{
  "id": 253075,
  "title": "Baseline🧠 - LB: 0.602",
  "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/253075",
  "author_name": "Yaroslav Isaienkov",
  "post_date": "2021-07-14T21:57:44.330000",
  "votes": 27,
  "comment_count": 13,
  "views": 0,
  "content": "<p>I experimented a bit with the data for this competition. I tried to use the data \"as is\" - and push it just into the EffNetB0.</p>\n<p>I took 8 slices (with equal steps) from each of the first three views of the structural multi-parametric MRI (mpMRI) scans:</p>\n<ul>\n<li>Fluid Attenuated Inversion Recovery (FLAIR)</li>\n<li>T1-weighted pre-contrast (T1w)</li>\n<li>T1-weighted post-contrast (T1Gd)</li>\n</ul>\n<p>Then I take a mean of slices of each type of scan and get 3 \"channels\": <br>\n<img src=\"https://i.imgur.com/GhT4iks.png\" alt=\"\"></p>\n<p>Then merge them into one \"image\" and push them to the EfficientNetB0 model.</p>\n<p>I trained it 2 epochs and after the second I have the next metrics:</p>\n<pre><code>[Epoch Train: 2] loss: 0.68856, score: 0.56197, time: 50 s\n[Epoch Valid: 2] loss: 0.70479, score: 0.52137, time: 12 s\n</code></pre>\n<p>As we can see my loss is equal to ~0.69. It means that the model is not good and the LB score is overfitted.</p>\n<p>From this experiment, I think this approach doesn't have the chance of success. Further improvement of the algorithm should be carried out in two directions:</p>\n<ul>\n<li>To use models that can accept 3D scans as input and are specially pre-trained for medical tasks.</li>\n<li>Cleaning and working with data, as the scans have different tilt angles and parameters.</li>\n</ul>\n<p>EDA and starter modeling you can find in my notebook:</p>\n<ul>\n<li><a href=\"https://www.kaggle.com/ihelon/brain-tumor-eda-with-animations-and-modeling\" target=\"_blank\">EDA &amp; train &amp; inference</a></li>\n</ul>",
  "messages": [
    {
      "id": 1388374,
      "postDate": "2021-07-14T21:57:44.330Z",
      "content": "<p>I experimented a bit with the data for this competition. I tried to use the data \"as is\" - and push it just into the EffNetB0.</p>\n<p>I took 8 slices (with equal steps) from each of the first three views of the structural multi-parametric MRI (mpMRI) scans:</p>\n<ul>\n<li>Fluid Attenuated Inversion Recovery (FLAIR)</li>\n<li>T1-weighted pre-contrast (T1w)</li>\n<li>T1-weighted post-contrast (T1Gd)</li>\n</ul>\n<p>Then I take a mean of slices of each type of scan and get 3 \"channels\": <br>\n<img src=\"https://i.imgur.com/GhT4iks.png\" alt=\"\"></p>\n<p>Then merge them into one \"image\" and push them to the EfficientNetB0 model.</p>\n<p>I trained it 2 epochs and after the second I have the next metrics:</p>\n<pre><code>[Epoch Train: 2] loss: 0.68856, score: 0.56197, time: 50 s\n[Epoch Valid: 2] loss: 0.70479, score: 0.52137, time: 12 s\n</code></pre>\n<p>As we can see my loss is equal to ~0.69. It means that the model is not good and the LB score is overfitted.</p>\n<p>From this experiment, I think this approach doesn't have the chance of success. Further improvement of the algorithm should be carried out in two directions:</p>\n<ul>\n<li>To use models that can accept 3D scans as input and are specially pre-trained for medical tasks.</li>\n<li>Cleaning and working with data, as the scans have different tilt angles and parameters.</li>\n</ul>\n<p>EDA and starter modeling you can find in my notebook:</p>\n<ul>\n<li><a href=\"https://www.kaggle.com/ihelon/brain-tumor-eda-with-animations-and-modeling\" target=\"_blank\">EDA &amp; train &amp; inference</a></li>\n</ul>",
      "rawMarkdown": "I experimented a bit with the data for this competition. I tried to use the data \"as is\" - and push it just into the EffNetB0.\n\nI took 8 slices (with equal steps) from each of the first three views of the structural multi-parametric MRI (mpMRI) scans:\n- Fluid Attenuated Inversion Recovery (FLAIR)\n- T1-weighted pre-contrast (T1w)\n- T1-weighted post-contrast (T1Gd)\n\nThen I take a mean of slices of each type of scan and get 3 \"channels\": \n![](https://i.imgur.com/GhT4iks.png)\n\nThen merge them into one \"image\" and push them to the EfficientNetB0 model.\n\nI trained it 2 epochs and after the second I have the next metrics:\n```\n[Epoch Train: 2] loss: 0.68856, score: 0.56197, time: 50 s\n[Epoch Valid: 2] loss: 0.70479, score: 0.52137, time: 12 s\n```\nAs we can see my loss is equal to ~0.69. It means that the model is not good and the LB score is overfitted.\n\nFrom this experiment, I think this approach doesn't have the chance of success. Further improvement of the algorithm should be carried out in two directions:\n- To use models that can accept 3D scans as input and are specially pre-trained for medical tasks.\n- Cleaning and working with data, as the scans have different tilt angles and parameters.\n\nEDA and starter modeling you can find in my notebook:\n- [EDA & train & inference](https://www.kaggle.com/ihelon/brain-tumor-eda-with-animations-and-modeling)",
      "votes": 27
    },
    {
      "id": 1390455,
      "postDate": "2021-07-16T16:56:04.330Z",
      "content": "<p>Hi,</p>\n<p>I think preprocessing the data in order to make them comparable (e.g. different plane) is going to be a key challenge here. </p>",
      "rawMarkdown": "Hi,\n\nI think preprocessing the data in order to make them comparable (e.g. different plane) is going to be a key challenge here. ",
      "votes": 3,
      "replies": [
        {
          "id": 1390479,
          "postDate": "2021-07-16T17:12:51.160Z",
          "content": "<p>Yes, it is a good point.</p>",
          "rawMarkdown": "Yes, it is a good point."
        },
        {
          "id": 1390748,
          "postDate": "2021-07-17T04:08:39.227Z",
          "content": "<p>Luca, what do you mean? Can you explain more thoroughly and give some reference?</p>",
          "rawMarkdown": "Luca, what do you mean? Can you explain more thoroughly and give some reference?"
        },
        {
          "id": 1391195,
          "postDate": "2021-07-17T11:35:15.517Z",
          "content": "<p>You have a sequence of picture representing a 3D-Model of the brain. </p>\n<p>Is is shown clearly in Yaroslavs notebook. In just a few of the pictures the tumor is seen. </p>\n<p>Then not every patient has the same plane:</p>\n<p><a href=\"https://www.kaggle.com/davidbroberts/determining-mr-image-planes\" target=\"_blank\">https://www.kaggle.com/davidbroberts/determining-mr-image-planes</a></p>\n<p>So in may point of view the first goal is going to be to preprocess the data in a way to make them comparable in order that a neural network can extract relevant information.</p>\n<p>A 3D-Reconstraction may an approach. </p>\n<p>Unfortunately I dont know yet how to do this.</p>",
          "rawMarkdown": "You have a sequence of picture representing a 3D-Model of the brain. \n\nIs is shown clearly in Yaroslavs notebook. In just a few of the pictures the tumor is seen. \n\nThen not every patient has the same plane:\n\nhttps://www.kaggle.com/davidbroberts/determining-mr-image-planes\n\nSo in may point of view the first goal is going to be to preprocess the data in a way to make them comparable in order that a neural network can extract relevant information.\n\nA 3D-Reconstraction may an approach. \n\nUnfortunately I dont know yet how to do this.",
          "votes": 3
        }
      ]
    },
    {
      "id": 1390391,
      "postDate": "2021-07-16T15:59:26.537Z",
      "content": "<p>Congrats <a href=\"https://www.kaggle.com/ihelon\" target=\"_blank\">@ihelon</a> on becoming Grandmaster!!!!, you deserve it, your notebooks deserve it</p>",
      "rawMarkdown": "Congrats @ihelon on becoming Grandmaster!!!!, you deserve it, your notebooks deserve it",
      "votes": 1,
      "replies": [
        {
          "id": 1390424,
          "postDate": "2021-07-16T16:21:44.097Z",
          "content": "<p><a href=\"https://www.kaggle.com/mrinath\" target=\"_blank\">@mrinath</a> Thank you! I will try to do it better and better!😊</p>",
          "rawMarkdown": "@mrinath Thank you! I will try to do it better and better!😊",
          "votes": 1
        }
      ]
    },
    {
      "id": 1388837,
      "postDate": "2021-07-15T08:55:28.160Z",
      "content": "<p>In the given 3 images, which one is the merged image? or you use these 3 to make a merged image?</p>",
      "rawMarkdown": "In the given 3 images, which one is the merged image? or you use these 3 to make a merged image?",
      "votes": 1,
      "replies": [
        {
          "id": 1388844,
          "postDate": "2021-07-15T09:02:42.217Z",
          "content": "<p><a href=\"https://www.kaggle.com/mrinath\" target=\"_blank\">@mrinath</a> each of the images on the graph is merged by selected scan type: </p>\n<ul>\n<li>Fluid Attenuated Inversion Recovery (FLAIR)</li>\n<li>T1-weighted pre-contrast (T1w)</li>\n<li>T1-weighted post-contrast (T1Gd)</li>\n</ul>\n<p>They look similar, but it is problem with the scan and I think we need to preprocess scans before any train procedures.</p>",
          "rawMarkdown": "@mrinath each of the images on the graph is merged by selected scan type: \n- Fluid Attenuated Inversion Recovery (FLAIR)\n- T1-weighted pre-contrast (T1w)\n- T1-weighted post-contrast (T1Gd)\n\nThey look similar, but it is problem with the scan and I think we need to preprocess scans before any train procedures.",
          "votes": 2
        },
        {
          "id": 1389852,
          "postDate": "2021-07-16T06:17:05.110Z",
          "content": "<p>why you are using only three categories? <a href=\"https://www.kaggle.com/ihelon\" target=\"_blank\">@ihelon</a>  any reason behind that?</p>",
          "rawMarkdown": "why you are using only three categories? @ihelon  any reason behind that?",
          "votes": 1
        },
        {
          "id": 1389983,
          "postDate": "2021-07-16T08:56:26.883Z",
          "content": "<p><a href=\"https://www.kaggle.com/santhoshkumarv\" target=\"_blank\">@santhoshkumarv</a> because the default EffNet model works with a three-channel image.</p>",
          "rawMarkdown": "@santhoshkumarv because the default EffNet model works with a three-channel image."
        }
      ]
    },
    {
      "id": 1388547,
      "postDate": "2021-07-15T04:12:29.077Z",
      "content": "<p>Thanks for such a great notebook. <br>\nIf understood correctly you took <strong>(FLAIR</strong>, <strong>T1w</strong>, <strong>T1Gd)</strong> these three folders of images and experimented.<br>\nDid you resize the images and take out the blank images?</p>",
      "rawMarkdown": "Thanks for such a great notebook. \nIf understood correctly you took **(FLAIR**, **T1w**, **T1Gd)** these three folders of images and experimented.\nDid you resize the images and take out the blank images?",
      "votes": 1,
      "replies": [
        {
          "id": 1388794,
          "postDate": "2021-07-15T08:29:09.350Z",
          "content": "<p>Thanks!<br>\nYes, it's rigth.<br>\nI resized them to 256*256 but I don't filter blank images. It's should be a little bit harder implementation. And I don't touch images on the corners (I think many blank images are there).</p>",
          "rawMarkdown": "Thanks!\nYes, it's rigth.\nI resized them to 256*256 but I don't filter blank images. It's should be a little bit harder implementation. And I don't touch images on the corners (I think many blank images are there).",
          "votes": 1
        }
      ]
    },
    {
      "id": 1389873,
      "postDate": "2021-07-16T06:58:31.903Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 1390455,
      "author_name": "LucaMTB",
      "author_url": "",
      "post_date": "2021-07-16T16:56:04.330000",
      "content": "<p>Hi,</p>\n<p>I think preprocessing the data in order to make them comparable (e.g. different plane) is going to be a key challenge here. </p>",
      "votes": 3,
      "replies": [
        {
          "id": 1390479,
          "author_name": "Yaroslav Isaienkov",
          "author_url": "",
          "post_date": "2021-07-16T17:12:51.160000",
          "content": "<p>Yes, it is a good point.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1390748,
          "author_name": "Arslanov",
          "author_url": "",
          "post_date": "2021-07-17T04:08:39.227000",
          "content": "<p>Luca, what do you mean? Can you explain more thoroughly and give some reference?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1391195,
          "author_name": "LucaMTB",
          "author_url": "",
          "post_date": "2021-07-17T11:35:15.517000",
          "content": "<p>You have a sequence of picture representing a 3D-Model of the brain. </p>\n<p>Is is shown clearly in Yaroslavs notebook. In just a few of the pictures the tumor is seen. </p>\n<p>Then not every patient has the same plane:</p>\n<p><a href=\"https://www.kaggle.com/davidbroberts/determining-mr-image-planes\" target=\"_blank\">https://www.kaggle.com/davidbroberts/determining-mr-image-planes</a></p>\n<p>So in may point of view the first goal is going to be to preprocess the data in a way to make them comparable in order that a neural network can extract relevant information.</p>\n<p>A 3D-Reconstraction may an approach. </p>\n<p>Unfortunately I dont know yet how to do this.</p>",
          "votes": 3,
          "replies": []
        }
      ]
    },
    {
      "id": 1390391,
      "author_name": "DeepUnderstanding",
      "author_url": "",
      "post_date": "2021-07-16T15:59:26.537000",
      "content": "<p>Congrats <a href=\"https://www.kaggle.com/ihelon\" target=\"_blank\">@ihelon</a> on becoming Grandmaster!!!!, you deserve it, your notebooks deserve it</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1390424,
          "author_name": "Yaroslav Isaienkov",
          "author_url": "",
          "post_date": "2021-07-16T16:21:44.097000",
          "content": "<p><a href=\"https://www.kaggle.com/mrinath\" target=\"_blank\">@mrinath</a> Thank you! I will try to do it better and better!😊</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1388837,
      "author_name": "DeepUnderstanding",
      "author_url": "",
      "post_date": "2021-07-15T08:55:28.160000",
      "content": "<p>In the given 3 images, which one is the merged image? or you use these 3 to make a merged image?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1388844,
          "author_name": "Yaroslav Isaienkov",
          "author_url": "",
          "post_date": "2021-07-15T09:02:42.217000",
          "content": "<p><a href=\"https://www.kaggle.com/mrinath\" target=\"_blank\">@mrinath</a> each of the images on the graph is merged by selected scan type: </p>\n<ul>\n<li>Fluid Attenuated Inversion Recovery (FLAIR)</li>\n<li>T1-weighted pre-contrast (T1w)</li>\n<li>T1-weighted post-contrast (T1Gd)</li>\n</ul>\n<p>They look similar, but it is problem with the scan and I think we need to preprocess scans before any train procedures.</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 1389852,
          "author_name": "Santhoshkumar",
          "author_url": "",
          "post_date": "2021-07-16T06:17:05.110000",
          "content": "<p>why you are using only three categories? <a href=\"https://www.kaggle.com/ihelon\" target=\"_blank\">@ihelon</a>  any reason behind that?</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1389983,
          "author_name": "Yaroslav Isaienkov",
          "author_url": "",
          "post_date": "2021-07-16T08:56:26.883000",
          "content": "<p><a href=\"https://www.kaggle.com/santhoshkumarv\" target=\"_blank\">@santhoshkumarv</a> because the default EffNet model works with a three-channel image.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1388547,
      "author_name": "tharun_01",
      "author_url": "",
      "post_date": "2021-07-15T04:12:29.077000",
      "content": "<p>Thanks for such a great notebook. <br>\nIf understood correctly you took <strong>(FLAIR</strong>, <strong>T1w</strong>, <strong>T1Gd)</strong> these three folders of images and experimented.<br>\nDid you resize the images and take out the blank images?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1388794,
          "author_name": "Yaroslav Isaienkov",
          "author_url": "",
          "post_date": "2021-07-15T08:29:09.350000",
          "content": "<p>Thanks!<br>\nYes, it's rigth.<br>\nI resized them to 256*256 but I don't filter blank images. It's should be a little bit harder implementation. And I don't touch images on the corners (I think many blank images are there).</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1389873,
      "author_name": "",
      "author_url": "",
      "post_date": "2021-07-16T06:58:31.903000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1388374": "I experimented a bit with the data for this competition. I tried to use the data \"as is\" - and push it just into the EffNetB0.\n\nI took 8 slices (with equal steps) from each of the first three views of the structural multi-parametric MRI (mpMRI) scans:\n- Fluid Attenuated Inversion Recovery (FLAIR)\n- T1-weighted pre-contrast (T1w)\n- T1-weighted post-contrast (T1Gd)\n\nThen I take a mean of slices of each type of scan and get 3 \"channels\": \n![](https://i.imgur.com/GhT4iks.png)\n\nThen merge them into one \"image\" and push them to the EfficientNetB0 model.\n\nI trained it 2 epochs and after the second I have the next metrics:\n```\n[Epoch Train: 2] loss: 0.68856, score: 0.56197, time: 50 s\n[Epoch Valid: 2] loss: 0.70479, score: 0.52137, time: 12 s\n```\nAs we can see my loss is equal to ~0.69. It means that the model is not good and the LB score is overfitted.\n\nFrom this experiment, I think this approach doesn't have the chance of success. Further improvement of the algorithm should be carried out in two directions:\n- To use models that can accept 3D scans as input and are specially pre-trained for medical tasks.\n- Cleaning and working with data, as the scans have different tilt angles and parameters.\n\nEDA and starter modeling you can find in my notebook:\n- [EDA & train & inference](https://www.kaggle.com/ihelon/brain-tumor-eda-with-animations-and-modeling)",
    "1390455": "Hi,\n\nI think preprocessing the data in order to make them comparable (e.g. different plane) is going to be a key challenge here. ",
    "1390391": "Congrats @ihelon on becoming Grandmaster!!!!, you deserve it, your notebooks deserve it",
    "1388837": "In the given 3 images, which one is the merged image? or you use these 3 to make a merged image?",
    "1388547": "Thanks for such a great notebook. \nIf understood correctly you took **(FLAIR**, **T1w**, **T1Gd)** these three folders of images and experimented.\nDid you resize the images and take out the blank images?",
    "1389873": ""
  }
}