{ "id": 47864, "title": "Availability of the images & how to cite", "url": "/competitions/data-science-bowl-2018/discussion/47864", "author_name": "Anne Carpenter", "post_date": "2018-01-19T22:29:28.826000", "votes": 54, "comment_count": 42, "views": 0, "content": "

Great news! I would like to confirm that all images for the challenge were contributed under Creative Commons license 0 (CC0 public domain). Although many other image sets exist out there we carefully collected only CC0 images so that the set would be maximally useful for future work - enjoy!

\n\n

Credits for the images are provided in this thread: https://www.kaggle.com/c/data-science-bowl-2018/discussion/54759

\n\n

Where to get the data set: The data set will live on at the BBBC, Broad Institute Broad Bioimage Benchmark Collection; we plan to release updated versions as well, with corrections: https://data.broadinstitute.org/bbbc/BBBC038/

\n\n

How to cite the data set: \"We used image set BBBC038v1 [Caicedo et al. https://www.ncbi.nlm.nih.gov/pubmed/31636459], available from the Broad Bioimage Benchmark Collection.\"

", "messages": [ { "id": 271223, "postDate": "2018-01-19T22:29:28.827Z", "content": "

Great news! I would like to confirm that all images for the challenge were contributed under Creative Commons license 0 (CC0 public domain). Although many other image sets exist out there we carefully collected only CC0 images so that the set would be maximally useful for future work - enjoy!

\n\n

Credits for the images are provided in this thread: https://www.kaggle.com/c/data-science-bowl-2018/discussion/54759

\n\n

Where to get the data set: The data set will live on at the BBBC, Broad Institute Broad Bioimage Benchmark Collection; we plan to release updated versions as well, with corrections: https://data.broadinstitute.org/bbbc/BBBC038/

\n\n

How to cite the data set: \"We used image set BBBC038v1 [Caicedo et al. https://www.ncbi.nlm.nih.gov/pubmed/31636459], available from the Broad Bioimage Benchmark Collection.\"

", "rawMarkdown": "Great news! I would like to confirm that all images for the challenge were contributed under Creative Commons license 0 (CC0 public domain). Although many other image sets exist out there we carefully collected only CC0 images so that the set would be maximally useful for future work - enjoy!\n\n**Credits for the images** are provided in this thread: https://www.kaggle.com/c/data-science-bowl-2018/discussion/54759\n\n**Where to get the data set**: The data set will live on at the BBBC, Broad Institute Broad Bioimage Benchmark Collection; we plan to release updated versions as well, with corrections: https://data.broadinstitute.org/bbbc/BBBC038/ \n\n**How to cite the data set**: \"We used image set BBBC038v1 [Caicedo et al. https://www.ncbi.nlm.nih.gov/pubmed/31636459], available from the Broad Bioimage Benchmark Collection.\"", "votes": 54 }, { "id": 279608, "postDate": "2018-02-08T10:51:46.923Z", "content": "

I have one question about RLE for submission. We have two ways to obtain the result: segmentation (left side) and instance segmentation (right side) in the

\n\n

\"Result\"

\n\n

Note that, these images are same only difference is that the left image summarizes all masks into a single image, while the right image distinguishes mask by mask. For example, you often see the left image in result of UNet, while the right image is in Mask-RCNN. So, we will use one of them to encode and submit. Which image should I chose to provide better performance? Thanks

", "rawMarkdown": "I have one question about RLE for submission. We have two ways to obtain the result: segmentation (left side) and instance segmentation (right side) in the\n\n![Result][1]\n \nNote that, these images are same only difference is that the left image summarizes all masks into a single image, while the right image distinguishes mask by mask. For example, you often see the left image in result of UNet, while the right image is in Mask-RCNN. So, we will use one of them to encode and submit. Which image should I chose to provide better performance? Thanks\n [1]: https://i.stack.imgur.com/ajYf8.png \"Left image: Segmentation , right image : Instance segmentation\"", "votes": 3, "replies": [ { "id": 279907, "postDate": "2018-02-08T22:57:54.207Z", "content": "

I was wondering about this too. Afaik the challenge is technically instance segmentation so a model like Mask-RCNN would usually work better, but since there is only one class and the predictions are to be submitted with run length encoding its unclear to me what the benefit of using instance segmentation over regular segmentation would be.

", "rawMarkdown": "I was wondering about this too. Afaik the challenge is technically instance segmentation so a model like Mask-RCNN would usually work better, but since there is only one class and the predictions are to be submitted with run length encoding its unclear to me what the benefit of using instance segmentation over regular segmentation would be. ", "votes": 1 }, { "id": 280163, "postDate": "2018-02-09T13:53:10.257Z", "content": "

Definitely, absolutely the one on the right. This is an instance segmentation problem.

\n\n

Distinguishing foreground/background is a relatively easy problem in most of these images. The hard part is separating touching nuclei. That's why the metric focuses on that problem.

", "rawMarkdown": "Definitely, absolutely the one on the right. This is an instance segmentation problem.\n\nDistinguishing foreground/background is a relatively easy problem in most of these images. The hard part is separating touching nuclei. That's why the metric focuses on that problem.", "votes": 4 }, { "id": 280178, "postDate": "2018-02-09T14:15:50.800Z", "content": "

Thanks for your information. However, if we use RLE to encode the result, the ouputs of RLE will be same for both right and left images. I used same input to generate left and right image

", "rawMarkdown": "Thanks for your information. However, if we use RLE to encode the result, the ouputs of RLE will be same for both right and left images. I used same input to generate left and right image" }, { "id": 280211, "postDate": "2018-02-09T15:08:35.733Z", "content": "

Thanks for the clarification Anne. I think the way I was formatting the submission file was throwing me off. I had only one RLE per image ID but taking what you said into account and looking at the example submission file from the notes again I see there can be multiple of the same image ID. So if I'm not still getting this wrong, if I detect 10 nuclei in some image 780659ca805eac then I would have 10 records in the submission file with that ID and RLE for each nucleus:

\n\n
ImageId,EncodedPixels  \n780659ca805eac,1 1\n780659ca805eac,2 3 8 9  \n....\n...\n...\n...\netc...\n
\n\n

Does this sound correct or am I still way off?

", "rawMarkdown": "Thanks for the clarification Anne. I think the way I was formatting the submission file was throwing me off. I had only one RLE per image ID but taking what you said into account and looking at the example submission file from the notes again I see there can be multiple of the same image ID. So if I'm not still getting this wrong, if I detect 10 nuclei in some image 780659ca805eac then I would have 10 records in the submission file with that ID and RLE for each nucleus:\n\n ImageId,EncodedPixels \n 780659ca805eac,1 1\n 780659ca805eac,2 3 8 9 \n ....\n ...\n ...\n ...\n etc...\n\nDoes this sound correct or am I still way off?\n", "votes": 3 }, { "id": 295956, "postDate": "2018-03-14T13:28:21.570Z", "content": "

@john1231983 thanks for bringing it up, so I don't understand, should we use RLE to generate CSV submission files or not? Both left and right will generate the same CSV tho.

", "rawMarkdown": "@john1231983 thanks for bringing it up, so I don't understand, should we use RLE to generate CSV submission files or not? Both left and right will generate the same CSV tho. " }, { "id": 295997, "postDate": "2018-03-14T15:00:20.007Z", "content": "

@Nicole: No. It is different. The left image gives you a binary image (maximum intensity is 1), so it cannot distinguish between cells. The right image gives you an instance segmentation (maximum intensity is the number of the cell), hence you can separate cell by cell. For each cell, you will encode it by using RLE likes

\n\n

```

\n\n
ID, encodeRLE\n\n1, 1 2 3\n\n1, 4 5 6\n\n1, 7 8 9\n
\n\n

```

\n\n

If you use the first image, it will give you like

\n\n

```

\n\n
ID, encodeRLE\n\n1, 1 2 3 4 5 6 7 8 9\n
\n\n

```

\n\n

The second image is the goal of the challenge

", "rawMarkdown": "@Nicole: No. It is different. The left image gives you a binary image (maximum intensity is 1), so it cannot distinguish between cells. The right image gives you an instance segmentation (maximum intensity is the number of the cell), hence you can separate cell by cell. For each cell, you will encode it by using RLE likes\n\n```\n\n ID, encodeRLE\n\n 1, 1 2 3\n\n 1, 4 5 6\n\n 1, 7 8 9\n\n```\n\nIf you use the first image, it will give you like\n\n```\n\n ID, encodeRLE\n\n 1, 1 2 3 4 5 6 7 8 9\n\n```\n\nThe second image is the goal of the challenge" }, { "id": 296045, "postDate": "2018-03-14T16:26:16.993Z", "content": "

@john1231983 cool, thanks for your confirmation. Then it's a similar approach as the code we STOLE from one of the kernels. :) Just we use skimage.morphology.label() to label segments with a pixel cutoff 0.5 since our output is sigmoid (U-net) that only predicts the probabilities of pixels. I need to revisit the approach to see if we should change the cutoff. So I guess your RLE code would be a bit different if you already have indices of segments. Thank you for sharing.

", "rawMarkdown": "@john1231983 cool, thanks for your confirmation. Then it's a similar approach as the code we STOLE from one of the kernels. :) Just we use `skimage.morphology.label()` to label segments with a pixel cutoff 0.5 since our output is sigmoid (U-net) that only predicts the probabilities of pixels. I need to revisit the approach to see if we should change the cutoff. So I guess your RLE code would be a bit different if you already have indices of segments. Thank you for sharing. \n" } ] }, { "id": 314778, "postDate": "2018-04-16T09:03:04.273Z", "content": "

Hello, Anne, if I didn't upload my model by now(cause I missed the deadlines) ,does this influence my final LB position?

", "rawMarkdown": "Hello, Anne, if I didn't upload my model by now(cause I missed the deadlines) ,does this influence my final LB position?", "votes": 1, "replies": [ { "id": 314786, "postDate": "2018-04-16T09:15:52.743Z", "content": "

If I recall the FAQ correctly, it will not change your LB position, you can still earn a medal, but you are out of the prize pool (ie if you won, you would not have the right for the 50K and groovy computer, that'd go on the 2nd place and so forth). https://www.kaggle.com/two-stage-faq

", "rawMarkdown": "If I recall the FAQ correctly, it will not change your LB position, you can still earn a medal, but you are out of the prize pool (ie if you won, you would not have the right for the 50K and groovy computer, that'd go on the 2nd place and so forth). https://www.kaggle.com/two-stage-faq ", "votes": 1 }, { "id": 314862, "postDate": "2018-04-16T11:35:41.543Z", "content": "

Got it! thx a lot!

", "rawMarkdown": "Got it! thx a lot!" } ] }, { "id": 276222, "postDate": "2018-01-31T02:00:48.610Z", "content": "

Hello，Anne，about the stage2，just release the test data，or test data and train data

", "rawMarkdown": "Hello，Anne，about the stage2，just release the test data，or test data and train data", "votes": 1, "replies": [ { "id": 276606, "postDate": "2018-01-31T22:08:35.360Z", "content": "

As noted on the Data tab: https://www.kaggle.com/c/data-science-bowl-2018/data

\n\n

\"The masks of the stage 1 test set will be released with the release of the stage 2 test set.\"

", "rawMarkdown": "As noted on the Data tab: https://www.kaggle.com/c/data-science-bowl-2018/data\n\n\"The masks of the stage 1 test set will be released with the release of the stage 2 test set.\"", "votes": 1 }, { "id": 276696, "postDate": "2018-02-01T07:20:20.413Z", "content": "

Thank you very much!

", "rawMarkdown": "Thank you very much!" }, { "id": 311495, "postDate": "2018-04-10T06:39:15.590Z", "content": "

~

", "rawMarkdown": "~" } ] }, { "id": 271415, "postDate": "2018-01-20T11:53:45.473Z", "content": "
\n
\n

@drannecarpenter Thanks a lot for letting us know about this!

\n
\n
", "rawMarkdown": ">> @drannecarpenter Thanks a lot for letting us know about this!", "votes": 1 }, { "id": 298752, "postDate": "2018-03-20T01:16:56.260Z", "content": "

I have a question about how to upload our final model into kaggle as stage 1 deadline approach? Thank you

", "rawMarkdown": " I have a question about how to upload our final model into kaggle as stage 1 deadline approach? Thank you", "votes": 2, "replies": [ { "id": 302050, "postDate": "2018-03-23T16:12:41.007Z", "content": "

Please search the other threads or ask in a new thread, since this Q isn't relevant to this thread's topic.

", "rawMarkdown": "Please search the other threads or ask in a new thread, since this Q isn't relevant to this thread's topic." }, { "id": 302070, "postDate": "2018-03-23T17:06:08.380Z", "content": "

@Jianhua, you can use this thread\nhttps://www.kaggle.com/c/data-science-bowl-2018/discussion/52385#298629

", "rawMarkdown": "@Jianhua, you can use this thread\nhttps://www.kaggle.com/c/data-science-bowl-2018/discussion/52385#298629" } ] }, { "id": 660581, "postDate": "2019-10-29T11:04:35.310Z", "content": "

The paper describing the challenge is now out! I have edited the pinned comment with how to cite the dataset.

", "rawMarkdown": "The paper describing the challenge is now out! I have edited the pinned comment with how to cite the dataset. " }, { "id": 314473, "postDate": "2018-04-15T16:27:22.200Z", "content": "

Hello! Sorry for offtop. Anne, could you send me the results of the first stage of the competition Data Science Bowl 2018? I forgot to save my final results for stage 1, and now it's stage 2, and the previous results were deleted. I need this to confirm my LB score for the online course.

", "rawMarkdown": "Hello! Sorry for offtop. Anne, could you send me the results of the first stage of the competition Data Science Bowl 2018? I forgot to save my final results for stage 1, and now it's stage 2, and the previous results were deleted. I need this to confirm my LB score for the online course.", "replies": [ { "id": 314793, "postDate": "2018-04-16T09:32:00.030Z", "content": "

The results for Stage 1 are currently available for download from the Data tab, as a file called stage1_solution.csv.zip. You can download this file and use it to compare the actual results against your solution, then score it against the provided metric to get the leaderboard score.

", "rawMarkdown": "The results for Stage 1 are currently available for download from the Data tab, as a file called stage1_solution.csv.zip. You can download this file and use it to compare the actual results against your solution, then score it against the provided metric to get the leaderboard score." } ] }, { "id": 313590, "postDate": "2018-04-13T14:29:35.273Z", "content": "

Just out of curiosity, please advise on the cells the 2 drawings in the Stage 2 test set (170bc41b2095177cccd3d4c8977c619147580f1d93b4fe9701eddd77736d4ece.png & 5390acefd575cf9b33413ddf6cbb9ce137ae07dc04616ba24c7b5fe476c827d2.png) are supposed to represent? Thanks.

", "rawMarkdown": "Just out of curiosity, please advise on the cells the 2 drawings in the Stage 2 test set (**170bc41b2095177cccd3d4c8977c619147580f1d93b4fe9701eddd77736d4ece.png** & **5390acefd575cf9b33413ddf6cbb9ce137ae07dc04616ba24c7b5fe476c827d2.png**) are supposed to represent? Thanks.", "replies": [ { "id": 314792, "postDate": "2018-04-16T09:29:22.010Z", "content": "

@Anne/any of the organizers: Please assist to address this query. Thanks.

", "rawMarkdown": "@Anne/any of the organizers: Please assist to address this query. Thanks." } ] }, { "id": 311896, "postDate": "2018-04-10T21:24:04.477Z", "content": "

After the competition is over, can/will the train/test datasets be available at https://www.kaggle.com/datasets? At this point, I have no hope of having a model ready by the deadline, but I still want to build one once I get my tools working and see how it would have scored on stage 2.

", "rawMarkdown": "After the competition is over, can/will the train/test datasets be available at https://www.kaggle.com/datasets? At this point, I have no hope of having a model ready by the deadline, but I still want to build one once I get my tools working and see how it would have scored on stage 2.", "replies": [ { "id": 311944, "postDate": "2018-04-11T00:30:34.553Z", "content": "

I'm not certain about that stage of the Kaggle platform but I can assure you the data will remain publicly available one way or the other! We have a copy in case it becomes unavailable at Kaggle.

", "rawMarkdown": "I'm not certain about that stage of the Kaggle platform but I can assure you the data will remain publicly available one way or the other! We have a copy in case it becomes unavailable at Kaggle." } ] }, { "id": 301916, "postDate": "2018-03-23T13:07:45.103Z", "content": "

Awesome work!!!!

", "rawMarkdown": "Awesome work!!!!" }, { "id": 301743, "postDate": "2018-03-23T07:22:26.787Z", "content": "

wow

", "rawMarkdown": "wow" }, { "id": 276228, "postDate": "2018-01-31T02:13:06.957Z", "content": "

Hi Anne, does that mean we could still continue using external datasets?

\n\n

There's been a few challenges over the years some still have active links

\n\n
    \n
  1. http://ludo17.free.fr/mitos_2012/download.html
    2. \n
  2. https://mitos-atypia-14.grand-challenge.org/donwload/
    3. \n
", "rawMarkdown": "Hi Anne, does that mean we could still continue using external datasets? \n\nThere's been a few challenges over the years some still have active links\n\n 1. http://ludo17.free.fr/mitos_2012/download.html\n 2. https://mitos-atypia-14.grand-challenge.org/donwload/\n\n", "replies": [ { "id": 276603, "postDate": "2018-01-31T22:05:52.883Z", "content": "

There is a thread for external data sets here: https://www.kaggle.com/c/data-science-bowl-2018/discussion/47572

", "rawMarkdown": "There is a thread for external data sets here: https://www.kaggle.com/c/data-science-bowl-2018/discussion/47572" }, { "id": 311497, "postDate": "2018-04-10T06:39:47.030Z", "content": "

~

", "rawMarkdown": "~" } ] }, { "id": 273205, "postDate": "2018-01-24T09:10:34.367Z", "content": "

Thanks Anne! Just a few quick questions:

\n\n
    \n
  1. How does a human expert recognize the Nuclei /Entire cell?\nCan you post some links to lectures/info on this topic for people without any background in biology?\nIn your expert opinion, how hard is it for anyone to become an expert when it comes to detecting cell/nuclei?
    2. \n
\n\n

2.You just said \"many other image sets exist\" - where can we find more relevant image sets, labelled or unlabeled?

", "rawMarkdown": "Thanks Anne! Just a few quick questions:\n\n1. How does a human expert recognize the Nuclei /Entire cell?\nCan you post some links to lectures/info on this topic for people without any background in biology?\nIn your expert opinion, how hard is it for anyone to become an expert when it comes to detecting cell/nuclei?\n\n2.You just said \"many other image sets exist\" - where can we find more relevant image sets, labelled or unlabeled?", "replies": [ { "id": 273422, "postDate": "2018-01-24T15:47:32.207Z", "content": "
\n

How does a human expert recognize the Nuclei /Entire cell? Can you post some links to lectures/info on this topic for people without any background in biology? In your expert opinion, how hard is it for anyone to become an expert when it comes to detecting cell/nuclei?

\n
\n\n

I can’t provide any exact details about how biologists recognize whether an object is or isn’t a nucleus. However, from my experience with working alongside biologists I’ve started to understand:

\n\n\n\n
\n

You just said \"many other image sets exist\" - where can we find more relevant image sets, labelled or unlabeled?

\n
\n\n

We have a thread for this exact subject. A number of images have already been shared!

", "rawMarkdown": "> How does a human expert recognize the Nuclei /Entire cell? Can you post some links to lectures/info on this topic for people without any background in biology? In your expert opinion, how hard is it for anyone to become an expert when it comes to detecting cell/nuclei?\n\nI can’t provide any exact details about how biologists recognize whether an object is or isn’t a nucleus. However, from my experience with working alongside biologists I’ve started to understand:\n\n* They’ve looked at many microscopy images in their lifetimes. I wouldn’t be surprised if Anne has seen tens (if not hundreds) of thousands of images in her lifetime!\n\n* They use contextual information that’s non-obvious to non-biologists (e.g. me). I often use the following analogy from computer vision and machine learning to explain this: it’s a “scene understanding” process rather than an object detection process. \n\n* It’s not particularly difficult to annotate an image. If you’re annotating bounding boxes (where overlapping objects is permitted), it’s straightforward. However, I’ve found that per-pixel annotation is difficult in situations where the cellular compartments (like nuclei) are touching. It’s also difficult to differentiate between debris and micro-nuclei or a dividing cell and overlapping cells.\n\n> You just said \"many other image sets exist\" - where can we find more relevant image sets, labelled or unlabeled?\n\nWe have a [thread][1] for this exact subject. A number of images have already been shared!\n\n [1]: https://www.kaggle.com/c/data-science-bowl-2018/discussion/47572\n\n", "votes": 3 }, { "id": 273569, "postDate": "2018-01-24T20:14:52.307Z", "content": "

Hi Yoni,\nFormer biologist here. You are basically looking for a bounded inner ovoid blob. It may be stained (colored) differently from the cellular background, darker than the cellular background, or unstained (lighter) then the cellular background. The hard part for humans, as well as the algorithm, is correct ID across conditions. I may get 98% in my own samples, where I know the stains, cell line, and microscopy, but only get 30% in another lab's samples.

", "rawMarkdown": "Hi Yoni,\nFormer biologist here. You are basically looking for a bounded inner ovoid blob. It may be stained (colored) differently from the cellular background, darker than the cellular background, or unstained (lighter) then the cellular background. The hard part for humans, as well as the algorithm, is correct ID across conditions. I may get 98% in my own samples, where I know the stains, cell line, and microscopy, but only get 30% in another lab's samples. ", "votes": 6 }, { "id": 273662, "postDate": "2018-01-25T00:06:11.733Z", "content": "
\n

How does a human expert recognize the Nuclei /Entire cell?

\n
\n\n

I'll take a stab at this one, with my usual disclaimer: I am not a cell biologist - I only play one on TV.

\n\n

Cells have incredibly complex textures, both inside and outside. Therefore, they almost never look completely homogenous, regardless of the exact stain used. However, it is important to note that different staining techniques and different microscopes will provide views of the cell that can be very unique.

\n\n

First, an amendment to @Allen Goodman already described: looking at cells and recognizing them is definitely a skill, but not in any kind of voodoo way. Cell types are different, so it helps knowing which ones we are dealing with. If you have dealt with a handful of cell types you would still be at a disadvantage compared to someone who has worked with dozens of different cells. Also, it helps knowing what cellular structure or what cellular protein is stained/labeled, as it makes the interpretation easier. One trick that helps is comparing brightfield and fluorescent images of the same field of cells. You may or may not be able to see cell contours in fluorescent images if cells are unlabeled or only partially labeled, but they are very defined in brightfield images.

\n\n

Here is how that looks like side by side (brightfield on the left):

\n\n

\"Cells

\n\n

Strong signals in the brightfield images, especially the ones that are not residing within cell boundaries, are potentially not cells. I've boxed one of those on the right. However, we don't have the benefit of accompanying brightfield images in this challenge.

\n\n

To expand a bit on my earlier point that cellular structures tend to be nuanced, here is an image of feature-rich nuclei:

\n\n

\"Granular

\n\n

Note also that it is not always nuclei that are stained. I am kind of partial to images that stain whole cells because it is easier to see the richness of their shapes.

\n\n

\"Whole

\n\n

Cells on the left have predominantly stained nuclei, but the rest of cell has a good deal of signal as well. Cells on the right would be the opposite of most images in our collection here, as the nuclei are excluded from staining.

\n\n

Note the \"hollow\" parts of nuclei in both of the previous two images, which seem like holes. In this particular case those are nucleoli. This further supports the idea that nuclear staining may not be uniform, and it often features either these \"holes\" or sometimes the inverse of it - the \"holes\" are stained more prominently than the rest.

\n\n

\"Inversion\"

\n\n

If you have multiple stains to use on nucleus and on everything else, then it is fairly easy to identify cells.

\n\n

\"Double-stained

\n\n

The point is that it is fairly easy to identify nuclei if: 1) you see enough images of the same cell type (like the ones above); 2) you have both bright and fluorescent images; 3) the images are focused; 4) you have a pretty good idea what part of the cell will be stained.

\n\n

The challenge we have is to annotate nuclei from different types of experiments (histology, fluorescence, brightfield), different types of cells, inconsistent magnifications, and with some images that are out of focus.

\n\n

Somewhat related to this topic: there is an ongoing discussion here about what I think are problems with image annotations. I still stand by everything I argued there with the exception of figure 3, which was explained by @Anne Carpenter to be unusual cells with vacuoles. While it is almost guaranteed that Dr Carpenter and her crew have greater cell biology expertise than anyone on Kaggle (including yours truly), it is up to you to decide whether the ground-truth is set correctly in all the images. On the other hand, maybe there is nothing to decide: given that the same experts most likely have annotated the stage 2 dataset, maybe we just ignore small inconsistencies and simply go with provided annotations.

", "rawMarkdown": "> How does a human expert recognize the Nuclei /Entire cell?\n\nI'll take a stab at this one, with my usual disclaimer: I am not a cell biologist - I only play one on TV.\n\nCells have incredibly complex textures, both inside and outside. Therefore, they almost never look completely homogenous, regardless of the exact stain used. However, it is important to note that different staining techniques and different microscopes will provide views of the cell that can be very unique.\n\nFirst, an amendment to @Allen Goodman already described: looking at cells and recognizing them is definitely a skill, but not in any kind of voodoo way. Cell types are different, so it helps knowing which ones we are dealing with. If you have dealt with a handful of cell types you would still be at a disadvantage compared to someone who has worked with dozens of different cells. Also, it helps knowing what cellular structure or what cellular protein is stained/labeled, as it makes the interpretation easier. One trick that helps is comparing brightfield and fluorescent images of the same field of cells. You may or may not be able to see cell contours in fluorescent images if cells are unlabeled or only partially labeled, but they are very defined in brightfield images.\n\nHere is how that looks like side by side (brightfield on the left):\n\n![Cells side by side][1]\n\nStrong signals in the brightfield images, especially the ones that are not residing within cell boundaries, are potentially not cells. I've boxed one of those on the right. However, we don't have the benefit of accompanying brightfield images in this challenge.\n\nTo expand a bit on my earlier point that cellular structures tend to be nuanced, here is an image of feature-rich nuclei:\n\n![Granular nuclei][2]\n\nNote also that it is not always nuclei that are stained. I am kind of partial to images that stain whole cells because it is easier to see the richness of their shapes.\n\n![Whole cell staining][3]\n\nCells on the left have predominantly stained nuclei, but the rest of cell has a good deal of signal as well. Cells on the right would be the opposite of most images in our collection here, as the nuclei are excluded from staining.\n\nNote the \"hollow\" parts of nuclei in both of the previous two images, which seem like holes. In this particular case those are nucleoli. This further supports the idea that nuclear staining may not be uniform, and it often features either these \"holes\" or sometimes the inverse of it - the \"holes\" are stained more prominently than the rest.\n\n![Inversion][4]\n\nIf you have multiple stains to use on nucleus and on everything else, then it is fairly easy to identify cells.\n\n![Double-stained cells][5]\n\nThe point is that it is fairly easy to identify nuclei if: 1) you see enough images of the same cell type (like the ones above); 2) you have both bright and fluorescent images; 3) the images are focused; 4) you have a pretty good idea what part of the cell will be stained.\n\nThe challenge we have is to annotate nuclei from different types of experiments (histology, fluorescence, brightfield), different types of cells, inconsistent magnifications, and with some images that are out of focus.\n\nSomewhat related to this topic: there is an ongoing discussion [__here__](https://www.kaggle.com/c/data-science-bowl-2018/discussion/47857) about what I think are problems with image annotations. I still stand by everything I argued there with the exception of figure 3, which was explained by @Anne Carpenter to be unusual cells with vacuoles. While it is almost guaranteed that Dr Carpenter and her crew have greater cell biology expertise than anyone on Kaggle (including yours truly), it is up to you to decide whether the ground-truth is set correctly in all the images. On the other hand, maybe there is nothing to decide: given that the same experts most likely have annotated the stage 2 dataset, maybe we just ignore small inconsistencies and simply go with provided annotations.\n\n\n [1]: https://i.imgur.com/UDlYMXG.png\n [2]: https://i.imgur.com/VaS5ilN.png\n [3]: https://i.imgur.com/r0WbWSL.png\n [4]: https://i.imgur.com/9lijaYz.png\n [5]: https://i.imgur.com/Bt6Q3F7.png", "votes": 18 }, { "id": 273670, "postDate": "2018-01-25T00:59:30.630Z", "content": "

Really awesome answer by Tilii (as always)

\n\n

Thank you !!

", "rawMarkdown": "Really awesome answer by Tilii (as always)\n\nThank you !!" }, { "id": 273688, "postDate": "2018-01-25T01:57:01.567Z", "content": "

Thank for explaining Tilii. This is a interesting challenge.

", "rawMarkdown": "Thank for explaining Tilii. This is a interesting challenge.", "votes": 1 }, { "id": 273794, "postDate": "2018-01-25T06:39:39.877Z", "content": "

Great explanation, Tilii! It is indeed a very interesting and challenging problem.

", "rawMarkdown": "Great explanation, Tilii! It is indeed a very interesting and challenging problem.", "votes": 1 }, { "id": 278266, "postDate": "2018-02-05T18:46:01.440Z", "content": "

Very clear explanation, thank you.

", "rawMarkdown": "Very clear explanation, thank you.", "votes": 1 } ] }, { "id": 313342, "postDate": "2018-04-13T06:03:33.070Z", "rawMarkdown": "", "isDeleted": true }, { "id": 274999, "postDate": "2018-01-28T01:28:03.377Z", "content": "

Thanks Tilii, great explanation.

", "rawMarkdown": " Thanks Tilii, great explanation.", "votes": 1 }, { "id": 302311, "postDate": "2018-03-23T22:28:49.343Z", "content": "

thanks

", "rawMarkdown": "thanks" } ], "comments": [ { "id": 279608, "author_name": "john1231983", "author_url": "", "post_date": "2018-02-08T10:51:46.923000", "content": "

I have one question about RLE for submission. We have two ways to obtain the result: segmentation (left side) and instance segmentation (right side) in the

\n\n

\"Result\"

\n\n

Note that, these images are same only difference is that the left image summarizes all masks into a single image, while the right image distinguishes mask by mask. For example, you often see the left image in result of UNet, while the right image is in Mask-RCNN. So, we will use one of them to encode and submit. Which image should I chose to provide better performance? Thanks

", "votes": 3, "replies": [ { "id": 279907, "author_name": "Jack Vial", "author_url": "", "post_date": "2018-02-08T22:57:54.207000", "content": "

I was wondering about this too. Afaik the challenge is technically instance segmentation so a model like Mask-RCNN would usually work better, but since there is only one class and the predictions are to be submitted with run length encoding its unclear to me what the benefit of using instance segmentation over regular segmentation would be.

", "votes": 1, "replies": [] }, { "id": 280163, "author_name": "Anne Carpenter", "author_url": "", "post_date": "2018-02-09T13:53:10.257000", "content": "

Definitely, absolutely the one on the right. This is an instance segmentation problem.

\n\n

Distinguishing foreground/background is a relatively easy problem in most of these images. The hard part is separating touching nuclei. That's why the metric focuses on that problem.

", "votes": 4, "replies": [] }, { "id": 280178, "author_name": "john1231983", "author_url": "", "post_date": "2018-02-09T14:15:50.800000", "content": "

Thanks for your information. However, if we use RLE to encode the result, the ouputs of RLE will be same for both right and left images. I used same input to generate left and right image

", "votes": 0, "replies": [] }, { "id": 280211, "author_name": "Jack Vial", "author_url": "", "post_date": "2018-02-09T15:08:35.733000", "content": "

Thanks for the clarification Anne. I think the way I was formatting the submission file was throwing me off. I had only one RLE per image ID but taking what you said into account and looking at the example submission file from the notes again I see there can be multiple of the same image ID. So if I'm not still getting this wrong, if I detect 10 nuclei in some image 780659ca805eac then I would have 10 records in the submission file with that ID and RLE for each nucleus:

\n\n
ImageId,EncodedPixels  \n780659ca805eac,1 1\n780659ca805eac,2 3 8 9  \n....\n...\n...\n...\netc...\n
\n\n

Does this sound correct or am I still way off?

", "votes": 3, "replies": [] }, { "id": 295956, "author_name": "Nicole Finnie", "author_url": "", "post_date": "2018-03-14T13:28:21.570000", "content": "

@john1231983 thanks for bringing it up, so I don't understand, should we use RLE to generate CSV submission files or not? Both left and right will generate the same CSV tho.

", "votes": 0, "replies": [] }, { "id": 295997, "author_name": "john1231983", "author_url": "", "post_date": "2018-03-14T15:00:20.007000", "content": "

@Nicole: No. It is different. The left image gives you a binary image (maximum intensity is 1), so it cannot distinguish between cells. The right image gives you an instance segmentation (maximum intensity is the number of the cell), hence you can separate cell by cell. For each cell, you will encode it by using RLE likes

\n\n

```

\n\n
ID, encodeRLE\n\n1, 1 2 3\n\n1, 4 5 6\n\n1, 7 8 9\n
\n\n

```

\n\n

If you use the first image, it will give you like

\n\n

```

\n\n
ID, encodeRLE\n\n1, 1 2 3 4 5 6 7 8 9\n
\n\n

```

\n\n

The second image is the goal of the challenge

", "votes": 0, "replies": [] }, { "id": 296045, "author_name": "Nicole Finnie", "author_url": "", "post_date": "2018-03-14T16:26:16.993000", "content": "

@john1231983 cool, thanks for your confirmation. Then it's a similar approach as the code we STOLE from one of the kernels. :) Just we use skimage.morphology.label() to label segments with a pixel cutoff 0.5 since our output is sigmoid (U-net) that only predicts the probabilities of pixels. I need to revisit the approach to see if we should change the cutoff. So I guess your RLE code would be a bit different if you already have indices of segments. Thank you for sharing.

", "votes": 0, "replies": [] } ] }, { "id": 314778, "author_name": "dingxianghu", "author_url": "", "post_date": "2018-04-16T09:03:04.273000", "content": "

Hello, Anne, if I didn't upload my model by now(cause I missed the deadlines) ,does this influence my final LB position?

", "votes": 1, "replies": [ { "id": 314786, "author_name": "Quark02", "author_url": "", "post_date": "2018-04-16T09:15:52.743000", "content": "

If I recall the FAQ correctly, it will not change your LB position, you can still earn a medal, but you are out of the prize pool (ie if you won, you would not have the right for the 50K and groovy computer, that'd go on the 2nd place and so forth). https://www.kaggle.com/two-stage-faq

", "votes": 1, "replies": [] }, { "id": 314862, "author_name": "dingxianghu", "author_url": "", "post_date": "2018-04-16T11:35:41.543000", "content": "

Got it! thx a lot!

", "votes": 0, "replies": [] } ] }, { "id": 276222, "author_name": "HaoChen", "author_url": "", "post_date": "2018-01-31T02:00:48.610000", "content": "

Hello，Anne，about the stage2，just release the test data，or test data and train data

", "votes": 1, "replies": [ { "id": 276606, "author_name": "Anne Carpenter", "author_url": "", "post_date": "2018-01-31T22:08:35.360000", "content": "

As noted on the Data tab: https://www.kaggle.com/c/data-science-bowl-2018/data

\n\n

\"The masks of the stage 1 test set will be released with the release of the stage 2 test set.\"

", "votes": 1, "replies": [] }, { "id": 276696, "author_name": "HaoChen", "author_url": "", "post_date": "2018-02-01T07:20:20.413000", "content": "

Thank you very much!

", "votes": 0, "replies": [] }, { "id": 311495, "author_name": "bennis", "author_url": "", "post_date": "2018-04-10T06:39:15.590000", "content": "

~

", "votes": 0, "replies": [] } ] }, { "id": 271415, "author_name": "Amit Kumar Jaiswal", "author_url": "", "post_date": "2018-01-20T11:53:45.473000", "content": "
\n
\n

@drannecarpenter Thanks a lot for letting us know about this!

\n
\n
", "votes": 1, "replies": [] }, { "id": 298752, "author_name": "Jianhua Ma", "author_url": "", "post_date": "2018-03-20T01:16:56.260000", "content": "

I have a question about how to upload our final model into kaggle as stage 1 deadline approach? Thank you

", "votes": 2, "replies": [ { "id": 302050, "author_name": "Anne Carpenter", "author_url": "", "post_date": "2018-03-23T16:12:41.007000", "content": "

Please search the other threads or ask in a new thread, since this Q isn't relevant to this thread's topic.

", "votes": 0, "replies": [] }, { "id": 302070, "author_name": "Nicole Finnie", "author_url": "", "post_date": "2018-03-23T17:06:08.380000", "content": "

@Jianhua, you can use this thread\nhttps://www.kaggle.com/c/data-science-bowl-2018/discussion/52385#298629

", "votes": 0, "replies": [] } ] }, { "id": 660581, "author_name": "Anne Carpenter", "author_url": "", "post_date": "2019-10-29T11:04:35.310000", "content": "

The paper describing the challenge is now out! I have edited the pinned comment with how to cite the dataset.

", "votes": 0, "replies": [] }, { "id": 314473, "author_name": "Freeman", "author_url": "", "post_date": "2018-04-15T16:27:22.200000", "content": "

Hello! Sorry for offtop. Anne, could you send me the results of the first stage of the competition Data Science Bowl 2018? I forgot to save my final results for stage 1, and now it's stage 2, and the previous results were deleted. I need this to confirm my LB score for the online course.

", "votes": 0, "replies": [ { "id": 314793, "author_name": "Shiraz S K", "author_url": "", "post_date": "2018-04-16T09:32:00.030000", "content": "

The results for Stage 1 are currently available for download from the Data tab, as a file called stage1_solution.csv.zip. You can download this file and use it to compare the actual results against your solution, then score it against the provided metric to get the leaderboard score.

", "votes": 0, "replies": [] } ] }, { "id": 313590, "author_name": "Shiraz S K", "author_url": "", "post_date": "2018-04-13T14:29:35.273000", "content": "

Just out of curiosity, please advise on the cells the 2 drawings in the Stage 2 test set (170bc41b2095177cccd3d4c8977c619147580f1d93b4fe9701eddd77736d4ece.png & 5390acefd575cf9b33413ddf6cbb9ce137ae07dc04616ba24c7b5fe476c827d2.png) are supposed to represent? Thanks.

", "votes": 0, "replies": [ { "id": 314792, "author_name": "Shiraz S K", "author_url": "", "post_date": "2018-04-16T09:29:22.010000", "content": "

@Anne/any of the organizers: Please assist to address this query. Thanks.

", "votes": 0, "replies": [] } ] }, { "id": 311896, "author_name": "Ian Oliver", "author_url": "", "post_date": "2018-04-10T21:24:04.477000", "content": "

After the competition is over, can/will the train/test datasets be available at https://www.kaggle.com/datasets? At this point, I have no hope of having a model ready by the deadline, but I still want to build one once I get my tools working and see how it would have scored on stage 2.

", "votes": 0, "replies": [ { "id": 311944, "author_name": "Anne Carpenter", "author_url": "", "post_date": "2018-04-11T00:30:34.553000", "content": "

I'm not certain about that stage of the Kaggle platform but I can assure you the data will remain publicly available one way or the other! We have a copy in case it becomes unavailable at Kaggle.

", "votes": 0, "replies": [] } ] }, { "id": 301916, "author_name": "ash16win", "author_url": "", "post_date": "2018-03-23T13:07:45.103000", "content": "

Awesome work!!!!

", "votes": 0, "replies": [] }, { "id": 301743, "author_name": "alancheg", "author_url": "", "post_date": "2018-03-23T07:22:26.787000", "content": "

wow

", "votes": 0, "replies": [] }, { "id": 276228, "author_name": "Wesley Goi", "author_url": "", "post_date": "2018-01-31T02:13:06.957000", "content": "

Hi Anne, does that mean we could still continue using external datasets?

\n\n

There's been a few challenges over the years some still have active links

\n\n
    \n
  1. http://ludo17.free.fr/mitos_2012/download.html
    2. \n
  2. https://mitos-atypia-14.grand-challenge.org/donwload/
    3. \n
", "votes": 0, "replies": [ { "id": 276603, "author_name": "Anne Carpenter", "author_url": "", "post_date": "2018-01-31T22:05:52.883000", "content": "

There is a thread for external data sets here: https://www.kaggle.com/c/data-science-bowl-2018/discussion/47572

", "votes": 0, "replies": [] }, { "id": 311497, "author_name": "bennis", "author_url": "", "post_date": "2018-04-10T06:39:47.030000", "content": "

~

", "votes": 0, "replies": [] } ] }, { "id": 273205, "author_name": "Yoni Keren", "author_url": "", "post_date": "2018-01-24T09:10:34.367000", "content": "

Thanks Anne! Just a few quick questions:

\n\n
    \n
  1. How does a human expert recognize the Nuclei /Entire cell?\nCan you post some links to lectures/info on this topic for people without any background in biology?\nIn your expert opinion, how hard is it for anyone to become an expert when it comes to detecting cell/nuclei?
    2. \n
\n\n

2.You just said \"many other image sets exist\" - where can we find more relevant image sets, labelled or unlabeled?

", "votes": 0, "replies": [ { "id": 273422, "author_name": "Allen Goodman", "author_url": "", "post_date": "2018-01-24T15:47:32.207000", "content": "
\n

How does a human expert recognize the Nuclei /Entire cell? Can you post some links to lectures/info on this topic for people without any background in biology? In your expert opinion, how hard is it for anyone to become an expert when it comes to detecting cell/nuclei?

\n
\n\n

I can’t provide any exact details about how biologists recognize whether an object is or isn’t a nucleus. However, from my experience with working alongside biologists I’ve started to understand:

\n\n\n\n
\n

You just said \"many other image sets exist\" - where can we find more relevant image sets, labelled or unlabeled?

\n
\n\n

We have a thread for this exact subject. A number of images have already been shared!

", "votes": 3, "replies": [] }, { "id": 273569, "author_name": "Kerry Donny-Clark", "author_url": "", "post_date": "2018-01-24T20:14:52.307000", "content": "

Hi Yoni,\nFormer biologist here. You are basically looking for a bounded inner ovoid blob. It may be stained (colored) differently from the cellular background, darker than the cellular background, or unstained (lighter) then the cellular background. The hard part for humans, as well as the algorithm, is correct ID across conditions. I may get 98% in my own samples, where I know the stains, cell line, and microscopy, but only get 30% in another lab's samples.

", "votes": 6, "replies": [] }, { "id": 273662, "author_name": "Tilii", "author_url": "", "post_date": "2018-01-25T00:06:11.733000", "content": "
\n

How does a human expert recognize the Nuclei /Entire cell?

\n
\n\n

I'll take a stab at this one, with my usual disclaimer: I am not a cell biologist - I only play one on TV.

\n\n

Cells have incredibly complex textures, both inside and outside. Therefore, they almost never look completely homogenous, regardless of the exact stain used. However, it is important to note that different staining techniques and different microscopes will provide views of the cell that can be very unique.

\n\n

First, an amendment to @Allen Goodman already described: looking at cells and recognizing them is definitely a skill, but not in any kind of voodoo way. Cell types are different, so it helps knowing which ones we are dealing with. If you have dealt with a handful of cell types you would still be at a disadvantage compared to someone who has worked with dozens of different cells. Also, it helps knowing what cellular structure or what cellular protein is stained/labeled, as it makes the interpretation easier. One trick that helps is comparing brightfield and fluorescent images of the same field of cells. You may or may not be able to see cell contours in fluorescent images if cells are unlabeled or only partially labeled, but they are very defined in brightfield images.

\n\n

Here is how that looks like side by side (brightfield on the left):

\n\n

\"Cells

\n\n

Strong signals in the brightfield images, especially the ones that are not residing within cell boundaries, are potentially not cells. I've boxed one of those on the right. However, we don't have the benefit of accompanying brightfield images in this challenge.

\n\n

To expand a bit on my earlier point that cellular structures tend to be nuanced, here is an image of feature-rich nuclei:

\n\n

\"Granular

\n\n

Note also that it is not always nuclei that are stained. I am kind of partial to images that stain whole cells because it is easier to see the richness of their shapes.

\n\n

\"Whole

\n\n

Cells on the left have predominantly stained nuclei, but the rest of cell has a good deal of signal as well. Cells on the right would be the opposite of most images in our collection here, as the nuclei are excluded from staining.

\n\n

Note the \"hollow\" parts of nuclei in both of the previous two images, which seem like holes. In this particular case those are nucleoli. This further supports the idea that nuclear staining may not be uniform, and it often features either these \"holes\" or sometimes the inverse of it - the \"holes\" are stained more prominently than the rest.

\n\n

\"Inversion\"

\n\n

If you have multiple stains to use on nucleus and on everything else, then it is fairly easy to identify cells.

\n\n

\"Double-stained

\n\n

The point is that it is fairly easy to identify nuclei if: 1) you see enough images of the same cell type (like the ones above); 2) you have both bright and fluorescent images; 3) the images are focused; 4) you have a pretty good idea what part of the cell will be stained.

\n\n

The challenge we have is to annotate nuclei from different types of experiments (histology, fluorescence, brightfield), different types of cells, inconsistent magnifications, and with some images that are out of focus.

\n\n

Somewhat related to this topic: there is an ongoing discussion here about what I think are problems with image annotations. I still stand by everything I argued there with the exception of figure 3, which was explained by @Anne Carpenter to be unusual cells with vacuoles. While it is almost guaranteed that Dr Carpenter and her crew have greater cell biology expertise than anyone on Kaggle (including yours truly), it is up to you to decide whether the ground-truth is set correctly in all the images. On the other hand, maybe there is nothing to decide: given that the same experts most likely have annotated the stage 2 dataset, maybe we just ignore small inconsistencies and simply go with provided annotations.

", "votes": 18, "replies": [] }, { "id": 273670, "author_name": "Serigne ", "author_url": "", "post_date": "2018-01-25T00:59:30.630000", "content": "

Really awesome answer by Tilii (as always)

\n\n

Thank you !!

", "votes": 0, "replies": [] }, { "id": 273688, "author_name": "William Green", "author_url": "", "post_date": "2018-01-25T01:57:01.567000", "content": "

Thank for explaining Tilii. This is a interesting challenge.

", "votes": 1, "replies": [] }, { "id": 273794, "author_name": "Mohamed Judi", "author_url": "", "post_date": "2018-01-25T06:39:39.877000", "content": "

Great explanation, Tilii! It is indeed a very interesting and challenging problem.

", "votes": 1, "replies": [] }, { "id": 278266, "author_name": "Jack Vial", "author_url": "", "post_date": "2018-02-05T18:46:01.440000", "content": "

Very clear explanation, thank you.

", "votes": 1, "replies": [] } ] }, { "id": 313342, "author_name": "", "author_url": "", "post_date": "2018-04-13T06:03:33.070000", "content": "", "votes": 0, "replies": [] }, { "id": 274999, "author_name": "YQM0nk3y", "author_url": "", "post_date": "2018-01-28T01:28:03.377000", "content": "

Thanks Tilii, great explanation.

", "votes": 1, "replies": [] }, { "id": 302311, "author_name": "Mert Arı", "author_url": "", "post_date": "2018-03-23T22:28:49.343000", "content": "

thanks

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "271223": "Great news! I would like to confirm that all images for the challenge were contributed under Creative Commons license 0 (CC0 public domain). Although many other image sets exist out there we carefully collected only CC0 images so that the set would be maximally useful for future work - enjoy!\n\n**Credits for the images** are provided in this thread: https://www.kaggle.com/c/data-science-bowl-2018/discussion/54759\n\n**Where to get the data set**: The data set will live on at the BBBC, Broad Institute Broad Bioimage Benchmark Collection; we plan to release updated versions as well, with corrections: https://data.broadinstitute.org/bbbc/BBBC038/ \n\n**How to cite the data set**: \"We used image set BBBC038v1 [Caicedo et al. https://www.ncbi.nlm.nih.gov/pubmed/31636459], available from the Broad Bioimage Benchmark Collection.\"", "279608": "I have one question about RLE for submission. We have two ways to obtain the result: segmentation (left side) and instance segmentation (right side) in the\n\n![Result][1]\n \nNote that, these images are same only difference is that the left image summarizes all masks into a single image, while the right image distinguishes mask by mask. For example, you often see the left image in result of UNet, while the right image is in Mask-RCNN. So, we will use one of them to encode and submit. Which image should I chose to provide better performance? Thanks\n [1]: https://i.stack.imgur.com/ajYf8.png \"Left image: Segmentation , right image : Instance segmentation\"", "314778": "Hello, Anne, if I didn't upload my model by now(cause I missed the deadlines) ,does this influence my final LB position?", "276222": "Hello，Anne，about the stage2，just release the test data，or test data and train data", "271415": ">> @drannecarpenter Thanks a lot for letting us know about this!", "298752": " I have a question about how to upload our final model into kaggle as stage 1 deadline approach? Thank you", "660581": "The paper describing the challenge is now out! I have edited the pinned comment with how to cite the dataset. ", "314473": "Hello! Sorry for offtop. Anne, could you send me the results of the first stage of the competition Data Science Bowl 2018? I forgot to save my final results for stage 1, and now it's stage 2, and the previous results were deleted. I need this to confirm my LB score for the online course.", "313590": "Just out of curiosity, please advise on the cells the 2 drawings in the Stage 2 test set (**170bc41b2095177cccd3d4c8977c619147580f1d93b4fe9701eddd77736d4ece.png** & **5390acefd575cf9b33413ddf6cbb9ce137ae07dc04616ba24c7b5fe476c827d2.png**) are supposed to represent? Thanks.", "311896": "After the competition is over, can/will the train/test datasets be available at https://www.kaggle.com/datasets? At this point, I have no hope of having a model ready by the deadline, but I still want to build one once I get my tools working and see how it would have scored on stage 2.", "301916": "Awesome work!!!!", "301743": "wow", "276228": "Hi Anne, does that mean we could still continue using external datasets? \n\nThere's been a few challenges over the years some still have active links\n\n 1. http://ludo17.free.fr/mitos_2012/download.html\n 2. https://mitos-atypia-14.grand-challenge.org/donwload/\n\n", "273205": "Thanks Anne! Just a few quick questions:\n\n1. How does a human expert recognize the Nuclei /Entire cell?\nCan you post some links to lectures/info on this topic for people without any background in biology?\nIn your expert opinion, how hard is it for anyone to become an expert when it comes to detecting cell/nuclei?\n\n2.You just said \"many other image sets exist\" - where can we find more relevant image sets, labelled or unlabeled?", "313342": "", "274999": " Thanks Tilii, great explanation.", "302311": "thanks" } }