{
  "id": 222528,
  "title": "Synthetized images from a domain expert, abnormal ETTs",
  "url": "/competitions/ranzcr-clip-catheter-line-classification/discussion/222528",
  "author_name": "",
  "post_date": "2021-02-27T15:29:48.686578300Z",
  "votes": 12,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi everyone,</p>\n<p>TL;DR</p>\n<blockquote>\n  <p>I share 144 + 56 pairs of <strong>abnormally placed ETT</strong>  + corresponding mask on x-ray with you -&gt; <a href=\"https://www.kaggle.com/sandorkonya/synthetized-endotracheal-tube-data-on-chest-xray\" target=\"_blank\">here</a></p>\n</blockquote>\n<p>First, i wrote a script that <strong>procedurally generates synthetized endotracheal tubes</strong>  with random settings (wall thickness, rotation) along a curved path, that i annotated manually on the images beforehand, and places these ETTs on the input x-ray images.</p>\n<p>The path's last point is <strong>always below the trachea bifurcation</strong>, this means, every generated image contains one <strong>abnormally positioned ETT</strong>!</p>\n<p>The dataset contains 2 sets of image + mask pairs: manually_blended(56 pairs) and automatically_blended(144 pairs)  - considering that we only have ~78ish abnormally placed ETT in the dataset, this should help to detect these.</p>\n<p>For visual types, the dataset contains the images on the right of the following figure: </p>\n<p><img src=\"https://i.postimg.cc/9fhBfp0Q/synthetic-ETT-2.jpg\" alt=\"creation of synthetic images\"></p>\n<p>I describe the difference between the manually blended and automatically blended ones in the dataset description.</p>\n<p><img src=\"https://i.postimg.cc/HnthdkbB/synthetic-ETT-3.jpg\" alt=\"difference between the two set\"></p>\n<p><strong>Note</strong>: if you use the set for training, every other annotation from the original data (NGT, CVC) should remain the same!</p>\n<p>would love to see 2 things:</p>\n<ol>\n<li>how does training with this additional set affect your LB?</li>\n<li>just out of curiosity - could we detect any ETT on unseen image by using solely the 144 images for training?</li>\n</ol>\n<p>Just for fun, some of the generated ETT (the color is only artistic touch :P )<br>\n<img src=\"https://i.postimg.cc/1zqsGbxs/etts.jpg\" alt=\"ETTs\"></p>\n<p>This was quite a time consuming job with the manual annotations and then the procedural ETT generation… so let me know if you like it!</p>",
  "messages": [
    {
      "id": "1220120",
      "postDate": "02/27/2021 15:29:48",
      "content": "<p>Hi everyone,</p>\n<p>TL;DR</p>\n<blockquote>\n  <p>I share 144 + 56 pairs of <strong>abnormally placed ETT</strong>  + corresponding mask on x-ray with you -&gt; <a href=\"https://www.kaggle.com/sandorkonya/synthetized-endotracheal-tube-data-on-chest-xray\" target=\"_blank\">here</a></p>\n</blockquote>\n<p>First, i wrote a script that <strong>procedurally generates synthetized endotracheal tubes</strong>  with random settings (wall thickness, rotation) along a curved path, that i annotated manually on the images beforehand, and places these ETTs on the input x-ray images.</p>\n<p>The path's last point is <strong>always below the trachea bifurcation</strong>, this means, every generated image contains one <strong>abnormally positioned ETT</strong>!</p>\n<p>The dataset contains 2 sets of image + mask pairs: manually_blended(56 pairs) and automatically_blended(144 pairs)  - considering that we only have ~78ish abnormally placed ETT in the dataset, this should help to detect these.</p>\n<p>For visual types, the dataset contains the images on the right of the following figure: </p>\n<p><img src=\"https://i.postimg.cc/9fhBfp0Q/synthetic-ETT-2.jpg\" alt=\"creation of synthetic images\"></p>\n<p>I describe the difference between the manually blended and automatically blended ones in the dataset description.</p>\n<p><img src=\"https://i.postimg.cc/HnthdkbB/synthetic-ETT-3.jpg\" alt=\"difference between the two set\"></p>\n<p><strong>Note</strong>: if you use the set for training, every other annotation from the original data (NGT, CVC) should remain the same!</p>\n<p>would love to see 2 things:</p>\n<ol>\n<li>how does training with this additional set affect your LB?</li>\n<li>just out of curiosity - could we detect any ETT on unseen image by using solely the 144 images for training?</li>\n</ol>\n<p>Just for fun, some of the generated ETT (the color is only artistic touch :P )<br>\n<img src=\"https://i.postimg.cc/1zqsGbxs/etts.jpg\" alt=\"ETTs\"></p>\n<p>This was quite a time consuming job with the manual annotations and then the procedural ETT generation… so let me know if you like it!</p>",
      "rawMarkdown": "Hi everyone,\n\nTL;DR\n> \nI share 144 + 56 pairs of **abnormally placed ETT**  + corresponding mask on x-ray with you -> [here](https://www.kaggle.com/sandorkonya/synthetized-endotracheal-tube-data-on-chest-xray)\n\n\nFirst, i wrote a script that **procedurally generates synthetized endotracheal tubes**  with random settings (wall thickness, rotation) along a curved path, that i annotated manually on the images beforehand, and places these ETTs on the input x-ray images.\n\nThe path's last point is **always below the trachea bifurcation**, this means, every generated image contains one **abnormally positioned ETT**!\n\nThe dataset contains 2 sets of image + mask pairs: manually_blended(56 pairs) and automatically_blended(144 pairs)  - considering that we only have ~78ish abnormally placed ETT in the dataset, this should help to detect these.\n\nFor visual types, the dataset contains the images on the right of the following figure: \n\n![creation of synthetic images](https://i.postimg.cc/9fhBfp0Q/synthetic-ETT-2.jpg)\n\nI describe the difference between the manually blended and automatically blended ones in the dataset description.\n\n![difference between the two set](https://i.postimg.cc/HnthdkbB/synthetic-ETT-3.jpg)\n\n**Note**: if you use the set for training, every other annotation from the original data (NGT, CVC) should remain the same!\n\n\nwould love to see 2 things:\n\n1. how does training with this additional set affect your LB?\n2. just out of curiosity - could we detect any ETT on unseen image by using solely the 144 images for training?\n\n\nJust for fun, some of the generated ETT (the color is only artistic touch :P )\n![ETTs](https://i.postimg.cc/1zqsGbxs/etts.jpg)\n\nThis was quite a time consuming job with the manual annotations and then the procedural ETT generation... so let me know if you like it!",
      "votes": null
    },
    {
      "id": "1220478",
      "postDate": "02/28/2021 03:48:30",
      "content": "<p>can you share the files of masks? Thank you</p>",
      "rawMarkdown": "can you share the files of masks? Thank you",
      "votes": null
    },
    {
      "id": "1220498",
      "postDate": "02/28/2021 04:21:25",
      "content": "<p>Hi,</p>\n<p>there are masks in both set with mask_xxxx.jpg!<br>\nRegards!</p>",
      "rawMarkdown": "Hi,\n\nthere are masks in both set with mask_xxxx.jpg!\nRegards!",
      "votes": null
    },
    {
      "id": "1220555",
      "postDate": "02/28/2021 05:25:51",
      "content": "<p>Thank you👍</p>",
      "rawMarkdown": "Thank you👍",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1220478,
      "author_name": "dldmw579",
      "author_url": "",
      "post_date": "02/28/2021 03:48:30",
      "content": "<p>can you share the files of masks? Thank you</p>",
      "votes": null,
      "replies": [
        {
          "id": 1220498,
          "author_name": "sandorkonya",
          "author_url": "",
          "post_date": "02/28/2021 04:21:25",
          "content": "<p>Hi,</p>\n<p>there are masks in both set with mask_xxxx.jpg!<br>\nRegards!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1220555,
          "author_name": "dldmw579",
          "author_url": "",
          "post_date": "02/28/2021 05:25:51",
          "content": "<p>Thank you👍</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1220120": "Hi everyone,\n\nTL;DR\n> \nI share 144 + 56 pairs of **abnormally placed ETT**  + corresponding mask on x-ray with you -> [here](https://www.kaggle.com/sandorkonya/synthetized-endotracheal-tube-data-on-chest-xray)\n\n\nFirst, i wrote a script that **procedurally generates synthetized endotracheal tubes**  with random settings (wall thickness, rotation) along a curved path, that i annotated manually on the images beforehand, and places these ETTs on the input x-ray images.\n\nThe path's last point is **always below the trachea bifurcation**, this means, every generated image contains one **abnormally positioned ETT**!\n\nThe dataset contains 2 sets of image + mask pairs: manually_blended(56 pairs) and automatically_blended(144 pairs)  - considering that we only have ~78ish abnormally placed ETT in the dataset, this should help to detect these.\n\nFor visual types, the dataset contains the images on the right of the following figure: \n\n![creation of synthetic images](https://i.postimg.cc/9fhBfp0Q/synthetic-ETT-2.jpg)\n\nI describe the difference between the manually blended and automatically blended ones in the dataset description.\n\n![difference between the two set](https://i.postimg.cc/HnthdkbB/synthetic-ETT-3.jpg)\n\n**Note**: if you use the set for training, every other annotation from the original data (NGT, CVC) should remain the same!\n\n\nwould love to see 2 things:\n\n1. how does training with this additional set affect your LB?\n2. just out of curiosity - could we detect any ETT on unseen image by using solely the 144 images for training?\n\n\nJust for fun, some of the generated ETT (the color is only artistic touch :P )\n![ETTs](https://i.postimg.cc/1zqsGbxs/etts.jpg)\n\nThis was quite a time consuming job with the manual annotations and then the procedural ETT generation... so let me know if you like it!",
    "1220478": "can you share the files of masks? Thank you",
    "1220498": "Hi,\n\nthere are masks in both set with mask_xxxx.jpg!\nRegards!",
    "1220555": "Thank you👍"
  },
  "source": "meta"
}