So looking through the data and doing a quick k-means clustering on the color channels reveals there are (at least) 3 different types of images in the dataset, which come from different types of microscopes. Why does it make sense to have them all in the same competition? It seems like any decent approach would have to divide these data and build a model for each since clearly grayscale images should be handled differently than color ones. Am I missing something here?
", "messages": [ { "id": "269744", "postDate": "01/17/2018 07:51:41", "content": "So looking through the data and doing a quick k-means clustering on the color channels reveals there are (at least) 3 different types of images in the dataset, which come from different types of microscopes. Why does it make sense to have them all in the same competition? It seems like any decent approach would have to divide these data and build a model for each since clearly grayscale images should be handled differently than color ones. Am I missing something here?
", "rawMarkdown": "So looking through the data and doing a quick k-means clustering on the color channels reveals there are (at least) 3 different types of images in the dataset, which come from different types of microscopes. Why does it make sense to have them all in the same competition? It seems like any decent approach would have to divide these data and build a model for each since clearly grayscale images should be handled differently than color ones. Am I missing something here?", "votes": null }, { "id": "269753", "postDate": "01/17/2018 08:06:43", "content": "Looks like this is on purpose. From the data page:
\n\n\n\n\nThe images were acquired under a variety of conditions and vary in the cell type, magnification, and imaging modality (brightfield vs. fluorescence). The dataset is designed to challenge an algorithm's ability to generalize across these variations.
\n
Expect the second stage to have even different modalities.
", "rawMarkdown": "Looks like this is on purpose. From the data page:\n\n> The images were acquired under a variety of conditions and vary in the cell type, magnification, and imaging modality (brightfield vs. fluorescence). The dataset is designed to challenge an algorithm's ability to generalize across these variations.\n\nExpect the second stage to have even different modalities.", "votes": null }, { "id": "269769", "postDate": "01/17/2018 09:03:59", "content": "Thanks, I had missed that line. It still seems odd to explicitly leave that information out since it is almost always known at experimental time
", "rawMarkdown": "Thanks, I had missed that line. It still seems odd to explicitly leave that information out since it is almost always known at experimental time", "votes": null }, { "id": "269879", "postDate": "01/17/2018 12:57:43", "content": "I was wondering the same and even asked that question on twitter. I am doing similar image analysis and one of the first things we did was to standardize the image acquisition. In my opinion, this will significantly improve the accuracy of the prediction and simplify the segmentation process. For all three types I would guess a standard segmentation algorithm would beat a NN, if the image acquisition is optimized (e.g. slide scanner).
\n\nI am very curious about the winning NN, how it will perform in the feature when image acquisition might change again...My experience with NN, if you change anything on the Hardware site, a NN gets worse while robust hand crafted scale and color invariant features keeps performing very well...
", "rawMarkdown": "I was wondering the same and even asked that question on twitter. I am doing similar image analysis and one of the first things we did was to standardize the image acquisition. In my opinion, this will significantly improve the accuracy of the prediction and simplify the segmentation process. For all three types I would guess a standard segmentation algorithm would beat a NN, if the image acquisition is optimized (e.g. slide scanner).\n\nI am very curious about the winning NN, how it will perform in the feature when image acquisition might change again...My experience with NN, if you change anything on the Hardware site, a NN gets worse while robust hand crafted scale and color invariant features keeps performing very well...", "votes": null }, { "id": "269946", "postDate": "01/17/2018 14:58:00", "content": "It’s a good question. But we’re interested in models that recognize or exploit features about nuclei rather than some particular experimental setup. Regardless, if the general problem is solved without metadata features, why would you bother with such features in the future?
", "rawMarkdown": "It’s a good question. But we’re interested in models that recognize or exploit features about nuclei rather than some particular experimental setup. Regardless, if the general problem is solved without metadata features, why would you bother with such features in the future?", "votes": null }, { "id": "269950", "postDate": "01/17/2018 15:04:47", "content": "The ideal model would be able to take any image as an input and return the segmentation of the nuclei within that image. The model would be a tool capable of being utilized by people with a diverse set of backgrounds and skills including those with little to no background in biology, image analysis, or machine learning. Such a model would be widely accessible to students and casual scientists, yet powerful enough to be deployed in demanding, high-throughput research environments.
\n\nA solution that does or does not include a step where images are sorted would be equally valid, because the end user does not need to be aware of the mechanics of the model.
", "rawMarkdown": "The ideal model would be able to take any image as an input and return the segmentation of the nuclei within that image. The model would be a tool capable of being utilized by people with a diverse set of backgrounds and skills including those with little to no background in biology, image analysis, or machine learning. Such a model would be widely accessible to students and casual scientists, yet powerful enough to be deployed in demanding, high-throughput research environments.\n\nA solution that does or does not include a step where images are sorted would be equally valid, because the end user does not need to be aware of the mechanics of the model.", "votes": null }, { "id": "269965", "postDate": "01/17/2018 15:31:29", "content": "I found that very simple U-Net has an ability of generalization to the unseen environment. So probably it could not be a problem. The way of the experiment is below.
\n\nFirst, I split train images into three clusters using k-means. Then I set one of the clusters as test data and trained with other two. Finally, I evaluated with the unseen test data (one of the clusters) and mean_iou were as good as other two clusters.
", "rawMarkdown": "I found that very simple U-Net has an ability of generalization to the unseen environment. So probably it could not be a problem. The way of the experiment is below.\n\nFirst, I split train images into three clusters using k-means. Then I set one of the clusters as test data and trained with other two. Finally, I evaluated with the unseen test data (one of the clusters) and mean_iou were as good as other two clusters.", "votes": null }, { "id": "270059", "postDate": "01/17/2018 17:49:08", "content": "\n\n\nAllen Goodman wrote
\n \n\n\nIt’s a good question. But we’re interested in models that recognize or exploit features about nuclei rather than some particular experimental setup. Regardless, if the general problem is solved without metadata features, why would you bother with such features in the future?
\n
It would be very interesting to have a list of the different experimental setups that we could find on the 2 stage. That way we could use data augmentation to model all the cases.
", "rawMarkdown": "> **Allen Goodman wrote**\n> \n> > It’s a good question. But we’re interested in models that recognize or exploit features about nuclei rather than some particular experimental setup. Regardless, if the general problem is solved without metadata features, why would you bother with such features in the future? \n> \n\nIt would be very interesting to have a list of the different experimental setups that we could find on the 2 stage. That way we could use data augmentation to model all the cases.", "votes": null }, { "id": "270079", "postDate": "01/17/2018 18:21:18", "content": "Yeah. I’d guess that augmentation will be super crucial. I’m excited to find out. 😎
\n\nIt’d be amazing if the competition inspired some new augmentation methods for robustly solving microscopy problems. I find microscopy really exciting from a computer vision perspective since it has a strictness that enables precise modeling of experimental setups inside augmentation methods (rather than relying on the extremely generalized methods used for natural images like channel shifts).
", "rawMarkdown": "Yeah. I’d guess that augmentation will be super crucial. I’m excited to find out. 😎\n\nIt’d be amazing if the competition inspired some new augmentation methods for robustly solving microscopy problems. I find microscopy really exciting from a computer vision perspective since it has a strictness that enables precise modeling of experimental setups inside augmentation methods (rather than relying on the extremely generalized methods used for natural images like channel shifts).", "votes": null }, { "id": "270096", "postDate": "01/17/2018 18:43:32", "content": "In @Kevin Mader's image annotation, group A are histological slides, group B are fluorescent images, and group C are bright-field images. Groups A and C can be acquired by a light microscope or some variation of it. Group B are taken by a fluorescent microscope.
", "rawMarkdown": "In @Kevin Mader's image annotation, group A are histological slides, group B are fluorescent images, and group C are bright-field images. Groups A and C can be acquired by a light microscope or some variation of it. Group B are taken by a fluorescent microscope.", "votes": null }, { "id": "270514", "postDate": "01/18/2018 11:36:27", "content": "In addition to the light & the fluorescent microscopes, are there any other microscopes that are commonly used ?
", "rawMarkdown": "In addition to the light & the fluorescent microscopes, are there any other microscopes that are commonly used ?", "votes": null }, { "id": "271302", "postDate": "01/20/2018 04:49:39", "content": "In the data page, it said \"The second stage dataset will contain images from unseen experimental conditions. \"\nDoes that mean the second stage dataset will have different types of images?
", "rawMarkdown": "In the data page, it said \"The second stage dataset will contain images from unseen experimental conditions. \"\nDoes that mean the second stage dataset will have different types of images?", "votes": null }, { "id": "271557", "postDate": "01/20/2018 19:17:47", "content": "Indeed. I'm an organizer. The second stage (test) dataset will indeed have image sets that are experimentally distinct from anything seen in the training and validation (stage 1) datasets. The goal of this competition is generalizability and robustness so the only way to test this well is to be sure that the final test image sets are distinct relative to the training data.
", "rawMarkdown": "Indeed. I'm an organizer. The second stage (test) dataset will indeed have image sets that are experimentally distinct from anything seen in the training and validation (stage 1) datasets. The goal of this competition is generalizability and robustness so the only way to test this well is to be sure that the final test image sets are distinct relative to the training data.", "votes": null }, { "id": "271558", "postDate": "01/20/2018 19:20:01", "content": "I'm an organizer. The answer here is that we are greedy!
\n\nWe want a single model that just works across all kinds of image modalities, no matter the size of the nuclei or the color scheme. Such a model could be built into software that biologists use with all kinds of microscopes and eliminate the need for them to train on their individual data or provide metadata about their cell type, microscope, resolution, etc.
", "rawMarkdown": "I'm an organizer. The answer here is that we are greedy! \n\nWe want a single model that just works across all kinds of image modalities, no matter the size of the nuclei or the color scheme. Such a model could be built into software that biologists use with all kinds of microscopes and eliminate the need for them to train on their individual data or provide metadata about their cell type, microscope, resolution, etc.", "votes": null }, { "id": "271843", "postDate": "01/21/2018 16:07:23", "content": "there's also electron microscope
", "rawMarkdown": "there's also electron microscope", "votes": null }, { "id": "271907", "postDate": "01/21/2018 20:57:05", "content": "That is true! But no electron microscopy in this challenge (or I should say, none of the images looked to be electron microscopy to MY eye, when annotating!)
", "rawMarkdown": "That is true! But no electron microscopy in this challenge (or I should say, none of the images looked to be electron microscopy to MY eye, when annotating!)", "votes": null }, { "id": "271965", "postDate": "01/22/2018 01:16:00", "content": "Has anyone done any research on different possible modalities? There's violet, fluorescent, and brightfield already in the train set. I don't see any other modalities in the stage 1 test set. I don't know if this is possible, but could one possibly be able to modify one modality to look like another. I think it might be possible. If you could do this you could possibly create more training data. If someone finds more modalities, could you modify images in the train set to look like modalities that could show up on the stage 2 test set? If it is possible, it could definitely make models much more robust.
", "rawMarkdown": "Has anyone done any research on different possible modalities? There's violet, fluorescent, and brightfield already in the train set. I don't see any other modalities in the stage 1 test set. I don't know if this is possible, but could one possibly be able to modify one modality to look like another. I think it might be possible. If you could do this you could possibly create more training data. If someone finds more modalities, could you modify images in the train set to look like modalities that could show up on the stage 2 test set? If it is possible, it could definitely make models much more robust.", "votes": null }, { "id": "277884", "postDate": "02/04/2018 18:07:50", "content": "Is it allowed to classify the modality of the test image first and then using trained models (for the detected modality) for segmentation? I mean to say that we want to train three different models for three different image modalities and in test phase, firstly, image modality will be detected and then segmented through the appropriate model?
", "rawMarkdown": "Is it allowed to classify the modality of the test image first and then using trained models (for the detected modality) for segmentation? I mean to say that we want to train three different models for three different image modalities and in test phase, firstly, image modality will be detected and then segmented through the appropriate model?", "votes": null }, { "id": "277944", "postDate": "02/04/2018 21:39:32", "content": "Yes. You can implement a modality detector if you want, and you can annotate the training set by hand if you need to. It is only forbidden to annotate the test set manually, it is allowed to annotate the training set.
\n\nThat said, it might not be advisable to do this here, since the second stage can possibly introduce new, unseen modalities.
", "rawMarkdown": "Yes. You can implement a modality detector if you want, and you can annotate the training set by hand if you need to. It is only forbidden to annotate the test set manually, it is allowed to annotate the training set.\n\nThat said, it might not be advisable to do this here, since the second stage can possibly introduce new, unseen modalities.", "votes": null }, { "id": "283972", "postDate": "02/16/2018 10:29:54", "content": "After simple KMean clusterization I found that we not only have 3 different modalities in train set, but also have 3 different modalities in test set. And they are DIFFERENT! Only one type of images from train is present in test, other images are completely different
", "rawMarkdown": "After simple KMean clusterization I found that we not only have 3 different modalities in train set, but also have 3 different modalities in test set. And they are DIFFERENT! Only one type of images from train is present in test, other images are completely different", "votes": null }, { "id": "284019", "postDate": "02/16/2018 12:56:52", "content": "Fourth test type in 4f ?
", "rawMarkdown": "Fourth test type in 4f ?", "votes": null }, { "id": "284043", "postDate": "02/16/2018 14:00:48", "content": "It is not so different to be another type imho
", "rawMarkdown": "It is not so different to be another type imho", "votes": null }, { "id": "284089", "postDate": "02/16/2018 15:40:58", "content": "Could you give the id of a train image that is most similar? I find 0287, but I feel it is not very similar at all.
", "rawMarkdown": "Could you give the id of a train image that is most similar? I find 0287, but I feel it is not very similar at all.", "votes": null }, { "id": "284152", "postDate": "02/16/2018 17:11:07", "content": "I don't understand what are you asking for :c
", "rawMarkdown": "I don't understand what are you asking for :c", "votes": null }, { "id": "298591", "postDate": "03/19/2018 19:26:00", "content": "Anne I know that you said we may see variations in stain... I am particularly worried about unstained cells. Did you annotate any of those?
", "rawMarkdown": "Anne I know that you said we may see variations in stain... I am particularly worried about unstained cells. Did you annotate any of those?", "votes": null }, { "id": "299173", "postDate": "03/20/2018 16:23:51", "content": "I literally do not remember if there are any in the competition!
", "rawMarkdown": "I literally do not remember if there are any in the competition!", "votes": null }, { "id": "778346", "postDate": "03/18/2020 11:08:06", "content": "That's great.
", "rawMarkdown": "That's great.", "votes": null } ], "comments": [ { "id": 269753, "author_name": "tobycheese", "author_url": "", "post_date": "01/17/2018 08:06:43", "content": "Looks like this is on purpose. From the data page:
\n\n\n\n\nThe images were acquired under a variety of conditions and vary in the cell type, magnification, and imaging modality (brightfield vs. fluorescence). The dataset is designed to challenge an algorithm's ability to generalize across these variations.
\n
Expect the second stage to have even different modalities.
", "votes": null, "replies": [ { "id": 269769, "author_name": "kmader", "author_url": "", "post_date": "01/17/2018 09:03:59", "content": "Thanks, I had missed that line. It still seems odd to explicitly leave that information out since it is almost always known at experimental time
", "votes": null, "replies": [] }, { "id": 271965, "author_name": "arpandhatt", "author_url": "", "post_date": "01/22/2018 01:16:00", "content": "Has anyone done any research on different possible modalities? There's violet, fluorescent, and brightfield already in the train set. I don't see any other modalities in the stage 1 test set. I don't know if this is possible, but could one possibly be able to modify one modality to look like another. I think it might be possible. If you could do this you could possibly create more training data. If someone finds more modalities, could you modify images in the train set to look like modalities that could show up on the stage 2 test set? If it is possible, it could definitely make models much more robust.
", "votes": null, "replies": [] } ] }, { "id": 269879, "author_name": "snowformatics", "author_url": "", "post_date": "01/17/2018 12:57:43", "content": "I was wondering the same and even asked that question on twitter. I am doing similar image analysis and one of the first things we did was to standardize the image acquisition. In my opinion, this will significantly improve the accuracy of the prediction and simplify the segmentation process. For all three types I would guess a standard segmentation algorithm would beat a NN, if the image acquisition is optimized (e.g. slide scanner).
\n\nI am very curious about the winning NN, how it will perform in the feature when image acquisition might change again...My experience with NN, if you change anything on the Hardware site, a NN gets worse while robust hand crafted scale and color invariant features keeps performing very well...
", "votes": null, "replies": [] }, { "id": 269946, "author_name": "agoodman", "author_url": "", "post_date": "01/17/2018 14:58:00", "content": "It’s a good question. But we’re interested in models that recognize or exploit features about nuclei rather than some particular experimental setup. Regardless, if the general problem is solved without metadata features, why would you bother with such features in the future?
", "votes": null, "replies": [ { "id": 270059, "author_name": "ironbar", "author_url": "", "post_date": "01/17/2018 17:49:08", "content": "\n\n\nAllen Goodman wrote
\n \n\n\nIt’s a good question. But we’re interested in models that recognize or exploit features about nuclei rather than some particular experimental setup. Regardless, if the general problem is solved without metadata features, why would you bother with such features in the future?
\n
It would be very interesting to have a list of the different experimental setups that we could find on the 2 stage. That way we could use data augmentation to model all the cases.
", "votes": null, "replies": [] }, { "id": 270079, "author_name": "agoodman", "author_url": "", "post_date": "01/17/2018 18:21:18", "content": "Yeah. I’d guess that augmentation will be super crucial. I’m excited to find out. 😎
\n\nIt’d be amazing if the competition inspired some new augmentation methods for robustly solving microscopy problems. I find microscopy really exciting from a computer vision perspective since it has a strictness that enables precise modeling of experimental setups inside augmentation methods (rather than relying on the extremely generalized methods used for natural images like channel shifts).
", "votes": null, "replies": [] } ] }, { "id": 269950, "author_name": "bitfrosting", "author_url": "", "post_date": "01/17/2018 15:04:47", "content": "The ideal model would be able to take any image as an input and return the segmentation of the nuclei within that image. The model would be a tool capable of being utilized by people with a diverse set of backgrounds and skills including those with little to no background in biology, image analysis, or machine learning. Such a model would be widely accessible to students and casual scientists, yet powerful enough to be deployed in demanding, high-throughput research environments.
\n\nA solution that does or does not include a step where images are sorted would be equally valid, because the end user does not need to be aware of the mechanics of the model.
", "votes": null, "replies": [] }, { "id": 269965, "author_name": "lyakaap", "author_url": "", "post_date": "01/17/2018 15:31:29", "content": "I found that very simple U-Net has an ability of generalization to the unseen environment. So probably it could not be a problem. The way of the experiment is below.
\n\nFirst, I split train images into three clusters using k-means. Then I set one of the clusters as test data and trained with other two. Finally, I evaluated with the unseen test data (one of the clusters) and mean_iou were as good as other two clusters.
", "votes": null, "replies": [] }, { "id": 270096, "author_name": "tilii7", "author_url": "", "post_date": "01/17/2018 18:43:32", "content": "In @Kevin Mader's image annotation, group A are histological slides, group B are fluorescent images, and group C are bright-field images. Groups A and C can be acquired by a light microscope or some variation of it. Group B are taken by a fluorescent microscope.
", "votes": null, "replies": [ { "id": 270514, "author_name": "ravivadapalli", "author_url": "", "post_date": "01/18/2018 11:36:27", "content": "In addition to the light & the fluorescent microscopes, are there any other microscopes that are commonly used ?
", "votes": null, "replies": [] }, { "id": 271843, "author_name": "weidelamancha", "author_url": "", "post_date": "01/21/2018 16:07:23", "content": "there's also electron microscope
", "votes": null, "replies": [] }, { "id": 271907, "author_name": "drannecarpenter", "author_url": "", "post_date": "01/21/2018 20:57:05", "content": "That is true! But no electron microscopy in this challenge (or I should say, none of the images looked to be electron microscopy to MY eye, when annotating!)
", "votes": null, "replies": [] }, { "id": 298591, "author_name": "bleuknight", "author_url": "", "post_date": "03/19/2018 19:26:00", "content": "Anne I know that you said we may see variations in stain... I am particularly worried about unstained cells. Did you annotate any of those?
", "votes": null, "replies": [] }, { "id": 299173, "author_name": "drannecarpenter", "author_url": "", "post_date": "03/20/2018 16:23:51", "content": "I literally do not remember if there are any in the competition!
", "votes": null, "replies": [] } ] }, { "id": 271302, "author_name": "luna316", "author_url": "", "post_date": "01/20/2018 04:49:39", "content": "In the data page, it said \"The second stage dataset will contain images from unseen experimental conditions. \"\nDoes that mean the second stage dataset will have different types of images?
", "votes": null, "replies": [ { "id": 271557, "author_name": "drannecarpenter", "author_url": "", "post_date": "01/20/2018 19:17:47", "content": "Indeed. I'm an organizer. The second stage (test) dataset will indeed have image sets that are experimentally distinct from anything seen in the training and validation (stage 1) datasets. The goal of this competition is generalizability and robustness so the only way to test this well is to be sure that the final test image sets are distinct relative to the training data.
", "votes": null, "replies": [] } ] }, { "id": 271558, "author_name": "drannecarpenter", "author_url": "", "post_date": "01/20/2018 19:20:01", "content": "I'm an organizer. The answer here is that we are greedy!
\n\nWe want a single model that just works across all kinds of image modalities, no matter the size of the nuclei or the color scheme. Such a model could be built into software that biologists use with all kinds of microscopes and eliminate the need for them to train on their individual data or provide metadata about their cell type, microscope, resolution, etc.
", "votes": null, "replies": [ { "id": 277884, "author_name": "anabik", "author_url": "", "post_date": "02/04/2018 18:07:50", "content": "Is it allowed to classify the modality of the test image first and then using trained models (for the detected modality) for segmentation? I mean to say that we want to train three different models for three different image modalities and in test phase, firstly, image modality will be detected and then segmented through the appropriate model?
", "votes": null, "replies": [] }, { "id": 277944, "author_name": "tobycheese", "author_url": "", "post_date": "02/04/2018 21:39:32", "content": "Yes. You can implement a modality detector if you want, and you can annotate the training set by hand if you need to. It is only forbidden to annotate the test set manually, it is allowed to annotate the training set.
\n\nThat said, it might not be advisable to do this here, since the second stage can possibly introduce new, unseen modalities.
", "votes": null, "replies": [] } ] }, { "id": 283972, "author_name": "bonlime", "author_url": "", "post_date": "02/16/2018 10:29:54", "content": "After simple KMean clusterization I found that we not only have 3 different modalities in train set, but also have 3 different modalities in test set. And they are DIFFERENT! Only one type of images from train is present in test, other images are completely different
", "votes": null, "replies": [ { "id": 284019, "author_name": "glimmung", "author_url": "", "post_date": "02/16/2018 12:56:52", "content": "Fourth test type in 4f ?
", "votes": null, "replies": [] }, { "id": 284043, "author_name": "bonlime", "author_url": "", "post_date": "02/16/2018 14:00:48", "content": "It is not so different to be another type imho
", "votes": null, "replies": [] }, { "id": 284089, "author_name": "glimmung", "author_url": "", "post_date": "02/16/2018 15:40:58", "content": "Could you give the id of a train image that is most similar? I find 0287, but I feel it is not very similar at all.
", "votes": null, "replies": [] }, { "id": 284152, "author_name": "bonlime", "author_url": "", "post_date": "02/16/2018 17:11:07", "content": "I don't understand what are you asking for :c
", "votes": null, "replies": [] } ] }, { "id": 778346, "author_name": "zeynepsevvaldvenci", "author_url": "", "post_date": "03/18/2020 11:08:06", "content": "That's great.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "269744": "So looking through the data and doing a quick k-means clustering on the color channels reveals there are (at least) 3 different types of images in the dataset, which come from different types of microscopes. Why does it make sense to have them all in the same competition? It seems like any decent approach would have to divide these data and build a model for each since clearly grayscale images should be handled differently than color ones. Am I missing something here?", "269753": "Looks like this is on purpose. From the data page:\n\n> The images were acquired under a variety of conditions and vary in the cell type, magnification, and imaging modality (brightfield vs. fluorescence). The dataset is designed to challenge an algorithm's ability to generalize across these variations.\n\nExpect the second stage to have even different modalities.", "269769": "Thanks, I had missed that line. It still seems odd to explicitly leave that information out since it is almost always known at experimental time", "269879": "I was wondering the same and even asked that question on twitter. I am doing similar image analysis and one of the first things we did was to standardize the image acquisition. In my opinion, this will significantly improve the accuracy of the prediction and simplify the segmentation process. For all three types I would guess a standard segmentation algorithm would beat a NN, if the image acquisition is optimized (e.g. slide scanner).\n\nI am very curious about the winning NN, how it will perform in the feature when image acquisition might change again...My experience with NN, if you change anything on the Hardware site, a NN gets worse while robust hand crafted scale and color invariant features keeps performing very well...", "269946": "It’s a good question. But we’re interested in models that recognize or exploit features about nuclei rather than some particular experimental setup. Regardless, if the general problem is solved without metadata features, why would you bother with such features in the future?", "269950": "The ideal model would be able to take any image as an input and return the segmentation of the nuclei within that image. The model would be a tool capable of being utilized by people with a diverse set of backgrounds and skills including those with little to no background in biology, image analysis, or machine learning. Such a model would be widely accessible to students and casual scientists, yet powerful enough to be deployed in demanding, high-throughput research environments.\n\nA solution that does or does not include a step where images are sorted would be equally valid, because the end user does not need to be aware of the mechanics of the model.", "269965": "I found that very simple U-Net has an ability of generalization to the unseen environment. So probably it could not be a problem. The way of the experiment is below.\n\nFirst, I split train images into three clusters using k-means. Then I set one of the clusters as test data and trained with other two. Finally, I evaluated with the unseen test data (one of the clusters) and mean_iou were as good as other two clusters.", "270059": "> **Allen Goodman wrote**\n> \n> > It’s a good question. But we’re interested in models that recognize or exploit features about nuclei rather than some particular experimental setup. Regardless, if the general problem is solved without metadata features, why would you bother with such features in the future? \n> \n\nIt would be very interesting to have a list of the different experimental setups that we could find on the 2 stage. That way we could use data augmentation to model all the cases.", "270079": "Yeah. I’d guess that augmentation will be super crucial. I’m excited to find out. 😎\n\nIt’d be amazing if the competition inspired some new augmentation methods for robustly solving microscopy problems. I find microscopy really exciting from a computer vision perspective since it has a strictness that enables precise modeling of experimental setups inside augmentation methods (rather than relying on the extremely generalized methods used for natural images like channel shifts).", "270096": "In @Kevin Mader's image annotation, group A are histological slides, group B are fluorescent images, and group C are bright-field images. Groups A and C can be acquired by a light microscope or some variation of it. Group B are taken by a fluorescent microscope.", "270514": "In addition to the light & the fluorescent microscopes, are there any other microscopes that are commonly used ?", "271302": "In the data page, it said \"The second stage dataset will contain images from unseen experimental conditions. \"\nDoes that mean the second stage dataset will have different types of images?", "271557": "Indeed. I'm an organizer. The second stage (test) dataset will indeed have image sets that are experimentally distinct from anything seen in the training and validation (stage 1) datasets. The goal of this competition is generalizability and robustness so the only way to test this well is to be sure that the final test image sets are distinct relative to the training data.", "271558": "I'm an organizer. The answer here is that we are greedy! \n\nWe want a single model that just works across all kinds of image modalities, no matter the size of the nuclei or the color scheme. Such a model could be built into software that biologists use with all kinds of microscopes and eliminate the need for them to train on their individual data or provide metadata about their cell type, microscope, resolution, etc.", "271843": "there's also electron microscope", "271907": "That is true! But no electron microscopy in this challenge (or I should say, none of the images looked to be electron microscopy to MY eye, when annotating!)", "271965": "Has anyone done any research on different possible modalities? There's violet, fluorescent, and brightfield already in the train set. I don't see any other modalities in the stage 1 test set. I don't know if this is possible, but could one possibly be able to modify one modality to look like another. I think it might be possible. If you could do this you could possibly create more training data. If someone finds more modalities, could you modify images in the train set to look like modalities that could show up on the stage 2 test set? If it is possible, it could definitely make models much more robust.", "277884": "Is it allowed to classify the modality of the test image first and then using trained models (for the detected modality) for segmentation? I mean to say that we want to train three different models for three different image modalities and in test phase, firstly, image modality will be detected and then segmented through the appropriate model?", "277944": "Yes. You can implement a modality detector if you want, and you can annotate the training set by hand if you need to. It is only forbidden to annotate the test set manually, it is allowed to annotate the training set.\n\nThat said, it might not be advisable to do this here, since the second stage can possibly introduce new, unseen modalities.", "283972": "After simple KMean clusterization I found that we not only have 3 different modalities in train set, but also have 3 different modalities in test set. And they are DIFFERENT! Only one type of images from train is present in test, other images are completely different", "284019": "Fourth test type in 4f ?", "284043": "It is not so different to be another type imho", "284089": "Could you give the id of a train image that is most similar? I find 0287, but I feel it is not very similar at all.", "284152": "I don't understand what are you asking for :c", "298591": "Anne I know that you said we may see variations in stain... I am particularly worried about unstained cells. Did you annotate any of those?", "299173": "I literally do not remember if there are any in the competition!", "778346": "That's great." }, "source": "meta" }