{ "id": 220275, "title": "Cell-level labels", "url": "/competitions/hpa-single-cell-image-classification/discussion/220275", "author_name": "Zepyhr99", "post_date": "2021-02-17T20:49:36.331000", "votes": 3, "comment_count": 13, "views": 0, "content": "

I just want to be clarified about what am I predicting. Do each cell has only single class included in the image-level labels or they can have multiple classes as the subset of image-level labels?

", "messages": [ { "id": 1207564, "postDate": "2021-02-17T23:21:13.933Z", "content": "

The single cell labels can also contain labels which are not included in the image-level labels, so it's not strictly subset of image-level labels.

", "rawMarkdown": "The single cell labels can also contain labels which are not included in the image-level labels, so it's not strictly subset of image-level labels.", "votes": 3, "replies": [ { "id": 1207739, "postDate": "2021-02-18T02:28:06.193Z", "content": "

Thanks for the clarification.
\nI thought, every cell has 0, 1 or 0&1 in the image of label 0|1, but this is incorrect according to your explanation…

", "rawMarkdown": "Thanks for the clarification.\nI thought, every cell has 0, 1 or 0&1 in the image of label 0|1, but this is incorrect according to your explanation..." }, { "id": 1207764, "postDate": "2021-02-18T03:01:14.780Z", "content": "

My understanding from this notebook is that the only class that can be labeled as such, which is not present in the image-level label but is present in the cell-level label, is Negative, but are you saying that classes other than Negative can also be labeled as such?

", "rawMarkdown": "My understanding from [this notebook](https://www.kaggle.com/lnhtrang/single-cell-patterns#SCV-and-multi-localization-examples) is that the only class that can be labeled as such, which is not present in the image-level label but is present in the cell-level label, is Negative, but are you saying that classes other than Negative can also be labeled as such?", "votes": 2 }, { "id": 1207765, "postDate": "2021-02-18T03:01:27.240Z", "content": "

Do you mean that every cell may has 3 or 4 in the image of label 0|1😨

", "rawMarkdown": "Do you mean that every cell may has 3 or 4 in the image of label 0|1😨" }, { "id": 1208836, "postDate": "2021-02-18T14:04:19.280Z", "content": "

Shouldn't be, else how are we supposed to train our model?

", "rawMarkdown": "Shouldn't be, else how are we supposed to train our model?" }, { "id": 1209739, "postDate": "2021-02-19T02:57:45.040Z", "content": "

\"However, for other images (especially in the test set), each cell will have only a part of the image-level label, or completely new label (18 Negative).\" quoting from https://www.kaggle.com/lnhtrang/single-cell-patterns#SCV-and-multi-localization-examples

\n

But the quote contradicts what you state here…

", "rawMarkdown": "\"However, for other images (especially in the test set), each cell will have only a part of the image-level label, or completely new label (18 Negative).\" quoting from [https://www.kaggle.com/lnhtrang/single-cell-patterns#SCV-and-multi-localization-examples](url)\n\nBut the quote contradicts what you state here...\n", "votes": 1 }, { "id": 1210863, "postDate": "2021-02-19T19:00:44.823Z", "content": "

Hi,

\n

Please let me clarify this. The statement in the notebook \"However, for other images (especially in the test set), each cell will have only a part of the image-level label, or completely new label (18 Negative).\" is mainly true. However, these images are manually labeled and there may be errors in terms of wrong labels and missing labels, but only to a minor extent. You are supposed to train your models using the image level labels.

\n

A difference that is important to take into consideration is that there is more variability between cell labels in the test set than in the training set (i.e. in the training set there is a much higher fraction of images where all cells have identical labels). \"This competition is a weakly-labeled challenge: based on the image level label, you build models to predict labels for each individual cell. For this competition, we specifically annotated every single cell in a subset of images with higher SCV compared to public HPA images, to act as test set. In other words, the test set will have higher SCV than the training set. (SCV = Single Cell Variability)\" as quoted from https://www.kaggle.com/lnhtrang/single-cell-patterns/notebook#Protein-localization-patterns.

\n

I hope this helps to clear things up.

", "rawMarkdown": "Hi,\n\nPlease let me clarify this. The statement in the notebook \"However, for other images (especially in the test set), each cell will have only a part of the image-level label, or completely new label (18 Negative).\" is mainly true. However, these images are manually labeled and there may be errors in terms of wrong labels and missing labels, but only to a minor extent. You are supposed to train your models using the image level labels.\n\nA difference that is important to take into consideration is that there is more variability between cell labels in the test set than in the training set (i.e. in the training set there is a much higher fraction of images where all cells have identical labels). \"This competition is a weakly-labeled challenge: based on the image level label, you build models to predict labels for each individual cell. For this competition, we specifically annotated every single cell in a subset of images with higher SCV compared to public HPA images, to act as test set. In other words, the test set will have higher SCV than the training set. (SCV = Single Cell Variability)\" as quoted from https://www.kaggle.com/lnhtrang/single-cell-patterns/notebook#Protein-localization-patterns.\n\nI hope this helps to clear things up.", "votes": 5 }, { "id": 1211154, "postDate": "2021-02-20T02:14:30.830Z", "content": "

Hi @emmalumpan, thanks for clearing it up.
\nJust to make sure I understand the problem correctly, the answer following question by @zstusnoopy :

\n
\n

Do you mean that every cell may has 3 or 4 in the image of label 0|1😨

\n
\n

is that it might be the case, but it's unintentional and mostly due to human error of labeling?

", "rawMarkdown": "Hi @emmalumpan, thanks for clearing it up.\nJust to make sure I understand the problem correctly, the answer following question by @zstusnoopy :\n\n> Do you mean that every cell may has 3 or 4 in the image of label 0|1😨\n\nis that it might be the case, but it's unintentional and mostly due to human error of labeling?\n\n" }, { "id": 1211173, "postDate": "2021-02-20T03:07:17.360Z", "content": "

Thanks for that clarification @emmalumpan.

", "rawMarkdown": "Thanks for that clarification @emmalumpan." }, { "id": 1211239, "postDate": "2021-02-20T04:56:50.237Z", "content": "
\n
\n

Do you mean that every cell may has 3 or 4 in the image of label 0|1😨

\n
\n

is that it might be the case, but it's unintentional and mostly due to human error of labelling?

\n
\n

@arka47 Yes, it can be the case but not common or rare for the training dataset. Basically, the single cells can have inconsistent or added labels compared to the image-level labels, due to unintentional labelling error, or sometimes SVC (Single Cell Variability). (Although, I wouldn't say every cell in the image but perhaps a few cells).

\n

However, as Emma pointed out, the test set has higher SVC than the training set, as a result, you should expect higher inconsistency between the cell-level and image-level labels in the test set, including the case where new labels are added to the cell labels (e.g. your predicted image-level labels may have 0 and 1, but the cell may actually have 0 and 4).

\n

This means, during inference, you should NOT completely exclude labels that are outside the range of your image-level predictions.

", "rawMarkdown": ">> Do you mean that every cell may has 3 or 4 in the image of label 0|1😨\n\n> is that it might be the case, but it's unintentional and mostly due to human error of labelling?\n\n@arka47 Yes, it can be the case but not common or rare for the training dataset. Basically, the single cells can have inconsistent or added labels compared to the image-level labels, due to unintentional labelling error, or sometimes SVC (Single Cell Variability). (Although, I wouldn't say `every cell` in the image but perhaps a few cells). \n\nHowever, as Emma pointed out, the test set has higher SVC than the training set, as a result, you should expect higher inconsistency between the cell-level and image-level labels in the test set, including the case where new labels are added to the cell labels (e.g. your predicted image-level labels may have 0 and 1, but the cell may actually have 0 and 4). \n\nThis means, during inference, you should NOT completely exclude labels that are outside the range of your image-level predictions.", "votes": 1 } ] }, { "id": 1207423, "postDate": "2021-02-17T20:49:36.333Z", "content": "

I just want to be clarified about what am I predicting. Do each cell has only single class included in the image-level labels or they can have multiple classes as the subset of image-level labels?

", "rawMarkdown": "I just want to be clarified about what am I predicting. Do each cell has only single class included in the image-level labels or they can have multiple classes as the subset of image-level labels?", "votes": 3 }, { "id": 1207506, "postDate": "2021-02-17T21:59:57.927Z", "content": "

Each cell can have multiple labels.

", "rawMarkdown": "Each cell can have multiple labels.", "replies": [ { "id": 1209843, "postDate": "2021-02-19T04:22:07.667Z", "content": "

So, according to my understanding, the model should be able to --
\n1) segment the cells from the image
\n2) Segment the organelles inside the cell
\n Please correct me if I am wrong

", "rawMarkdown": "So, according to my understanding, the model should be able to --\n1) segment the cells from the image \n2) Segment the organelles **inside** the cell \n Please correct me if I am wrong" }, { "id": 1210065, "postDate": "2021-02-19T07:23:49.597Z", "content": "

Hi,

\n

Yes you are nearly correct. The model should be able to
\n1) Segment the cells from the image
\n2) Classify the protein localization pattern(s) for each cell.

\n

Regarding #2 you could segment the different organelles as part of your approach, but we only require labels per cell for submission.

", "rawMarkdown": "Hi,\n\nYes you are nearly correct. The model should be able to\n1) Segment the cells from the image\n2) **Classify** the protein localization pattern(s) for each cell. \n\nRegarding #2 you could segment the different organelles as part of your approach, but we only require labels per cell for submission.", "votes": 1 } ] } ], "comments": [ { "id": 1207564, "author_name": "Wei Ouyang", "author_url": "", "post_date": "2021-02-17T23:21:13.933000", "content": "

The single cell labels can also contain labels which are not included in the image-level labels, so it's not strictly subset of image-level labels.

", "votes": 3, "replies": [ { "id": 1207739, "author_name": "cool_rabbit", "author_url": "", "post_date": "2021-02-18T02:28:06.193000", "content": "

Thanks for the clarification.
\nI thought, every cell has 0, 1 or 0&1 in the image of label 0|1, but this is incorrect according to your explanation…

", "votes": 0, "replies": [] }, { "id": 1207764, "author_name": "KaizaburoChubachi", "author_url": "", "post_date": "2021-02-18T03:01:14.780000", "content": "

My understanding from this notebook is that the only class that can be labeled as such, which is not present in the image-level label but is present in the cell-level label, is Negative, but are you saying that classes other than Negative can also be labeled as such?

", "votes": 2, "replies": [] }, { "id": 1207765, "author_name": "Linye Li", "author_url": "", "post_date": "2021-02-18T03:01:27.240000", "content": "

Do you mean that every cell may has 3 or 4 in the image of label 0|1😨

", "votes": 0, "replies": [] }, { "id": 1208836, "author_name": "Arka Saha", "author_url": "", "post_date": "2021-02-18T14:04:19.280000", "content": "

Shouldn't be, else how are we supposed to train our model?

", "votes": 0, "replies": [] }, { "id": 1209739, "author_name": "Darkate", "author_url": "", "post_date": "2021-02-19T02:57:45.040000", "content": "

\"However, for other images (especially in the test set), each cell will have only a part of the image-level label, or completely new label (18 Negative).\" quoting from https://www.kaggle.com/lnhtrang/single-cell-patterns#SCV-and-multi-localization-examples

\n

But the quote contradicts what you state here…

", "votes": 1, "replies": [] }, { "id": 1210863, "author_name": "Emma Lundberg", "author_url": "", "post_date": "2021-02-19T19:00:44.823000", "content": "

Hi,

\n

Please let me clarify this. The statement in the notebook \"However, for other images (especially in the test set), each cell will have only a part of the image-level label, or completely new label (18 Negative).\" is mainly true. However, these images are manually labeled and there may be errors in terms of wrong labels and missing labels, but only to a minor extent. You are supposed to train your models using the image level labels.

\n

A difference that is important to take into consideration is that there is more variability between cell labels in the test set than in the training set (i.e. in the training set there is a much higher fraction of images where all cells have identical labels). \"This competition is a weakly-labeled challenge: based on the image level label, you build models to predict labels for each individual cell. For this competition, we specifically annotated every single cell in a subset of images with higher SCV compared to public HPA images, to act as test set. In other words, the test set will have higher SCV than the training set. (SCV = Single Cell Variability)\" as quoted from https://www.kaggle.com/lnhtrang/single-cell-patterns/notebook#Protein-localization-patterns.

\n

I hope this helps to clear things up.

", "votes": 5, "replies": [] }, { "id": 1211154, "author_name": "Arka Saha", "author_url": "", "post_date": "2021-02-20T02:14:30.830000", "content": "

Hi @emmalumpan, thanks for clearing it up.
\nJust to make sure I understand the problem correctly, the answer following question by @zstusnoopy :

\n
\n

Do you mean that every cell may has 3 or 4 in the image of label 0|1😨

\n
\n

is that it might be the case, but it's unintentional and mostly due to human error of labeling?

", "votes": 0, "replies": [] }, { "id": 1211173, "author_name": "KaizaburoChubachi", "author_url": "", "post_date": "2021-02-20T03:07:17.360000", "content": "

Thanks for that clarification @emmalumpan.

", "votes": 0, "replies": [] }, { "id": 1211239, "author_name": "Wei Ouyang", "author_url": "", "post_date": "2021-02-20T04:56:50.237000", "content": "
\n
\n

Do you mean that every cell may has 3 or 4 in the image of label 0|1😨

\n
\n

is that it might be the case, but it's unintentional and mostly due to human error of labelling?

\n
\n

@arka47 Yes, it can be the case but not common or rare for the training dataset. Basically, the single cells can have inconsistent or added labels compared to the image-level labels, due to unintentional labelling error, or sometimes SVC (Single Cell Variability). (Although, I wouldn't say every cell in the image but perhaps a few cells).

\n

However, as Emma pointed out, the test set has higher SVC than the training set, as a result, you should expect higher inconsistency between the cell-level and image-level labels in the test set, including the case where new labels are added to the cell labels (e.g. your predicted image-level labels may have 0 and 1, but the cell may actually have 0 and 4).

\n

This means, during inference, you should NOT completely exclude labels that are outside the range of your image-level predictions.

", "votes": 1, "replies": [] } ] }, { "id": 1207506, "author_name": "Emma Lundberg", "author_url": "", "post_date": "2021-02-17T21:59:57.927000", "content": "

Each cell can have multiple labels.

", "votes": 0, "replies": [ { "id": 1209843, "author_name": "adeeb10abbas", "author_url": "", "post_date": "2021-02-19T04:22:07.667000", "content": "

So, according to my understanding, the model should be able to --
\n1) segment the cells from the image
\n2) Segment the organelles inside the cell
\n Please correct me if I am wrong

", "votes": 0, "replies": [] }, { "id": 1210065, "author_name": "Emma Lundberg", "author_url": "", "post_date": "2021-02-19T07:23:49.597000", "content": "

Hi,

\n

Yes you are nearly correct. The model should be able to
\n1) Segment the cells from the image
\n2) Classify the protein localization pattern(s) for each cell.

\n

Regarding #2 you could segment the different organelles as part of your approach, but we only require labels per cell for submission.

", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "1207564": "The single cell labels can also contain labels which are not included in the image-level labels, so it's not strictly subset of image-level labels.", "1207423": "I just want to be clarified about what am I predicting. Do each cell has only single class included in the image-level labels or they can have multiple classes as the subset of image-level labels?", "1207506": "Each cell can have multiple labels." } }