{
  "id": 252837,
  "title": "prior research on the same subject.",
  "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/252837",
  "author_name": "",
  "post_date": "2021-07-13T23:59:04.852180700Z",
  "votes": 41,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Hi, kagglers!</p>\n<p>I did a quick search and found that there is a paper that came out just recently on the same subject.<br>\nHowever, in this paper, only T2 is used, but in this competition, there are four shooting methods. This may be the key point of this competition.<br>\n<a href=\"http://www.ajnr.org/content/42/5/845\" target=\"_blank\">http://www.ajnr.org/content/42/5/845</a></p>\n<blockquote>\n  <p>Fluid Attenuated Inversion Recovery (FLAIR)<br>\n  T1-weighted pre-contrast (T1w)<br>\n  T1-weighted post-contrast (T1Gd)<br>\n  T2-weighted (T2) &lt;-only used in the above paper.</p>\n</blockquote>",
  "messages": [
    {
      "id": "1387077",
      "postDate": "07/13/2021 23:59:04",
      "content": "<p>Hi, kagglers!</p>\n<p>I did a quick search and found that there is a paper that came out just recently on the same subject.<br>\nHowever, in this paper, only T2 is used, but in this competition, there are four shooting methods. This may be the key point of this competition.<br>\n<a href=\"http://www.ajnr.org/content/42/5/845\" target=\"_blank\">http://www.ajnr.org/content/42/5/845</a></p>\n<blockquote>\n  <p>Fluid Attenuated Inversion Recovery (FLAIR)<br>\n  T1-weighted pre-contrast (T1w)<br>\n  T1-weighted post-contrast (T1Gd)<br>\n  T2-weighted (T2) &lt;-only used in the above paper.</p>\n</blockquote>",
      "rawMarkdown": "Hi, kagglers!\n\nI did a quick search and found that there is a paper that came out just recently on the same subject.\nHowever, in this paper, only T2 is used, but in this competition, there are four shooting methods. This may be the key point of this competition.\nhttp://www.ajnr.org/content/42/5/845\n\n> Fluid Attenuated Inversion Recovery (FLAIR)\n> T1-weighted pre-contrast (T1w)\n> T1-weighted post-contrast (T1Gd)\n> T2-weighted (T2) <-only used in the above paper.",
      "votes": null
    },
    {
      "id": "1387098",
      "postDate": "07/14/2021 00:36:57",
      "content": "<p>Thank you for sharing this article! Great resource!<br>\nEspecially noticed a statement: \"The benefits of using <strong>T2WIs</strong> are that they are routinely acquired, they can be obtained quickly, and high quality T2WI can even be obtained in the setting of motion degradation.\"</p>\n<p>Are you aware of any references/information which discuss characteristic/benefits/downsides of the other three - FLAIR, T1w and T1Gd?</p>",
      "rawMarkdown": "Thank you for sharing this article! Great resource!\nEspecially noticed a statement: \"The benefits of using **T2WIs** are that they are routinely acquired, they can be obtained quickly, and high quality T2WI can even be obtained in the setting of motion degradation.\"\n\nAre you aware of any references/information which discuss characteristic/benefits/downsides of the other three - FLAIR, T1w and T1Gd?",
      "votes": null
    },
    {
      "id": "1387152",
      "postDate": "07/14/2021 02:22:18",
      "content": "<p>I have just made a post on the discussion. <br>\nHowever, the description of T1Gd is currently being written. Please wait for updates.</p>\n<p><a href=\"https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/252843\" target=\"_blank\">https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/252843</a></p>",
      "rawMarkdown": "I have just made a post on the discussion. \nHowever, the description of T1Gd is currently being written. Please wait for updates.\n\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/252843",
      "votes": null
    },
    {
      "id": "1387546",
      "postDate": "07/14/2021 09:04:08",
      "content": "<p>nice projjj</p>",
      "rawMarkdown": "nice projjj",
      "votes": null
    },
    {
      "id": "1387907",
      "postDate": "07/14/2021 14:03:06",
      "content": "<p>Great resource. Thanks for share🔥🔥🔥🔥🔥</p>",
      "rawMarkdown": "Great resource. Thanks for share🔥🔥🔥🔥🔥",
      "votes": null
    },
    {
      "id": "1394558",
      "postDate": "07/20/2021 12:43:19",
      "content": "<p>Thank you for sharing this useful paper 😄<br>\nI have tried to briefly summarize the essence of the paper that may be relevant to this competition.</p>\n<p>I thought the most significant difference between this competition and this one is the labels that are given.<br>\nWhile the labels used in the research in the paper are masks for segmentation, in this competition, there is only one binary label for each patient's cubic data.</p>\n<hr>\n<p><strong>Overview</strong></p>\n<p><img src=\"https://f.easyuploader.app/20210720214124_78485a55.JPG\" alt=\"\"></p>\n<ul>\n<li><p>Segmentation-based approach (needs mask labels!)</p></li>\n<li><p>3-D Unet with DenseNet encoder/decoder (called MGMT-net in the paper)</p></li>\n<li><p>Pre-trained with other datasets (in author’s previous study)<br>\n<a href=\"https://academic.oup.com/neuro-oncology/article/22/3/402/5602247\" target=\"_blank\">https://academic.oup.com/neuro-oncology/article/22/3/402/5602247</a></p></li>\n<li><p>AUC<br>\nFold1: 0.9574<br>\nFold2: 0.8978<br>\nFold3: 0.939</p></li>\n<li><p>pre-processing: </p>\n<ul>\n<li>removing radiofrequency inhomogeneity using N4 Bias Field Correction <a href=\"https://simpleitk.readthedocs.io/en/master/link_N4BiasFieldCorrection_docs.html\" target=\"_blank\">https://simpleitk.readthedocs.io/en/master/link_N4BiasFieldCorrection_docs.html</a></li>\n<li>normalizing intensity to zero-mean and unit variance</li></ul></li>\n<li><p>Data augmentation: </p>\n<ul>\n<li>horizontal and vertical flipping</li>\n<li>random and translational rotation</li>\n<li>salt and pepper noise</li>\n<li>Gaussian noise</li>\n<li>projective transformation</li>\n<li>down-sampling images by 50% and 25% (reducing the voxel resolution to 2 and 4 mm^3)</li></ul></li>\n<li><p>Training:</p>\n<ul>\n<li>Adam: Initial learning rate = 10-5</li>\n<li>batch size: 15</li>\n<li>maximal epochs: 100 for each fold</li>\n<li>75%  overlapping 3D patches (size: 32 × 32 × 32 voxels) as a translation in the x-y-z-plane in training and validation</li>\n<li>During training, only patches with at least 1 tumor voxel were included; thus, the number of patches included per training varied depending on the size of the tumor.</li></ul></li>\n<li><p>Post processing:</p>\n<ul>\n<li>combined 2 outputs (methylated and unmethylated MGMT voxel predictions)</li>\n<li>Majority voting over the voxelwise classes for determining MGMT type (0 or 1)</li></ul></li>\n</ul>",
      "rawMarkdown": "Thank you for sharing this useful paper 😄\nI have tried to briefly summarize the essence of the paper that may be relevant to this competition.\n\nI thought the most significant difference between this competition and this one is the labels that are given.\nWhile the labels used in the research in the paper are masks for segmentation, in this competition, there is only one binary label for each patient's cubic data.\n\n---\n**Overview**\n\n![](https://f.easyuploader.app/20210720214124_78485a55.JPG)\n\n- Segmentation-based approach (needs mask labels!)\n\n- 3-D Unet with DenseNet encoder/decoder (called MGMT-net in the paper)\n  \n- Pre-trained with other datasets (in author’s previous study)\nhttps://academic.oup.com/neuro-oncology/article/22/3/402/5602247\n  \n- AUC\nFold1: 0.9574\nFold2: 0.8978\nFold3: 0.939\n  \n- pre-processing: \n - removing radiofrequency inhomogeneity using N4 Bias Field Correction https://simpleitk.readthedocs.io/en/master/link_N4BiasFieldCorrection_docs.html\n - normalizing intensity to zero-mean and unit variance\n  \n- Data augmentation: \n - horizontal and vertical flipping\n - random and translational rotation\n - salt and pepper noise\n - Gaussian noise\n - projective transformation\n - down-sampling images by 50% and 25% (reducing the voxel resolution to 2 and 4 mm^3)\n  \n- Training:\n - Adam: Initial learning rate = 10-5\n - batch size: 15\n - maximal epochs: 100 for each fold\n - 75%  overlapping 3D patches (size: 32 × 32 × 32 voxels) as a translation in the x-y-z-plane in training and validation\n - During training, only patches with at least 1 tumor voxel were included; thus, the number of patches included per training varied depending on the size of the tumor.\n  \n- Post processing:\n - combined 2 outputs (methylated and unmethylated MGMT voxel predictions)\n - Majority voting over the voxelwise classes for determining MGMT type (0 or 1)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1387098,
      "author_name": "elenaeb",
      "author_url": "",
      "post_date": "07/14/2021 00:36:57",
      "content": "<p>Thank you for sharing this article! Great resource!<br>\nEspecially noticed a statement: \"The benefits of using <strong>T2WIs</strong> are that they are routinely acquired, they can be obtained quickly, and high quality T2WI can even be obtained in the setting of motion degradation.\"</p>\n<p>Are you aware of any references/information which discuss characteristic/benefits/downsides of the other three - FLAIR, T1w and T1Gd?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1387152,
          "author_name": "dlbsabu",
          "author_url": "",
          "post_date": "07/14/2021 02:22:18",
          "content": "<p>I have just made a post on the discussion. <br>\nHowever, the description of T1Gd is currently being written. Please wait for updates.</p>\n<p><a href=\"https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/252843\" target=\"_blank\">https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/252843</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1387546,
      "author_name": "mikeyclark",
      "author_url": "",
      "post_date": "07/14/2021 09:04:08",
      "content": "<p>nice projjj</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1387907,
      "author_name": "sanikamal",
      "author_url": "",
      "post_date": "07/14/2021 14:03:06",
      "content": "<p>Great resource. Thanks for share🔥🔥🔥🔥🔥</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1394558,
      "author_name": "maxwell110",
      "author_url": "",
      "post_date": "07/20/2021 12:43:19",
      "content": "<p>Thank you for sharing this useful paper 😄<br>\nI have tried to briefly summarize the essence of the paper that may be relevant to this competition.</p>\n<p>I thought the most significant difference between this competition and this one is the labels that are given.<br>\nWhile the labels used in the research in the paper are masks for segmentation, in this competition, there is only one binary label for each patient's cubic data.</p>\n<hr>\n<p><strong>Overview</strong></p>\n<p><img src=\"https://f.easyuploader.app/20210720214124_78485a55.JPG\" alt=\"\"></p>\n<ul>\n<li><p>Segmentation-based approach (needs mask labels!)</p></li>\n<li><p>3-D Unet with DenseNet encoder/decoder (called MGMT-net in the paper)</p></li>\n<li><p>Pre-trained with other datasets (in author’s previous study)<br>\n<a href=\"https://academic.oup.com/neuro-oncology/article/22/3/402/5602247\" target=\"_blank\">https://academic.oup.com/neuro-oncology/article/22/3/402/5602247</a></p></li>\n<li><p>AUC<br>\nFold1: 0.9574<br>\nFold2: 0.8978<br>\nFold3: 0.939</p></li>\n<li><p>pre-processing: </p>\n<ul>\n<li>removing radiofrequency inhomogeneity using N4 Bias Field Correction <a href=\"https://simpleitk.readthedocs.io/en/master/link_N4BiasFieldCorrection_docs.html\" target=\"_blank\">https://simpleitk.readthedocs.io/en/master/link_N4BiasFieldCorrection_docs.html</a></li>\n<li>normalizing intensity to zero-mean and unit variance</li></ul></li>\n<li><p>Data augmentation: </p>\n<ul>\n<li>horizontal and vertical flipping</li>\n<li>random and translational rotation</li>\n<li>salt and pepper noise</li>\n<li>Gaussian noise</li>\n<li>projective transformation</li>\n<li>down-sampling images by 50% and 25% (reducing the voxel resolution to 2 and 4 mm^3)</li></ul></li>\n<li><p>Training:</p>\n<ul>\n<li>Adam: Initial learning rate = 10-5</li>\n<li>batch size: 15</li>\n<li>maximal epochs: 100 for each fold</li>\n<li>75%  overlapping 3D patches (size: 32 × 32 × 32 voxels) as a translation in the x-y-z-plane in training and validation</li>\n<li>During training, only patches with at least 1 tumor voxel were included; thus, the number of patches included per training varied depending on the size of the tumor.</li></ul></li>\n<li><p>Post processing:</p>\n<ul>\n<li>combined 2 outputs (methylated and unmethylated MGMT voxel predictions)</li>\n<li>Majority voting over the voxelwise classes for determining MGMT type (0 or 1)</li></ul></li>\n</ul>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1387077": "Hi, kagglers!\n\nI did a quick search and found that there is a paper that came out just recently on the same subject.\nHowever, in this paper, only T2 is used, but in this competition, there are four shooting methods. This may be the key point of this competition.\nhttp://www.ajnr.org/content/42/5/845\n\n> Fluid Attenuated Inversion Recovery (FLAIR)\n> T1-weighted pre-contrast (T1w)\n> T1-weighted post-contrast (T1Gd)\n> T2-weighted (T2) <-only used in the above paper.",
    "1387098": "Thank you for sharing this article! Great resource!\nEspecially noticed a statement: \"The benefits of using **T2WIs** are that they are routinely acquired, they can be obtained quickly, and high quality T2WI can even be obtained in the setting of motion degradation.\"\n\nAre you aware of any references/information which discuss characteristic/benefits/downsides of the other three - FLAIR, T1w and T1Gd?",
    "1387152": "I have just made a post on the discussion. \nHowever, the description of T1Gd is currently being written. Please wait for updates.\n\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/252843",
    "1387546": "nice projjj",
    "1387907": "Great resource. Thanks for share🔥🔥🔥🔥🔥",
    "1394558": "Thank you for sharing this useful paper 😄\nI have tried to briefly summarize the essence of the paper that may be relevant to this competition.\n\nI thought the most significant difference between this competition and this one is the labels that are given.\nWhile the labels used in the research in the paper are masks for segmentation, in this competition, there is only one binary label for each patient's cubic data.\n\n---\n**Overview**\n\n![](https://f.easyuploader.app/20210720214124_78485a55.JPG)\n\n- Segmentation-based approach (needs mask labels!)\n\n- 3-D Unet with DenseNet encoder/decoder (called MGMT-net in the paper)\n  \n- Pre-trained with other datasets (in author’s previous study)\nhttps://academic.oup.com/neuro-oncology/article/22/3/402/5602247\n  \n- AUC\nFold1: 0.9574\nFold2: 0.8978\nFold3: 0.939\n  \n- pre-processing: \n - removing radiofrequency inhomogeneity using N4 Bias Field Correction https://simpleitk.readthedocs.io/en/master/link_N4BiasFieldCorrection_docs.html\n - normalizing intensity to zero-mean and unit variance\n  \n- Data augmentation: \n - horizontal and vertical flipping\n - random and translational rotation\n - salt and pepper noise\n - Gaussian noise\n - projective transformation\n - down-sampling images by 50% and 25% (reducing the voxel resolution to 2 and 4 mm^3)\n  \n- Training:\n - Adam: Initial learning rate = 10-5\n - batch size: 15\n - maximal epochs: 100 for each fold\n - 75%  overlapping 3D patches (size: 32 × 32 × 32 voxels) as a translation in the x-y-z-plane in training and validation\n - During training, only patches with at least 1 tumor voxel were included; thus, the number of patches included per training varied depending on the size of the tumor.\n  \n- Post processing:\n - combined 2 outputs (methylated and unmethylated MGMT voxel predictions)\n - Majority voting over the voxelwise classes for determining MGMT type (0 or 1)"
  },
  "source": "meta"
}