capacities don't do a good job of this. But we have a training set, since we have images where we know from the labels that the centrosome is visible on the green layer and we have the microtubules on the red layer, and we have further examples where the Golgi apparatus is labelled, that also lies near the centrosome. \n\nOur biologists tell us that the centrosome is at the “origin” of the microtubules.  It looks like there must be a pattern, but I can’t spot the solution.\n\nIt seems to me that this would be a valuable problem to solve because the cells look otherwise pretty much like unoriented round shapes, but since they only have one centrosome, there is a potential unique orientation.\n\n( I also asked a question on this under https://www.kaggle.com/lnhtrang/single-cell-patterns)\n",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1240207": "Hi All,\nI would like to throw out this little sub-challenge I have been thinking about.\n\nMuch of the HPA competition seems to be tractable to non-machine-learning computer vision techniques, so that some extra effort in hand-coded feature detection could potentially solve, say 80% of the problem with 20% of the CPU/GPU power (see discussion here https://www.kaggle.com/lukemerrick/who-needs-a-gpu-shallow-learning-in-julia).\n\nSo what would we do with the other 80% CPU/GPU capacity ?\n\nHow about this one ?\n\n**Given the red microtubules layer of a cell, identify the probable position of the centrosome.**\n\nLooking at the images I find that my limited human vision capacities don't do a good job of this. But we have a training set, since we have images where we know from the labels that the centrosome is visible on the green layer and we have the microtubules on the red layer, and we have further examples where the Golgi apparatus is labelled, that also lies near the centrosome. \n\nOur biologists tell us that the centrosome is at the “origin” of the microtubules.  It looks like there must be a pattern, but I can’t spot the solution.\n\nIt seems to me that this would be a valuable problem to solve because the cells look otherwise pretty much like unoriented round shapes, but since they only have one centrosome, there is a potential unique orientation.\n\n( I also asked a question on this under https://www.kaggle.com/lnhtrang/single-cell-patterns)\n"
  }
}