![\"enter](\"https://i.imgur.com/jWkhRw3.jpg\")
I've been debating internally for few days whether to write this or not, and this post finally convinced me to do it. In that post @Heng CherKeng correctly identifies that there are several cells not included in the mask, but that's not the only problem with it. A separate problem in these images is that there lots of signals identified as nuclei that are most definitely not cells. Here are the same image and its mask side by side, courtesy of @Sashikanth Dareddy and this post:
\n\n
You may need to open these images in a separate tab to see them in original size.
\n\nAlthough there are some exceptions, it is a rule that signals that do not look like cells in terms of shape, and especially in terms of intensity, are not cells. I have marked three such signals on the left side and there is at least one more that I didn't mark. All of them are present in the mask. As I've described separately, sometimes there is literally dust or some other buffer impurity that will create these artifacts. Now, one could reasonably argue that mis-identifying 3-4 cells in a field of 100 is not a big deal, but this happens even in images that have fewer cells. Besides, it is a type of false signal that is very easy to spot and eliminate because of its different shape and intensity.
\n\nAnother example where things are identified poorly even at high magnification:
\n\n![\"enter](\"https://i.imgur.com/ye316tI.png\")
The part in the middle is most definitely not a nucleus, even though the signal density seems about right. First, nuclei are round and the signal is continuously distributed in them as you can see from other cells. At best, this is a nucleus that has exploded and its contents are now diffusing.
\n\nLastly, this is one of several images where the nucleus is not stained. I don't know exactly what the apical structure with strong signal is, but an educated guess is Golgi apparatus. Nuclei are clearly seen as round structures next to that signal, yet they are not what is identified in masks.
\n\n![\"enter](\"https://i.imgur.com/IoTVMtN.png\")
This is also problematic because in several instances there are two or even multiple cells that have these signals overlapping, so one would severely underestimate the cell number in this image.
\n\nThis brings me to the question posed in title: Are we identifying cells or nuclei? Competition description says nuclei, but image masks say otherwise. To me, this seems like a cell-counting challenge rather than nuclei identification. If that's not the real intent, I think that quite a few image masks need to be fixed.
", "messages": [ { "id": "271166", "postDate": "01/19/2018 20:22:43", "content": "I've been debating internally for few days whether to write this or not, and this post finally convinced me to do it. In that post @Heng CherKeng correctly identifies that there are several cells not included in the mask, but that's not the only problem with it. A separate problem in these images is that there lots of signals identified as nuclei that are most definitely not cells. Here are the same image and its mask side by side, courtesy of @Sashikanth Dareddy and this post:
\n\n
You may need to open these images in a separate tab to see them in original size.
\n\nAlthough there are some exceptions, it is a rule that signals that do not look like cells in terms of shape, and especially in terms of intensity, are not cells. I have marked three such signals on the left side and there is at least one more that I didn't mark. All of them are present in the mask. As I've described separately, sometimes there is literally dust or some other buffer impurity that will create these artifacts. Now, one could reasonably argue that mis-identifying 3-4 cells in a field of 100 is not a big deal, but this happens even in images that have fewer cells. Besides, it is a type of false signal that is very easy to spot and eliminate because of its different shape and intensity.
\n\nAnother example where things are identified poorly even at high magnification:
\n\n
The part in the middle is most definitely not a nucleus, even though the signal density seems about right. First, nuclei are round and the signal is continuously distributed in them as you can see from other cells. At best, this is a nucleus that has exploded and its contents are now diffusing.
\n\nLastly, this is one of several images where the nucleus is not stained. I don't know exactly what the apical structure with strong signal is, but an educated guess is Golgi apparatus. Nuclei are clearly seen as round structures next to that signal, yet they are not what is identified in masks.
\n\n
This is also problematic because in several instances there are two or even multiple cells that have these signals overlapping, so one would severely underestimate the cell number in this image.
\n\nThis brings me to the question posed in title: Are we identifying cells or nuclei? Competition description says nuclei, but image masks say otherwise. To me, this seems like a cell-counting challenge rather than nuclei identification. If that's not the real intent, I think that quite a few image masks need to be fixed.
", "rawMarkdown": "I've been debating internally for few days whether to write this or not, and this [__post__](https://www.kaggle.com/c/data-science-bowl-2018/discussion/47770#271020) finally convinced me to do it. In that post @Heng CherKeng correctly identifies that there are several cells not included in the mask, but that's not the only problem with it. A separate problem in these images is that there lots of signals identified as nuclei that are most definitely not cells. Here are the same image and its mask side by side, courtesy of @Sashikanth Dareddy and [__this post__](https://www.kaggle.com/c/data-science-bowl-2018/discussion/47672):\n\n![enter image description here][1]\n\n**You may need to open these images in a separate tab to see them in original size.**\n\nAlthough there are some exceptions, it is a rule that signals that do not look like cells in terms of shape, and especially in terms of intensity, are not cells. I have marked three such signals on the left side and there is at least one more that I didn't mark. All of them are present in the mask. As I've described [__separately__](https://www.kaggle.com/c/data-science-bowl-2018/discussion/47681), sometimes there is literally dust or some other buffer impurity that will create these artifacts. Now, one could reasonably argue that mis-identifying 3-4 cells in a field of 100 is not a big deal, but this happens even in images that have fewer cells. Besides, it is a type of false signal that is very easy to spot and eliminate because of its different shape and intensity.\n\nAnother example where things are identified poorly even at high magnification:\n\n![enter image description here][2]\n\nThe part in the middle is most definitely not a nucleus, even though the signal density seems about right. First, nuclei are round and the signal is continuously distributed in them as you can see from other cells. At best, this is a nucleus that has exploded and its contents are now diffusing.\n\nLastly, this is one of several images where the nucleus is not stained. I don't know exactly what the apical structure with strong signal is, but an educated guess is [__Golgi apparatus__](https://en.wikipedia.org/wiki/Golgi_apparatus). Nuclei are clearly seen as round structures next to that signal, yet they are not what is identified in masks.\n\n![enter image description here][3]\n\nThis is also problematic because in several instances there are two or even multiple cells that have these signals overlapping, so one would severely underestimate the cell number in this image.\n\nThis brings me to the question posed in title: Are we identifying cells or nuclei? Competition description says nuclei, but image masks say otherwise. To me, this seems like a cell-counting challenge rather than nuclei identification. If that's not the real intent, I think that quite a few image masks need to be fixed.\n\n [1]: https://i.imgur.com/jWkhRw3.jpg \"Non-cellular specks\"\n [2]: https://i.imgur.com/ye316tI.png \"Exploding cell\"\n [3]: https://i.imgur.com/IoTVMtN.png \"Most likely Golgi\"", "votes": null }, { "id": "271188", "postDate": "01/19/2018 21:18:46", "content": "The definition of nuclei in this competition can be thought of generally as any independent body of DNA. Frequently, experimental conditions will make cells sick enough that their nuclei become deformed or fragmented, and nuclei no longer resemble the round, uniform shapes of healthy cells. Therefore, unusual shapes of nuclei are to be expected. Furthermore, when looking at a cluster of stressed cells, it can be difficult to discern the boundaries between nuclei. In fact, there may even be disagreement between biologists when deciding the boundaries. Multiple biologists were responsible for the annotation of the training data, so, while acknowledging that some annotations will be seen differently by individuals, the variety of opinions between biologists will be represented when considering the data set as a whole.
\n\nIn addition, the fragments of a deformed nuclei, the region that Tilli refers to as possibly the golgi, are known as micronuclei. The presence of micronuclei can be informative to the overall health of the cell and therefore it is desirable to segment them. However, the small size of the micronuclei posed a challenge to their annotation, and not every micronuclei in the image set will be annotated.
", "rawMarkdown": "The definition of nuclei in this competition can be thought of generally as any independent body of DNA. Frequently, experimental conditions will make cells sick enough that their nuclei become deformed or fragmented, and nuclei no longer resemble the round, uniform shapes of healthy cells. Therefore, unusual shapes of nuclei are to be expected. Furthermore, when looking at a cluster of stressed cells, it can be difficult to discern the boundaries between nuclei. In fact, there may even be disagreement between biologists when deciding the boundaries. Multiple biologists were responsible for the annotation of the training data, so, while acknowledging that some annotations will be seen differently by individuals, the variety of opinions between biologists will be represented when considering the data set as a whole.\n\nIn addition, the fragments of a deformed nuclei, the region that Tilli refers to as possibly the golgi, are known as micronuclei. The presence of micronuclei can be informative to the overall health of the cell and therefore it is desirable to segment them. However, the small size of the micronuclei posed a challenge to their annotation, and not every micronuclei in the image set will be annotated.", "votes": null }, { "id": "271190", "postDate": "01/19/2018 21:27:13", "content": "Kyle is on the Broad Institute team that annotated these so he is a reliable source; as another member of that team, I would like to add to what he has said:
\n\nThe first image has major problems, nuclei are missing. That is bad.
\n\nThe second image has micronuclei or other DNA regions and we made the conscious decision to INCLUDE them; a model ought to identify them and if biologists want to exclude them for their particular analysis it would be easy to do so by simple filtering based on morphological properties.
\n\nThe third image IS absolutely nuclei, with a peculiar appearance because they are adjacent to vacuoles. This is exactly the kind of variation we expect a model to be able to handle.
", "rawMarkdown": "Kyle is on the Broad Institute team that annotated these so he is a reliable source; as another member of that team, I would like to add to what he has said: \n\nThe first image has major problems, nuclei are missing. That is bad.\n\nThe second image has micronuclei or other DNA regions and we made the conscious decision to INCLUDE them; a model ought to identify them and if biologists want to exclude them for their particular analysis it would be easy to do so by simple filtering based on morphological properties.\n\nThe third image IS absolutely nuclei, with a peculiar appearance because they are adjacent to vacuoles. This is exactly the kind of variation we expect a model to be able to handle.", "votes": null }, { "id": "271221", "postDate": "01/19/2018 22:21:54", "content": "The notion that all the signals in images could be cells is OK. The idea that all signals in images must be cells is simply not true. Back to image #1, I think there are several bright signals in it that I consider not to be cells, no matter how stressed.
\n\nI still think #2 is at best an exploded nucleus.
\n\nI'll buy the vacuole explanation, as I already though that could be it.
\n\nSo I can learn something, here is yet another image mask I consider problematic:
\n\n![\"enter](\"https://i.imgur.com/d0Xl41w.png\")
Just by looking at it, I'd say that this cell has specular cytoplasmic staining and diffuse nuclear stain. But let's say for the argument sake that the whole oval thing is weirdly stained nucleus. What exactly would be the hollow part inside of nucleus? Because the image mask is telling us that there is something in the middle of nucleus that actually isn't nucleus.
", "rawMarkdown": "The notion that all the signals in images could be cells is OK. The idea that all signals in images must be cells is simply not true. Back to image #1, I think there are several bright signals in it that I consider not to be cells, no matter how stressed.\n\nI still think #2 is at best an exploded nucleus.\n\nI'll buy the vacuole explanation, as I already though that could be it.\n\nSo I can learn something, here is yet another image mask I consider problematic:\n\n![enter image description here][1]\n\nJust by looking at it, I'd say that this cell has specular cytoplasmic staining and diffuse nuclear stain. But let's say for the argument sake that the whole oval thing is weirdly stained nucleus. What exactly would be the hollow part inside of nucleus? Because the image mask is telling us that there is something in the middle of nucleus that actually isn't nucleus.\n\n [1]: https://i.imgur.com/d0Xl41w.png", "votes": null }, { "id": "271563", "postDate": "01/20/2018 19:40:49", "content": "Hi there, organizer here. Glad you are open to learning. That image captured an embryo with multiple nuclei. Beautiful, isn't it?
", "rawMarkdown": "Hi there, organizer here. Glad you are open to learning. That image captured an embryo with multiple nuclei. Beautiful, isn't it?", "votes": null }, { "id": "271868", "postDate": "01/21/2018 18:16:45", "content": "@Anne Carpenter Thank you for explaining the embryo image.
\n\nI found a movie of an exploding nucleus (smack in the middle):
\n\n![\"enter](\"https://i.imgur.com/KzHDWDb.gif\")
I know it isn't exactly apples to apples, but I think that the final result is fairly similar to the micronuclei image above.
", "rawMarkdown": "Anne Carpenter Thank you for explaining the embryo image.\n\nI found a movie of an exploding nucleus (smack in the middle):\n\n![enter image description here][1]\n\nI know it isn't exactly apples to apples, but I think that the final result is fairly similar to the micronuclei image above.\n\n [1]: https://i.imgur.com/KzHDWDb.gif", "votes": null }, { "id": "271905", "postDate": "01/21/2018 20:56:42", "content": "Bummer for that cell, but fun to watch :D
", "rawMarkdown": "Bummer for that cell, but fun to watch :D", "votes": null } ], "comments": [ { "id": 271188, "author_name": "bitfrosting", "author_url": "", "post_date": "01/19/2018 21:18:46", "content": "The definition of nuclei in this competition can be thought of generally as any independent body of DNA. Frequently, experimental conditions will make cells sick enough that their nuclei become deformed or fragmented, and nuclei no longer resemble the round, uniform shapes of healthy cells. Therefore, unusual shapes of nuclei are to be expected. Furthermore, when looking at a cluster of stressed cells, it can be difficult to discern the boundaries between nuclei. In fact, there may even be disagreement between biologists when deciding the boundaries. Multiple biologists were responsible for the annotation of the training data, so, while acknowledging that some annotations will be seen differently by individuals, the variety of opinions between biologists will be represented when considering the data set as a whole.
\n\nIn addition, the fragments of a deformed nuclei, the region that Tilli refers to as possibly the golgi, are known as micronuclei. The presence of micronuclei can be informative to the overall health of the cell and therefore it is desirable to segment them. However, the small size of the micronuclei posed a challenge to their annotation, and not every micronuclei in the image set will be annotated.
", "votes": null, "replies": [ { "id": 271190, "author_name": "drannecarpenter", "author_url": "", "post_date": "01/19/2018 21:27:13", "content": "Kyle is on the Broad Institute team that annotated these so he is a reliable source; as another member of that team, I would like to add to what he has said:
\n\nThe first image has major problems, nuclei are missing. That is bad.
\n\nThe second image has micronuclei or other DNA regions and we made the conscious decision to INCLUDE them; a model ought to identify them and if biologists want to exclude them for their particular analysis it would be easy to do so by simple filtering based on morphological properties.
\n\nThe third image IS absolutely nuclei, with a peculiar appearance because they are adjacent to vacuoles. This is exactly the kind of variation we expect a model to be able to handle.
", "votes": null, "replies": [] }, { "id": 271221, "author_name": "tilii7", "author_url": "", "post_date": "01/19/2018 22:21:54", "content": "The notion that all the signals in images could be cells is OK. The idea that all signals in images must be cells is simply not true. Back to image #1, I think there are several bright signals in it that I consider not to be cells, no matter how stressed.
\n\nI still think #2 is at best an exploded nucleus.
\n\nI'll buy the vacuole explanation, as I already though that could be it.
\n\nSo I can learn something, here is yet another image mask I consider problematic:
\n\n
Just by looking at it, I'd say that this cell has specular cytoplasmic staining and diffuse nuclear stain. But let's say for the argument sake that the whole oval thing is weirdly stained nucleus. What exactly would be the hollow part inside of nucleus? Because the image mask is telling us that there is something in the middle of nucleus that actually isn't nucleus.
", "votes": null, "replies": [] }, { "id": 271563, "author_name": "drannecarpenter", "author_url": "", "post_date": "01/20/2018 19:40:49", "content": "Hi there, organizer here. Glad you are open to learning. That image captured an embryo with multiple nuclei. Beautiful, isn't it?
", "votes": null, "replies": [] }, { "id": 271868, "author_name": "tilii7", "author_url": "", "post_date": "01/21/2018 18:16:45", "content": "@Anne Carpenter Thank you for explaining the embryo image.
\n\nI found a movie of an exploding nucleus (smack in the middle):
\n\n
I know it isn't exactly apples to apples, but I think that the final result is fairly similar to the micronuclei image above.
", "votes": null, "replies": [] }, { "id": 271905, "author_name": "drannecarpenter", "author_url": "", "post_date": "01/21/2018 20:56:42", "content": "Bummer for that cell, but fun to watch :D
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "271166": "I've been debating internally for few days whether to write this or not, and this [__post__](https://www.kaggle.com/c/data-science-bowl-2018/discussion/47770#271020) finally convinced me to do it. In that post @Heng CherKeng correctly identifies that there are several cells not included in the mask, but that's not the only problem with it. A separate problem in these images is that there lots of signals identified as nuclei that are most definitely not cells. Here are the same image and its mask side by side, courtesy of @Sashikanth Dareddy and [__this post__](https://www.kaggle.com/c/data-science-bowl-2018/discussion/47672):\n\n![enter image description here][1]\n\n**You may need to open these images in a separate tab to see them in original size.**\n\nAlthough there are some exceptions, it is a rule that signals that do not look like cells in terms of shape, and especially in terms of intensity, are not cells. I have marked three such signals on the left side and there is at least one more that I didn't mark. All of them are present in the mask. As I've described [__separately__](https://www.kaggle.com/c/data-science-bowl-2018/discussion/47681), sometimes there is literally dust or some other buffer impurity that will create these artifacts. Now, one could reasonably argue that mis-identifying 3-4 cells in a field of 100 is not a big deal, but this happens even in images that have fewer cells. Besides, it is a type of false signal that is very easy to spot and eliminate because of its different shape and intensity.\n\nAnother example where things are identified poorly even at high magnification:\n\n![enter image description here][2]\n\nThe part in the middle is most definitely not a nucleus, even though the signal density seems about right. First, nuclei are round and the signal is continuously distributed in them as you can see from other cells. At best, this is a nucleus that has exploded and its contents are now diffusing.\n\nLastly, this is one of several images where the nucleus is not stained. I don't know exactly what the apical structure with strong signal is, but an educated guess is [__Golgi apparatus__](https://en.wikipedia.org/wiki/Golgi_apparatus). Nuclei are clearly seen as round structures next to that signal, yet they are not what is identified in masks.\n\n![enter image description here][3]\n\nThis is also problematic because in several instances there are two or even multiple cells that have these signals overlapping, so one would severely underestimate the cell number in this image.\n\nThis brings me to the question posed in title: Are we identifying cells or nuclei? Competition description says nuclei, but image masks say otherwise. To me, this seems like a cell-counting challenge rather than nuclei identification. If that's not the real intent, I think that quite a few image masks need to be fixed.\n\n [1]: https://i.imgur.com/jWkhRw3.jpg \"Non-cellular specks\"\n [2]: https://i.imgur.com/ye316tI.png \"Exploding cell\"\n [3]: https://i.imgur.com/IoTVMtN.png \"Most likely Golgi\"", "271188": "The definition of nuclei in this competition can be thought of generally as any independent body of DNA. Frequently, experimental conditions will make cells sick enough that their nuclei become deformed or fragmented, and nuclei no longer resemble the round, uniform shapes of healthy cells. Therefore, unusual shapes of nuclei are to be expected. Furthermore, when looking at a cluster of stressed cells, it can be difficult to discern the boundaries between nuclei. In fact, there may even be disagreement between biologists when deciding the boundaries. Multiple biologists were responsible for the annotation of the training data, so, while acknowledging that some annotations will be seen differently by individuals, the variety of opinions between biologists will be represented when considering the data set as a whole.\n\nIn addition, the fragments of a deformed nuclei, the region that Tilli refers to as possibly the golgi, are known as micronuclei. The presence of micronuclei can be informative to the overall health of the cell and therefore it is desirable to segment them. However, the small size of the micronuclei posed a challenge to their annotation, and not every micronuclei in the image set will be annotated.", "271190": "Kyle is on the Broad Institute team that annotated these so he is a reliable source; as another member of that team, I would like to add to what he has said: \n\nThe first image has major problems, nuclei are missing. That is bad.\n\nThe second image has micronuclei or other DNA regions and we made the conscious decision to INCLUDE them; a model ought to identify them and if biologists want to exclude them for their particular analysis it would be easy to do so by simple filtering based on morphological properties.\n\nThe third image IS absolutely nuclei, with a peculiar appearance because they are adjacent to vacuoles. This is exactly the kind of variation we expect a model to be able to handle.", "271221": "The notion that all the signals in images could be cells is OK. The idea that all signals in images must be cells is simply not true. Back to image #1, I think there are several bright signals in it that I consider not to be cells, no matter how stressed.\n\nI still think #2 is at best an exploded nucleus.\n\nI'll buy the vacuole explanation, as I already though that could be it.\n\nSo I can learn something, here is yet another image mask I consider problematic:\n\n![enter image description here][1]\n\nJust by looking at it, I'd say that this cell has specular cytoplasmic staining and diffuse nuclear stain. But let's say for the argument sake that the whole oval thing is weirdly stained nucleus. What exactly would be the hollow part inside of nucleus? Because the image mask is telling us that there is something in the middle of nucleus that actually isn't nucleus.\n\n [1]: https://i.imgur.com/d0Xl41w.png", "271563": "Hi there, organizer here. Glad you are open to learning. That image captured an embryo with multiple nuclei. Beautiful, isn't it?", "271868": "Anne Carpenter Thank you for explaining the embryo image.\n\nI found a movie of an exploding nucleus (smack in the middle):\n\n![enter image description here][1]\n\nI know it isn't exactly apples to apples, but I think that the final result is fairly similar to the micronuclei image above.\n\n [1]: https://i.imgur.com/KzHDWDb.gif", "271905": "Bummer for that cell, but fun to watch :D" }, "source": "meta" }