{
  "id": 223934,
  "title": "Training image's protein channel suffers from translational artifact",
  "url": "/competitions/hpa-single-cell-image-classification/discussion/223934",
  "author_name": "Milo Toor",
  "post_date": "2021-03-06T07:00:41.416000",
  "votes": 10,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I noticed that at least one image in the training set appears to suffer from some kind of processing artifact, such that its protein channel (i.e. the green image) is visibly distorted relative to the other channels. Image <code>dda15e7a-bbbb-11e8-b2ba-ac1f6b6435d0</code> has the label <code>1|0</code>--\"Nuclear membrane\" and \"Nucleoplasm\", respectively. However, the green channel and the blue channel are clearly misaligned in portions of the image. It would appear that the protein channel is translated relative to the others. See links below for images.</p>\n<p><a href=\"https://ibb.co/zb1xW82\" target=\"_blank\">Blue = nucleoplasm, green = protein</a><br>\n<a href=\"https://ibb.co/ccZLc5H\" target=\"_blank\">Red = microtubules, green = ER, blue = nucleoplasm</a></p>\n<p>As you can see, the non-protein channels are all aligned well to one another and stack on top of each other nicely-- it's just the protein channel that is off. It seems to me that the error is more severe higher in the image than it is in the lower section, but the translation is non-negligible everywhere.</p>\n<p>Because the CellSegmentator is oblivious to the protein channel (at least that's my understanding), the masks are resilient to the translation artifact. However it's not clear to me what the proper way of labeling the masks would be, especially in the image's upper section when the cells' masks and their associated protein fluorescence are completely disjoint.</p>\n<p>My inclination is to just remove this image from my training set, though I'm curious to know if this is in fact an imaging artifact and not some weird cellular phenomenon. In any case I thought it would be a good idea to bring this to folks' attention, thanks for hearing me out!</p>",
  "messages": [
    {
      "id": 1228173,
      "postDate": "2021-03-06T07:00:41.417Z",
      "content": "<p>I noticed that at least one image in the training set appears to suffer from some kind of processing artifact, such that its protein channel (i.e. the green image) is visibly distorted relative to the other channels. Image <code>dda15e7a-bbbb-11e8-b2ba-ac1f6b6435d0</code> has the label <code>1|0</code>--\"Nuclear membrane\" and \"Nucleoplasm\", respectively. However, the green channel and the blue channel are clearly misaligned in portions of the image. It would appear that the protein channel is translated relative to the others. See links below for images.</p>\n<p><a href=\"https://ibb.co/zb1xW82\" target=\"_blank\">Blue = nucleoplasm, green = protein</a><br>\n<a href=\"https://ibb.co/ccZLc5H\" target=\"_blank\">Red = microtubules, green = ER, blue = nucleoplasm</a></p>\n<p>As you can see, the non-protein channels are all aligned well to one another and stack on top of each other nicely-- it's just the protein channel that is off. It seems to me that the error is more severe higher in the image than it is in the lower section, but the translation is non-negligible everywhere.</p>\n<p>Because the CellSegmentator is oblivious to the protein channel (at least that's my understanding), the masks are resilient to the translation artifact. However it's not clear to me what the proper way of labeling the masks would be, especially in the image's upper section when the cells' masks and their associated protein fluorescence are completely disjoint.</p>\n<p>My inclination is to just remove this image from my training set, though I'm curious to know if this is in fact an imaging artifact and not some weird cellular phenomenon. In any case I thought it would be a good idea to bring this to folks' attention, thanks for hearing me out!</p>",
      "rawMarkdown": "I noticed that at least one image in the training set appears to suffer from some kind of processing artifact, such that its protein channel (i.e. the green image) is visibly distorted relative to the other channels. Image `dda15e7a-bbbb-11e8-b2ba-ac1f6b6435d0` has the label `1|0`--\"Nuclear membrane\" and \"Nucleoplasm\", respectively. However, the green channel and the blue channel are clearly misaligned in portions of the image. It would appear that the protein channel is translated relative to the others. See links below for images.\n\n[Blue = nucleoplasm, green = protein](https://ibb.co/zb1xW82)\n[Red = microtubules, green = ER, blue = nucleoplasm](https://ibb.co/ccZLc5H)\n\nAs you can see, the non-protein channels are all aligned well to one another and stack on top of each other nicely-- it's just the protein channel that is off. It seems to me that the error is more severe higher in the image than it is in the lower section, but the translation is non-negligible everywhere.\n\nBecause the CellSegmentator is oblivious to the protein channel (at least that's my understanding), the masks are resilient to the translation artifact. However it's not clear to me what the proper way of labeling the masks would be, especially in the image's upper section when the cells' masks and their associated protein fluorescence are completely disjoint.\n\nMy inclination is to just remove this image from my training set, though I'm curious to know if this is in fact an imaging artifact and not some weird cellular phenomenon. In any case I thought it would be a good idea to bring this to folks' attention, thanks for hearing me out!",
      "votes": 9
    },
    {
      "id": 1228689,
      "postDate": "2021-03-06T16:43:13.727Z",
      "content": "<p>Great find!</p>\n<p>I know they encountered similar issues to this in the previous competition as well!</p>",
      "rawMarkdown": "Great find!\n\nI know they encountered similar issues to this in the previous competition as well!",
      "votes": 3
    }
  ],
  "comments": [
    {
      "id": 1228689,
      "author_name": "Darien Schettler",
      "author_url": "",
      "post_date": "2021-03-06T16:43:13.727000",
      "content": "<p>Great find!</p>\n<p>I know they encountered similar issues to this in the previous competition as well!</p>",
      "votes": 3,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1228173": "I noticed that at least one image in the training set appears to suffer from some kind of processing artifact, such that its protein channel (i.e. the green image) is visibly distorted relative to the other channels. Image `dda15e7a-bbbb-11e8-b2ba-ac1f6b6435d0` has the label `1|0`--\"Nuclear membrane\" and \"Nucleoplasm\", respectively. However, the green channel and the blue channel are clearly misaligned in portions of the image. It would appear that the protein channel is translated relative to the others. See links below for images.\n\n[Blue = nucleoplasm, green = protein](https://ibb.co/zb1xW82)\n[Red = microtubules, green = ER, blue = nucleoplasm](https://ibb.co/ccZLc5H)\n\nAs you can see, the non-protein channels are all aligned well to one another and stack on top of each other nicely-- it's just the protein channel that is off. It seems to me that the error is more severe higher in the image than it is in the lower section, but the translation is non-negligible everywhere.\n\nBecause the CellSegmentator is oblivious to the protein channel (at least that's my understanding), the masks are resilient to the translation artifact. However it's not clear to me what the proper way of labeling the masks would be, especially in the image's upper section when the cells' masks and their associated protein fluorescence are completely disjoint.\n\nMy inclination is to just remove this image from my training set, though I'm curious to know if this is in fact an imaging artifact and not some weird cellular phenomenon. In any case I thought it would be a good idea to bring this to folks' attention, thanks for hearing me out!",
    "1228689": "Great find!\n\nI know they encountered similar issues to this in the previous competition as well!"
  }
}