{
  "id": 332714,
  "title": "is there going to be a problem?  HPA data and HuBMAP data ",
  "url": "/competitions/hubmap-organ-segmentation/discussion/332714",
  "author_name": "hengck23",
  "post_date": "2022-06-23T03:25:21.897000",
  "votes": 78,
  "comment_count": 17,
  "views": 0,
  "content": "<p>It is mentioned </p>\n<p>training dataset  = public HPA data, <br>\npublic test set =  private HPA data and HuBMAP data (hidden)<br>\nprivate test set = only HuBMAP data (hidden)</p>\n<hr>\n<p>we do not see the HuBMAP data at all !!!<br>\nhow do we know \"how smiliar are the two dataset\"? This affects how to design our algorithms (e.g. augmentation or domain adaption)</p>\n<p>is there any sample HuBMAP data that we can look at, e.g. from <a href=\"https://portal.hubmapconsortium.org\" target=\"_blank\">https://portal.hubmapconsortium.org</a></p>\n<p>i wonder is it a good way to design the dataset for this competition? </p>\n<hr>\n<p>note:</p>\n<ul>\n<li>(1) for each model, we probably need to submit twice: one for  csv column source == HPA and another for source != HPA.<br>\nso now CV correlates to two LB <br>\n(we also need to probe for the num of HPA and HuBMAP in public test)</li>\n</ul>\n<hr>\n<ul>\n<li>(2) the image scale and colors are different!</li>\n</ul>\n<p><a href=\"https://www.kaggle.com/code/yashvrdnjain/background-information/notebook\" target=\"_blank\">https://www.kaggle.com/code/yashvrdnjain/background-information/notebook</a><br>\nHPA Data : DAB and counterstained with hematoxylin. <br>\nHuBMAP Data  : PAS / H&amp;E stain</p>\n<hr>\n<p>magnification :<br>\npixel_size - The height/width of a single pixel from this image in micrometers. All HPA images have a pixel size of 0.4 µm. For Hubmap imagery the pixel size is 0.5 µm for kidney, 0.2290 µm for large intestine, 0.7562 µm for lung, 0.4945 µm for spleen, and 6.263 µm for prostate.</p>\n<p>tissue_thickness - The thickness of the biopsy sample in micrometers. All HPA images have a thickness of 4 µm. The Hubmap samples have tissue slice thicknesses 10 µm for kidney, 8 µm for large intestine, 4 µm for spleen, 5 µm for lung, and 5 µm for prostate.</p>\n<hr>\n<p>presence of noise ???</p>\n<ul>\n<li>tissue folding, cracks, ????</li>\n</ul>\n<p><img src=\"https://i.ibb.co/chgQ9gM/Selection-058.png\" alt=\"https://i.ibb.co/chgQ9gM/Selection-058.png\"></p>",
  "messages": [
    {
      "id": 1829908,
      "postDate": "2022-06-23T03:25:21.897Z",
      "content": "<p>It is mentioned </p>\n<p>training dataset  = public HPA data, <br>\npublic test set =  private HPA data and HuBMAP data (hidden)<br>\nprivate test set = only HuBMAP data (hidden)</p>\n<hr>\n<p>we do not see the HuBMAP data at all !!!<br>\nhow do we know \"how smiliar are the two dataset\"? This affects how to design our algorithms (e.g. augmentation or domain adaption)</p>\n<p>is there any sample HuBMAP data that we can look at, e.g. from <a href=\"https://portal.hubmapconsortium.org\" target=\"_blank\">https://portal.hubmapconsortium.org</a></p>\n<p>i wonder is it a good way to design the dataset for this competition? </p>\n<hr>\n<p>note:</p>\n<ul>\n<li>(1) for each model, we probably need to submit twice: one for  csv column source == HPA and another for source != HPA.<br>\nso now CV correlates to two LB <br>\n(we also need to probe for the num of HPA and HuBMAP in public test)</li>\n</ul>\n<hr>\n<ul>\n<li>(2) the image scale and colors are different!</li>\n</ul>\n<p><a href=\"https://www.kaggle.com/code/yashvrdnjain/background-information/notebook\" target=\"_blank\">https://www.kaggle.com/code/yashvrdnjain/background-information/notebook</a><br>\nHPA Data : DAB and counterstained with hematoxylin. <br>\nHuBMAP Data  : PAS / H&amp;E stain</p>\n<hr>\n<p>magnification :<br>\npixel_size - The height/width of a single pixel from this image in micrometers. All HPA images have a pixel size of 0.4 µm. For Hubmap imagery the pixel size is 0.5 µm for kidney, 0.2290 µm for large intestine, 0.7562 µm for lung, 0.4945 µm for spleen, and 6.263 µm for prostate.</p>\n<p>tissue_thickness - The thickness of the biopsy sample in micrometers. All HPA images have a thickness of 4 µm. The Hubmap samples have tissue slice thicknesses 10 µm for kidney, 8 µm for large intestine, 4 µm for spleen, 5 µm for lung, and 5 µm for prostate.</p>\n<hr>\n<p>presence of noise ???</p>\n<ul>\n<li>tissue folding, cracks, ????</li>\n</ul>\n<p><img src=\"https://i.ibb.co/chgQ9gM/Selection-058.png\" alt=\"https://i.ibb.co/chgQ9gM/Selection-058.png\"></p>",
      "rawMarkdown": "It is mentioned \n\ntraining dataset  = public HPA data, \npublic test set =  private HPA data and HuBMAP data (hidden)\nprivate test set = only HuBMAP data (hidden)\n\n---\n\nwe do not see the HuBMAP data at all !!!\nhow do we know \"how smiliar are the two dataset\"? This affects how to design our algorithms (e.g. augmentation or domain adaption)\n\nis there any sample HuBMAP data that we can look at, e.g. from https://portal.hubmapconsortium.org\n\ni wonder is it a good way to design the dataset for this competition? \n\n---\n\nnote:\n- (1) for each model, we probably need to submit twice: one for  csv column source == HPA and another for source != HPA.\n  so now CV correlates to two LB \n (we also need to probe for the num of HPA and HuBMAP in public test)\n\n\n---\n\n- (2) the image scale and colors are different!\n\nhttps://www.kaggle.com/code/yashvrdnjain/background-information/notebook\nHPA Data : DAB and counterstained with hematoxylin. \nHuBMAP Data  : PAS / H&E stain\n\n---\n\nmagnification :\npixel_size - The height/width of a single pixel from this image in micrometers. All HPA images have a pixel size of 0.4 µm. For Hubmap imagery the pixel size is 0.5 µm for kidney, 0.2290 µm for large intestine, 0.7562 µm for lung, 0.4945 µm for spleen, and 6.263 µm for prostate.\n\ntissue_thickness - The thickness of the biopsy sample in micrometers. All HPA images have a thickness of 4 µm. The Hubmap samples have tissue slice thicknesses 10 µm for kidney, 8 µm for large intestine, 4 µm for spleen, 5 µm for lung, and 5 µm for prostate.\n\n\n---\n\n\npresence of noise ???\n- tissue folding, cracks, ????\n\n\n![https://i.ibb.co/chgQ9gM/Selection-058.png](https://i.ibb.co/chgQ9gM/Selection-058.png)",
      "votes": 74
    },
    {
      "id": 1830919,
      "postDate": "2022-06-23T19:29:19.950Z",
      "content": "<p>Hello <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a> ,<br>\nGeneralization is an important goal in the context of building a reference Atlas (which is the main goal of HuBMAP). The data comes from various different sources (different researchers, different labs, different equipment etc), using different methodologies, and all has to be integrated. Additionally, HuBMAP focusses on many organs, not just a single organ. Hence, generalizability of ML algorithms is a crucial goal.<br>\nThe hurdles you mention such as resolution differences, color differences, artifacts, different staining etc, are all real issues that HubMAP faces. </p>",
      "rawMarkdown": "Hello @hengck23 ,\nGeneralization is an important goal in the context of building a reference Atlas (which is the main goal of HuBMAP). The data comes from various different sources (different researchers, different labs, different equipment etc), using different methodologies, and all has to be integrated. Additionally, HuBMAP focusses on many organs, not just a single organ. Hence, generalizability of ML algorithms is a crucial goal.\nThe hurdles you mention such as resolution differences, color differences, artifacts, different staining etc, are all real issues that HubMAP faces. \n",
      "votes": 3,
      "replies": [
        {
          "id": 1831013,
          "postDate": "2022-06-23T21:04:37.393Z",
          "content": "<p><a href=\"https://www.kaggle.com/yashvrdnjain\" target=\"_blank\">@yashvrdnjain</a> <br>\nthanks for the reply. I am thinking that the NLP pretrain/finetune transformer approach may be a good solution:</p>\n<ul>\n<li>pretrain with unsupervised learning (e.g. masked patch prediction) using many, many various stains, scales, etc</li>\n<li>finetune with few-shot supervised learning (e.g. specific lab, organ … )</li>\n</ul>\n<p>e.g.<br>\nBEiT: BERT Pre-Training of Image Transformers<br>\n<a href=\"https://openreview.net/pdf?id=p-BhZSz59o4\" target=\"_blank\">https://openreview.net/pdf?id=p-BhZSz59o4</a></p>",
          "rawMarkdown": "@yashvrdnjain \nthanks for the reply. I am thinking that the NLP pretrain/finetune transformer approach may be a good solution:\n- pretrain with unsupervised learning (e.g. masked patch prediction) using many, many various stains, scales, etc\n- finetune with few-shot supervised learning (e.g. specific lab, organ ... )\n\ne.g.\nBEiT: BERT Pre-Training of Image Transformers\nhttps://openreview.net/pdf?id=p-BhZSz59o4",
          "votes": 1
        },
        {
          "id": 1832933,
          "postDate": "2022-06-25T15:06:08.280Z",
          "content": "<p>Could you also comment on the quality of the test labels?</p>\n<p><a href=\"https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/332958\" target=\"_blank\">https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/332958</a></p>\n<p>The training labels are very imprecise and noisey. If the test labels are also like this then there's no point in trying to get the model to produce tighter/more precise segmentations. If not, there's little incentive to improve the model. In fact, learning human labelling error/noise might be the optimal strategy in such regime.</p>",
          "rawMarkdown": "Could you also comment on the quality of the test labels?\n\nhttps://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/332958\n\nThe training labels are very imprecise and noisey. If the test labels are also like this then there's no point in trying to get the model to produce tighter/more precise segmentations. If not, there's little incentive to improve the model. In fact, learning human labelling error/noise might be the optimal strategy in such regime.",
          "votes": 2
        }
      ]
    },
    {
      "id": 1830355,
      "postDate": "2022-06-23T11:45:10.517Z",
      "content": "<p>This document describes the process used by subject matter experts (SMEs) for downloading HuBMAP image data, generating manual segmentations of structures such as functional tissueunits (FTUs), and uploading the segmentations to a storage location. The audience for this SOP is the group of subject matter experts (SMEs) who are tasked with segmenting FTUs in image data,typically a histologist recommended by the PI associated with the organ of interest. The current tool preferred for this process is QuPath</p>\n<p><a href=\"https://zenodo.org/record/6629522/preview/SOP_%20Manual%20Segmentation%20of%20Tissue_v1.2.0.pdf\" target=\"_blank\">https://zenodo.org/record/6629522/preview/SOP_%20Manual%20Segmentation%20of%20Tissue_v1.2.0.pdf</a><br>\n<a href=\"https://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/229186\" target=\"_blank\">https://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/229186</a></p>",
      "rawMarkdown": "This document describes the process used by subject matter experts (SMEs) for downloading HuBMAP image data, generating manual segmentations of structures such as functional tissueunits (FTUs), and uploading the segmentations to a storage location. The audience for this SOP is the group of subject matter experts (SMEs) who are tasked with segmenting FTUs in image data,typically a histologist recommended by the PI associated with the organ of interest. The current tool preferred for this process is QuPath\n\nhttps://zenodo.org/record/6629522/preview/SOP_%20Manual%20Segmentation%20of%20Tissue_v1.2.0.pdf\nhttps://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/229186",
      "votes": 4
    },
    {
      "id": 1829989,
      "postDate": "2022-06-23T04:56:58.713Z",
      "content": "<p>maybe this is a hidden test image :)<br>\n<a href=\"https://storage.googleapis.com/kaggle-competitions/kaggle/34547/logos/header.png?t=2022-02-15-22-37-27\" target=\"_blank\">https://storage.googleapis.com/kaggle-competitions/kaggle/34547/logos/header.png?t=2022-02-15-22-37-27</a></p>",
      "rawMarkdown": "maybe this is a hidden test image :)\nhttps://storage.googleapis.com/kaggle-competitions/kaggle/34547/logos/header.png?t=2022-02-15-22-37-27",
      "votes": 4,
      "replies": [
        {
          "id": 1830211,
          "postDate": "2022-06-23T09:08:56.763Z",
          "content": "<p>FTUs in this competition are:<br>\n(a) glomeruli in kidney  : </p>\n<ul>\n<li>refer to last year competition for  HuBMAP data</li>\n<li>refer to HuBMAP website<br>\n<a href=\"https://portal.hubmapconsortium.org/browse/collection/4964d24bbc6668a72c4cbb5e0393a6bc\" target=\"_blank\">https://portal.hubmapconsortium.org/browse/collection/4964d24bbc6668a72c4cbb5e0393a6bc</a></li>\n</ul>\n<p><a href=\"https://ibb.co/RvfSgQ6\"><img src=\"https://i.ibb.co/M2rhgP1/Selection-011.png\" alt=\"Selection-011\"></a><br>\n<a href=\"https://ibb.co/TK1P8mx\"><img src=\"https://i.ibb.co/YDXfkpF/Selection-010.png\" alt=\"Selection-010\"></a> </p>\n<p>(b) crypt in the large intestine : </p>\n<ul>\n<li>refer to here?  crypt in colon <br>\n(i check the meta data, these are from standford TMD. they are actually tag as \"large intestine\" in Hubmap database)</li>\n</ul>\n<p><a href=\"https://cns-iu.github.io/ccf-research-kaggle-2021/\" target=\"_blank\">https://cns-iu.github.io/ccf-research-kaggle-2021/</a><br>\n<a href=\"https://github.com/cns-iu/ccf-research-ftu\" target=\"_blank\">https://github.com/cns-iu/ccf-research-ftu</a></p>\n<p><img src=\"https://i.ibb.co/TYxYN1X/Selection-009.png\" alt=\"https://i.ibb.co/TYxYN1X/Selection-009.png\"></p>\n<p>(c) alveolus in the lung, <br>\n(d) glandular acinus in the prostate, <br>\n(e) white pulp in the spleen.</p>\n<ul>\n<li>we have one spleen test slide in the download page</li>\n</ul>\n<p>actually HPA also have H&amp;E stain</p>\n<p><img src=\"https://i.ibb.co/dcbbYvv/Selection-036.png\" alt=\"https://i.ibb.co/dcbbYvv/Selection-036.png\"></p>",
          "rawMarkdown": "FTUs in this competition are:\n(a) glomeruli in kidney  : \n- refer to last year competition for  HuBMAP data\n- refer to HuBMAP website\nhttps://portal.hubmapconsortium.org/browse/collection/4964d24bbc6668a72c4cbb5e0393a6bc\n\n<a href=\"https://ibb.co/RvfSgQ6\"><img src=\"https://i.ibb.co/M2rhgP1/Selection-011.png\" alt=\"Selection-011\" border=\"0\"></a>\n<a href=\"https://ibb.co/TK1P8mx\"><img src=\"https://i.ibb.co/YDXfkpF/Selection-010.png\" alt=\"Selection-010\" border=\"0\"></a> \n\n(b) crypt in the large intestine : \n- refer to here?  crypt in colon \n(i check the meta data, these are from standford TMD. they are actually tag as \"large intestine\" in Hubmap database)\n\nhttps://cns-iu.github.io/ccf-research-kaggle-2021/\nhttps://github.com/cns-iu/ccf-research-ftu\n\n\n![https://i.ibb.co/TYxYN1X/Selection-009.png](https://i.ibb.co/TYxYN1X/Selection-009.png)\n\n(c) alveolus in the lung, \n(d) glandular acinus in the prostate, \n(e) white pulp in the spleen.\n- we have one spleen test slide in the download page\n\nactually HPA also have H&E stain\n\n![https://i.ibb.co/dcbbYvv/Selection-036.png](https://i.ibb.co/dcbbYvv/Selection-036.png)",
          "votes": 3
        },
        {
          "id": 1835626,
          "postDate": "2022-06-28T00:01:01.467Z",
          "content": "<p>another site for external data (can be used for e.g. unlabel self-supervised, etc)<br>\n<a href=\"https://gtexportal.org\" target=\"_blank\">https://gtexportal.org</a><br>\n<img src=\"https://i.ibb.co/gVbLZ79/Selection-055.png\" alt=\"https://i.ibb.co/gVbLZ79/Selection-055.png\"></p>",
          "rawMarkdown": "another site for external data (can be used for e.g. unlabel self-supervised, etc)\nhttps://gtexportal.org\n![https://i.ibb.co/gVbLZ79/Selection-055.png](https://i.ibb.co/gVbLZ79/Selection-055.png)",
          "votes": 1
        }
      ]
    },
    {
      "id": 1912598,
      "postDate": "2022-08-24T21:06:02.083Z",
      "content": "<p>some 3d data. in theory you can virtual stain them in any color (google for papers)</p>\n<p>paper: <br>\nImaging intact human organs with local resolution of cellular structures using hierarchical phase-contrast tomography<br>\n<a href=\"https://www.nature.com/articles/s41592-021-01317-x\" target=\"_blank\">https://www.nature.com/articles/s41592-021-01317-x</a></p>\n<p>data:  <br>\n<a href=\"https://human-organ-atlas.esrf.eu/\" target=\"_blank\">https://human-organ-atlas.esrf.eu/</a><br>\n<a href=\"https://www.youtube.com/watch?v=WTFshuoyUDQ\" target=\"_blank\">https://www.youtube.com/watch?v=WTFshuoyUDQ</a></p>\n<p><img src=\"https://media.springernature.com/lw685/springer-static/image/art%3A10.1038%2Fs41592-021-01317-x/MediaObjects/41592_2021_1317_Fig3_HTML.png\" alt=\"https://media.springernature.com/lw685/springer-static/image/art%3A10.1038%2Fs41592-021-01317-x/MediaObjects/41592_2021_1317_Fig3_HTML.png\"></p>\n<p>e.g. spleen<br>\n<a href=\"https://human-organ-atlas.esrf.eu/datasets/572244401\" target=\"_blank\">https://human-organ-atlas.esrf.eu/datasets/572244401</a><br>\n(the video helps you to visualize how 2d slice is made from the 3d organ)<br>\n<a href=\"https://icatplus.esrf.fr/resource/abbd4d8d-9b72-44be-9716-646037ad64ac/file/download?resourceId=6172d6633d89500012beeac6\" target=\"_blank\">https://icatplus.esrf.fr/resource/abbd4d8d-9b72-44be-9716-646037ad64ac/file/download?resourceId=6172d6633d89500012beeac6</a></p>",
      "rawMarkdown": "some 3d data. in theory you can virtual stain them in any color (google for papers)\n\npaper: \nImaging intact human organs with local resolution of cellular structures using hierarchical phase-contrast tomography\nhttps://www.nature.com/articles/s41592-021-01317-x\n\ndata:  \nhttps://human-organ-atlas.esrf.eu/\nhttps://www.youtube.com/watch?v=WTFshuoyUDQ\n\n![https://media.springernature.com/lw685/springer-static/image/art%3A10.1038%2Fs41592-021-01317-x/MediaObjects/41592_2021_1317_Fig3_HTML.png](https://media.springernature.com/lw685/springer-static/image/art%3A10.1038%2Fs41592-021-01317-x/MediaObjects/41592_2021_1317_Fig3_HTML.png)\n\ne.g. spleen\nhttps://human-organ-atlas.esrf.eu/datasets/572244401\n(the video helps you to visualize how 2d slice is made from the 3d organ)\nhttps://icatplus.esrf.fr/resource/abbd4d8d-9b72-44be-9716-646037ad64ac/file/download?resourceId=6172d6633d89500012beeac6",
      "votes": 1
    },
    {
      "id": 1874634,
      "postDate": "2022-07-28T12:00:20.650Z",
      "content": "<p>one of the best images for lung:<br>\n<a href=\"https://histology.medicine.umich.edu/resources/respiratory-system\" target=\"_blank\">https://histology.medicine.umich.edu/resources/respiratory-system</a><br>\nscroll to:</p>\n<p>III. Respiratory Portion of the Lung<br>\nSlide 129_20x (lung, H&amp;E) View Virtual Slide<br>\ne.g. <a href=\"https://histologyslides.med.umich.edu/Histology/Respiratory%20System/129_HISTO_20X.htm\" target=\"_blank\">https://histologyslides.med.umich.edu/Histology/Respiratory%20System/129_HISTO_20X.htm</a></p>\n<p><a href=\"https://ibb.co/NryLZDh\"><img src=\"https://i.ibb.co/tzJb8jf/Selection-098.png\" alt=\"Selection-098\"></a> </p>\n<p>Slide 130_20x (lung, H&amp;E) View Virtual Slide</p>\n<p>Slide 130-1_40x (lung, H&amp;E) View Virtual Slide</p>\n<p>Slide 130-2_40x (lung, H&amp;E) View Virtual Slide</p>\n<p>Slide 132_20x (lung, H&amp;E) View Virtual Slide</p>\n<p>Slide 132_40x (lung, H&amp;E) View Virtual Slide</p>\n<p>also:<br>\n<a href=\"https://undergraduate.vetmed.wsu.edu/courses/vph-308/histology/lab-9-respiratory-cardiovascular/lab-9-respiratory\" target=\"_blank\">https://undergraduate.vetmed.wsu.edu/courses/vph-308/histology/lab-9-respiratory-cardiovascular/lab-9-respiratory</a></p>",
      "rawMarkdown": "one of the best images for lung:\nhttps://histology.medicine.umich.edu/resources/respiratory-system\nscroll to:\n\nIII. Respiratory Portion of the Lung\nSlide 129_20x (lung, H&E) View Virtual Slide\ne.g. https://histologyslides.med.umich.edu/Histology/Respiratory%20System/129_HISTO_20X.htm\n\n<a href=\"https://ibb.co/NryLZDh\"><img src=\"https://i.ibb.co/tzJb8jf/Selection-098.png\" alt=\"Selection-098\" border=\"0\"></a> \n\nSlide 130_20x (lung, H&E) View Virtual Slide\n\nSlide 130-1_40x (lung, H&E) View Virtual Slide\n\nSlide 130-2_40x (lung, H&E) View Virtual Slide\n\nSlide 132_20x (lung, H&E) View Virtual Slide\n\nSlide 132_40x (lung, H&E) View Virtual Slide\n\n\nalso:\nhttps://undergraduate.vetmed.wsu.edu/courses/vph-308/histology/lab-9-respiratory-cardiovascular/lab-9-respiratory\n",
      "votes": 2,
      "replies": [
        {
          "id": 1874716,
          "postDate": "2022-07-28T12:41:47.363Z",
          "content": "<p>Thanks for sharing</p>",
          "rawMarkdown": "Thanks for sharing"
        }
      ]
    },
    {
      "id": 1890383,
      "postDate": "2022-08-08T17:47:50.553Z",
      "content": "<p>hey <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a> , could you please explain how's this to be done? thanks</p>\n<p>(1) for each model, we probably need to submit twice: one for csv column source == HPA and another for source != HPA.<br>\nso now CV correlates to two LB<br>\n(we also need to probe for the num of HPA and HuBMAP in public test)</p>",
      "rawMarkdown": "hey @hengck23 , could you please explain how's this to be done? thanks\n\n(1) for each model, we probably need to submit twice: one for csv column source == HPA and another for source != HPA.\nso now CV correlates to two LB\n(we also need to probe for the num of HPA and HuBMAP in public test)"
    },
    {
      "id": 1880446,
      "postDate": "2022-08-01T18:11:49.107Z",
      "content": "<p>Including HuBMAP 2021's data to actual training data and training with the mixed data might cause (or would cause?) unbalanced dataset issues. But since we are predicting the same thing, maybe it wouldn't be a problem(?).</p>",
      "rawMarkdown": "Including HuBMAP 2021's data to actual training data and training with the mixed data might cause (or would cause?) unbalanced dataset issues. But since we are predicting the same thing, maybe it wouldn't be a problem(?)."
    },
    {
      "id": 1870619,
      "postDate": "2022-07-25T17:12:51.677Z",
      "content": "<p>Hi hengck23, thank you for all your informative posts.</p>\n<p>For the public leaderboard, do you think people should be validating on HPA at all? If the the private test set is Hubmap only, then the public test set shouldn't have HPA in it. I'm thinking of only checking Hubmap score on the public test set. Thank you.</p>",
      "rawMarkdown": "Hi hengck23, thank you for all your informative posts.\n\nFor the public leaderboard, do you think people should be validating on HPA at all? If the the private test set is Hubmap only, then the public test set shouldn't have HPA in it. I'm thinking of only checking Hubmap score on the public test set. Thank you.",
      "replies": [
        {
          "id": 1870632,
          "postDate": "2022-07-25T17:21:32.270Z",
          "content": "<p>correlation between score of HPA and Hubmap could be higher than you think.</p>\n<p>at least for kidney FTU, we have last year data. You can verify if your model improve Hubmap when HPA score gets better.</p>\n<p>both HPA and Hubmap have a common stain component (the blue H stain). HPA is blue-brown, while Hubmap is blue-pink.</p>\n<p>it goes back to is the discrminative information is in shape, color or texture? </p>",
          "rawMarkdown": "correlation between score of HPA and Hubmap could be higher than you think.\n\nat least for kidney FTU, we have last year data. You can verify if your model improve Hubmap when HPA score gets better.\n\nboth HPA and Hubmap have a common stain component (the blue H stain). HPA is blue-brown, while Hubmap is blue-pink.\n\nit goes back to is the discrminative information is in shape, color or texture? ",
          "votes": 1
        }
      ]
    },
    {
      "id": 1840485,
      "postDate": "2022-07-02T09:30:04.670Z",
      "content": "<p>I was confused about using two datasets together, but I'm getting help from your good discussion.</p>",
      "rawMarkdown": "I was confused about using two datasets together, but I'm getting help from your good discussion."
    },
    {
      "id": 1835732,
      "postDate": "2022-06-28T03:31:31.887Z",
      "content": "<p>glomeruli detection model<br>\n<a href=\"https://www.orbit.bio/deep-learning-object-segmentation/\" target=\"_blank\">https://www.orbit.bio/deep-learning-object-segmentation/</a></p>",
      "rawMarkdown": " glomeruli detection model\nhttps://www.orbit.bio/deep-learning-object-segmentation/"
    },
    {
      "id": 1847652,
      "postDate": "2022-07-08T03:36:09.390Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 1830919,
      "author_name": "Yashvardhan Jain",
      "author_url": "",
      "post_date": "2022-06-23T19:29:19.950000",
      "content": "<p>Hello <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a> ,<br>\nGeneralization is an important goal in the context of building a reference Atlas (which is the main goal of HuBMAP). The data comes from various different sources (different researchers, different labs, different equipment etc), using different methodologies, and all has to be integrated. Additionally, HuBMAP focusses on many organs, not just a single organ. Hence, generalizability of ML algorithms is a crucial goal.<br>\nThe hurdles you mention such as resolution differences, color differences, artifacts, different staining etc, are all real issues that HubMAP faces. </p>",
      "votes": 3,
      "replies": [
        {
          "id": 1831013,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2022-06-23T21:04:37.393000",
          "content": "<p><a href=\"https://www.kaggle.com/yashvrdnjain\" target=\"_blank\">@yashvrdnjain</a> <br>\nthanks for the reply. I am thinking that the NLP pretrain/finetune transformer approach may be a good solution:</p>\n<ul>\n<li>pretrain with unsupervised learning (e.g. masked patch prediction) using many, many various stains, scales, etc</li>\n<li>finetune with few-shot supervised learning (e.g. specific lab, organ … )</li>\n</ul>\n<p>e.g.<br>\nBEiT: BERT Pre-Training of Image Transformers<br>\n<a href=\"https://openreview.net/pdf?id=p-BhZSz59o4\" target=\"_blank\">https://openreview.net/pdf?id=p-BhZSz59o4</a></p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1832933,
          "author_name": "Jack Shi",
          "author_url": "",
          "post_date": "2022-06-25T15:06:08.280000",
          "content": "<p>Could you also comment on the quality of the test labels?</p>\n<p><a href=\"https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/332958\" target=\"_blank\">https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/332958</a></p>\n<p>The training labels are very imprecise and noisey. If the test labels are also like this then there's no point in trying to get the model to produce tighter/more precise segmentations. If not, there's little incentive to improve the model. In fact, learning human labelling error/noise might be the optimal strategy in such regime.</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 1830355,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2022-06-23T11:45:10.517000",
      "content": "<p>This document describes the process used by subject matter experts (SMEs) for downloading HuBMAP image data, generating manual segmentations of structures such as functional tissueunits (FTUs), and uploading the segmentations to a storage location. The audience for this SOP is the group of subject matter experts (SMEs) who are tasked with segmenting FTUs in image data,typically a histologist recommended by the PI associated with the organ of interest. The current tool preferred for this process is QuPath</p>\n<p><a href=\"https://zenodo.org/record/6629522/preview/SOP_%20Manual%20Segmentation%20of%20Tissue_v1.2.0.pdf\" target=\"_blank\">https://zenodo.org/record/6629522/preview/SOP_%20Manual%20Segmentation%20of%20Tissue_v1.2.0.pdf</a><br>\n<a href=\"https://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/229186\" target=\"_blank\">https://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/229186</a></p>",
      "votes": 4,
      "replies": []
    },
    {
      "id": 1829989,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2022-06-23T04:56:58.713000",
      "content": "<p>maybe this is a hidden test image :)<br>\n<a href=\"https://storage.googleapis.com/kaggle-competitions/kaggle/34547/logos/header.png?t=2022-02-15-22-37-27\" target=\"_blank\">https://storage.googleapis.com/kaggle-competitions/kaggle/34547/logos/header.png?t=2022-02-15-22-37-27</a></p>",
      "votes": 4,
      "replies": [
        {
          "id": 1830211,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2022-06-23T09:08:56.763000",
          "content": "<p>FTUs in this competition are:<br>\n(a) glomeruli in kidney  : </p>\n<ul>\n<li>refer to last year competition for  HuBMAP data</li>\n<li>refer to HuBMAP website<br>\n<a href=\"https://portal.hubmapconsortium.org/browse/collection/4964d24bbc6668a72c4cbb5e0393a6bc\" target=\"_blank\">https://portal.hubmapconsortium.org/browse/collection/4964d24bbc6668a72c4cbb5e0393a6bc</a></li>\n</ul>\n<p><a href=\"https://ibb.co/RvfSgQ6\"><img src=\"https://i.ibb.co/M2rhgP1/Selection-011.png\" alt=\"Selection-011\"></a><br>\n<a href=\"https://ibb.co/TK1P8mx\"><img src=\"https://i.ibb.co/YDXfkpF/Selection-010.png\" alt=\"Selection-010\"></a> </p>\n<p>(b) crypt in the large intestine : </p>\n<ul>\n<li>refer to here?  crypt in colon <br>\n(i check the meta data, these are from standford TMD. they are actually tag as \"large intestine\" in Hubmap database)</li>\n</ul>\n<p><a href=\"https://cns-iu.github.io/ccf-research-kaggle-2021/\" target=\"_blank\">https://cns-iu.github.io/ccf-research-kaggle-2021/</a><br>\n<a href=\"https://github.com/cns-iu/ccf-research-ftu\" target=\"_blank\">https://github.com/cns-iu/ccf-research-ftu</a></p>\n<p><img src=\"https://i.ibb.co/TYxYN1X/Selection-009.png\" alt=\"https://i.ibb.co/TYxYN1X/Selection-009.png\"></p>\n<p>(c) alveolus in the lung, <br>\n(d) glandular acinus in the prostate, <br>\n(e) white pulp in the spleen.</p>\n<ul>\n<li>we have one spleen test slide in the download page</li>\n</ul>\n<p>actually HPA also have H&amp;E stain</p>\n<p><img src=\"https://i.ibb.co/dcbbYvv/Selection-036.png\" alt=\"https://i.ibb.co/dcbbYvv/Selection-036.png\"></p>",
          "votes": 3,
          "replies": []
        },
        {
          "id": 1835626,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2022-06-28T00:01:01.467000",
          "content": "<p>another site for external data (can be used for e.g. unlabel self-supervised, etc)<br>\n<a href=\"https://gtexportal.org\" target=\"_blank\">https://gtexportal.org</a><br>\n<img src=\"https://i.ibb.co/gVbLZ79/Selection-055.png\" alt=\"https://i.ibb.co/gVbLZ79/Selection-055.png\"></p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1912598,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2022-08-24T21:06:02.083000",
      "content": "<p>some 3d data. in theory you can virtual stain them in any color (google for papers)</p>\n<p>paper: <br>\nImaging intact human organs with local resolution of cellular structures using hierarchical phase-contrast tomography<br>\n<a href=\"https://www.nature.com/articles/s41592-021-01317-x\" target=\"_blank\">https://www.nature.com/articles/s41592-021-01317-x</a></p>\n<p>data:  <br>\n<a href=\"https://human-organ-atlas.esrf.eu/\" target=\"_blank\">https://human-organ-atlas.esrf.eu/</a><br>\n<a href=\"https://www.youtube.com/watch?v=WTFshuoyUDQ\" target=\"_blank\">https://www.youtube.com/watch?v=WTFshuoyUDQ</a></p>\n<p><img src=\"https://media.springernature.com/lw685/springer-static/image/art%3A10.1038%2Fs41592-021-01317-x/MediaObjects/41592_2021_1317_Fig3_HTML.png\" alt=\"https://media.springernature.com/lw685/springer-static/image/art%3A10.1038%2Fs41592-021-01317-x/MediaObjects/41592_2021_1317_Fig3_HTML.png\"></p>\n<p>e.g. spleen<br>\n<a href=\"https://human-organ-atlas.esrf.eu/datasets/572244401\" target=\"_blank\">https://human-organ-atlas.esrf.eu/datasets/572244401</a><br>\n(the video helps you to visualize how 2d slice is made from the 3d organ)<br>\n<a href=\"https://icatplus.esrf.fr/resource/abbd4d8d-9b72-44be-9716-646037ad64ac/file/download?resourceId=6172d6633d89500012beeac6\" target=\"_blank\">https://icatplus.esrf.fr/resource/abbd4d8d-9b72-44be-9716-646037ad64ac/file/download?resourceId=6172d6633d89500012beeac6</a></p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1874634,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2022-07-28T12:00:20.650000",
      "content": "<p>one of the best images for lung:<br>\n<a href=\"https://histology.medicine.umich.edu/resources/respiratory-system\" target=\"_blank\">https://histology.medicine.umich.edu/resources/respiratory-system</a><br>\nscroll to:</p>\n<p>III. Respiratory Portion of the Lung<br>\nSlide 129_20x (lung, H&amp;E) View Virtual Slide<br>\ne.g. <a href=\"https://histologyslides.med.umich.edu/Histology/Respiratory%20System/129_HISTO_20X.htm\" target=\"_blank\">https://histologyslides.med.umich.edu/Histology/Respiratory%20System/129_HISTO_20X.htm</a></p>\n<p><a href=\"https://ibb.co/NryLZDh\"><img src=\"https://i.ibb.co/tzJb8jf/Selection-098.png\" alt=\"Selection-098\"></a> </p>\n<p>Slide 130_20x (lung, H&amp;E) View Virtual Slide</p>\n<p>Slide 130-1_40x (lung, H&amp;E) View Virtual Slide</p>\n<p>Slide 130-2_40x (lung, H&amp;E) View Virtual Slide</p>\n<p>Slide 132_20x (lung, H&amp;E) View Virtual Slide</p>\n<p>Slide 132_40x (lung, H&amp;E) View Virtual Slide</p>\n<p>also:<br>\n<a href=\"https://undergraduate.vetmed.wsu.edu/courses/vph-308/histology/lab-9-respiratory-cardiovascular/lab-9-respiratory\" target=\"_blank\">https://undergraduate.vetmed.wsu.edu/courses/vph-308/histology/lab-9-respiratory-cardiovascular/lab-9-respiratory</a></p>",
      "votes": 2,
      "replies": [
        {
          "id": 1874716,
          "author_name": "Robson",
          "author_url": "",
          "post_date": "2022-07-28T12:41:47.363000",
          "content": "<p>Thanks for sharing</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1890383,
      "author_name": "JINO ROHIT",
      "author_url": "",
      "post_date": "2022-08-08T17:47:50.553000",
      "content": "<p>hey <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a> , could you please explain how's this to be done? thanks</p>\n<p>(1) for each model, we probably need to submit twice: one for csv column source == HPA and another for source != HPA.<br>\nso now CV correlates to two LB<br>\n(we also need to probe for the num of HPA and HuBMAP in public test)</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1880446,
      "author_name": "Furkan K",
      "author_url": "",
      "post_date": "2022-08-01T18:11:49.107000",
      "content": "<p>Including HuBMAP 2021's data to actual training data and training with the mixed data might cause (or would cause?) unbalanced dataset issues. But since we are predicting the same thing, maybe it wouldn't be a problem(?).</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1870619,
      "author_name": "datadote",
      "author_url": "",
      "post_date": "2022-07-25T17:12:51.677000",
      "content": "<p>Hi hengck23, thank you for all your informative posts.</p>\n<p>For the public leaderboard, do you think people should be validating on HPA at all? If the the private test set is Hubmap only, then the public test set shouldn't have HPA in it. I'm thinking of only checking Hubmap score on the public test set. Thank you.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1870632,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2022-07-25T17:21:32.270000",
          "content": "<p>correlation between score of HPA and Hubmap could be higher than you think.</p>\n<p>at least for kidney FTU, we have last year data. You can verify if your model improve Hubmap when HPA score gets better.</p>\n<p>both HPA and Hubmap have a common stain component (the blue H stain). HPA is blue-brown, while Hubmap is blue-pink.</p>\n<p>it goes back to is the discrminative information is in shape, color or texture? </p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1840485,
      "author_name": "feelgom",
      "author_url": "",
      "post_date": "2022-07-02T09:30:04.670000",
      "content": "<p>I was confused about using two datasets together, but I'm getting help from your good discussion.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1835732,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2022-06-28T03:31:31.887000",
      "content": "<p>glomeruli detection model<br>\n<a href=\"https://www.orbit.bio/deep-learning-object-segmentation/\" target=\"_blank\">https://www.orbit.bio/deep-learning-object-segmentation/</a></p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1847652,
      "author_name": "",
      "author_url": "",
      "post_date": "2022-07-08T03:36:09.390000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1829908": "It is mentioned \n\ntraining dataset  = public HPA data, \npublic test set =  private HPA data and HuBMAP data (hidden)\nprivate test set = only HuBMAP data (hidden)\n\n---\n\nwe do not see the HuBMAP data at all !!!\nhow do we know \"how smiliar are the two dataset\"? This affects how to design our algorithms (e.g. augmentation or domain adaption)\n\nis there any sample HuBMAP data that we can look at, e.g. from https://portal.hubmapconsortium.org\n\ni wonder is it a good way to design the dataset for this competition? \n\n---\n\nnote:\n- (1) for each model, we probably need to submit twice: one for  csv column source == HPA and another for source != HPA.\n  so now CV correlates to two LB \n (we also need to probe for the num of HPA and HuBMAP in public test)\n\n\n---\n\n- (2) the image scale and colors are different!\n\nhttps://www.kaggle.com/code/yashvrdnjain/background-information/notebook\nHPA Data : DAB and counterstained with hematoxylin. \nHuBMAP Data  : PAS / H&E stain\n\n---\n\nmagnification :\npixel_size - The height/width of a single pixel from this image in micrometers. All HPA images have a pixel size of 0.4 µm. For Hubmap imagery the pixel size is 0.5 µm for kidney, 0.2290 µm for large intestine, 0.7562 µm for lung, 0.4945 µm for spleen, and 6.263 µm for prostate.\n\ntissue_thickness - The thickness of the biopsy sample in micrometers. All HPA images have a thickness of 4 µm. The Hubmap samples have tissue slice thicknesses 10 µm for kidney, 8 µm for large intestine, 4 µm for spleen, 5 µm for lung, and 5 µm for prostate.\n\n\n---\n\n\npresence of noise ???\n- tissue folding, cracks, ????\n\n\n![https://i.ibb.co/chgQ9gM/Selection-058.png](https://i.ibb.co/chgQ9gM/Selection-058.png)",
    "1830919": "Hello @hengck23 ,\nGeneralization is an important goal in the context of building a reference Atlas (which is the main goal of HuBMAP). The data comes from various different sources (different researchers, different labs, different equipment etc), using different methodologies, and all has to be integrated. Additionally, HuBMAP focusses on many organs, not just a single organ. Hence, generalizability of ML algorithms is a crucial goal.\nThe hurdles you mention such as resolution differences, color differences, artifacts, different staining etc, are all real issues that HubMAP faces. \n",
    "1830355": "This document describes the process used by subject matter experts (SMEs) for downloading HuBMAP image data, generating manual segmentations of structures such as functional tissueunits (FTUs), and uploading the segmentations to a storage location. The audience for this SOP is the group of subject matter experts (SMEs) who are tasked with segmenting FTUs in image data,typically a histologist recommended by the PI associated with the organ of interest. The current tool preferred for this process is QuPath\n\nhttps://zenodo.org/record/6629522/preview/SOP_%20Manual%20Segmentation%20of%20Tissue_v1.2.0.pdf\nhttps://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/229186",
    "1829989": "maybe this is a hidden test image :)\nhttps://storage.googleapis.com/kaggle-competitions/kaggle/34547/logos/header.png?t=2022-02-15-22-37-27",
    "1912598": "some 3d data. in theory you can virtual stain them in any color (google for papers)\n\npaper: \nImaging intact human organs with local resolution of cellular structures using hierarchical phase-contrast tomography\nhttps://www.nature.com/articles/s41592-021-01317-x\n\ndata:  \nhttps://human-organ-atlas.esrf.eu/\nhttps://www.youtube.com/watch?v=WTFshuoyUDQ\n\n![https://media.springernature.com/lw685/springer-static/image/art%3A10.1038%2Fs41592-021-01317-x/MediaObjects/41592_2021_1317_Fig3_HTML.png](https://media.springernature.com/lw685/springer-static/image/art%3A10.1038%2Fs41592-021-01317-x/MediaObjects/41592_2021_1317_Fig3_HTML.png)\n\ne.g. spleen\nhttps://human-organ-atlas.esrf.eu/datasets/572244401\n(the video helps you to visualize how 2d slice is made from the 3d organ)\nhttps://icatplus.esrf.fr/resource/abbd4d8d-9b72-44be-9716-646037ad64ac/file/download?resourceId=6172d6633d89500012beeac6",
    "1874634": "one of the best images for lung:\nhttps://histology.medicine.umich.edu/resources/respiratory-system\nscroll to:\n\nIII. Respiratory Portion of the Lung\nSlide 129_20x (lung, H&E) View Virtual Slide\ne.g. https://histologyslides.med.umich.edu/Histology/Respiratory%20System/129_HISTO_20X.htm\n\n<a href=\"https://ibb.co/NryLZDh\"><img src=\"https://i.ibb.co/tzJb8jf/Selection-098.png\" alt=\"Selection-098\" border=\"0\"></a> \n\nSlide 130_20x (lung, H&E) View Virtual Slide\n\nSlide 130-1_40x (lung, H&E) View Virtual Slide\n\nSlide 130-2_40x (lung, H&E) View Virtual Slide\n\nSlide 132_20x (lung, H&E) View Virtual Slide\n\nSlide 132_40x (lung, H&E) View Virtual Slide\n\n\nalso:\nhttps://undergraduate.vetmed.wsu.edu/courses/vph-308/histology/lab-9-respiratory-cardiovascular/lab-9-respiratory\n",
    "1890383": "hey @hengck23 , could you please explain how's this to be done? thanks\n\n(1) for each model, we probably need to submit twice: one for csv column source == HPA and another for source != HPA.\nso now CV correlates to two LB\n(we also need to probe for the num of HPA and HuBMAP in public test)",
    "1880446": "Including HuBMAP 2021's data to actual training data and training with the mixed data might cause (or would cause?) unbalanced dataset issues. But since we are predicting the same thing, maybe it wouldn't be a problem(?).",
    "1870619": "Hi hengck23, thank you for all your informative posts.\n\nFor the public leaderboard, do you think people should be validating on HPA at all? If the the private test set is Hubmap only, then the public test set shouldn't have HPA in it. I'm thinking of only checking Hubmap score on the public test set. Thank you.",
    "1840485": "I was confused about using two datasets together, but I'm getting help from your good discussion.",
    "1835732": " glomeruli detection model\nhttps://www.orbit.bio/deep-learning-object-segmentation/",
    "1847652": ""
  }
}