{
  "id": 431890,
  "title": "2D Pipeline for Bowel and Extravasation",
  "url": "/competitions/rsna-2023-abdominal-trauma-detection/discussion/431890",
  "author_name": "JanGlinko2",
  "post_date": "2023-08-15T09:26:12.845000",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I wanted to start building my pipeline from a 2D pipeline for Bowel and Extravasation classification.<br>\nI'm assuming, that all scans containing positive samples are listed in <code>image_level_label.csv</code>, while the rest are negative samples. <br>\nHowever, I'm afraid of some dataset characteristics, namely:</p>\n<ul>\n<li>246 / 3147 patients have bowel, extravasation, or both injuries,</li>\n<li>0 of 246 patients have all scan slices positive in that context,</li>\n<li>hard cases, where only 1 positive slice out of all slices is 1.0 in ground truth data (I'm assuming, that scoring will work in the same way), i.e.:<ul>\n<li>14018, 1 out of 199</li>\n<li>15472, 2 out of 233</li>\n<li>1675, 1 out of 665</li>\n<li>19249, 1 out of 513</li>\n<li>22755, 1 out of 222</li>\n<li>26324, 1 out of 105</li>\n<li>36686, 1 out of 326<br>\nand many more with less than 10 out of N.</li></ul></li>\n</ul>\n<p>To sum up, the amount of Bowel or Extravasation is very low among all patients and I think it's very easy to predict a false positive, while 1 positive slice is enough for positive classification expectation.</p>\n<p>For me, both Bowel and Extravasation are perfect candidates for hardcoded predictions. </p>\n<p>Have you proceeded with any experiments in this field already? I would like to hear your opinions.</p>",
  "messages": [
    {
      "id": 2392051,
      "postDate": "2023-08-15T12:46:43.953Z",
      "content": "<p>Bowel was easier for me to predict while there wasn't much signal for extrasavation injuries. I'll try to make my first submission today to see if my model is working or not.</p>",
      "rawMarkdown": "Bowel was easier for me to predict while there wasn't much signal for extrasavation injuries. I'll try to make my first submission today to see if my model is working or not.",
      "votes": 1
    },
    {
      "id": 2391658,
      "postDate": "2023-08-15T09:26:12.847Z",
      "content": "<p>I wanted to start building my pipeline from a 2D pipeline for Bowel and Extravasation classification.<br>\nI'm assuming, that all scans containing positive samples are listed in <code>image_level_label.csv</code>, while the rest are negative samples. <br>\nHowever, I'm afraid of some dataset characteristics, namely:</p>\n<ul>\n<li>246 / 3147 patients have bowel, extravasation, or both injuries,</li>\n<li>0 of 246 patients have all scan slices positive in that context,</li>\n<li>hard cases, where only 1 positive slice out of all slices is 1.0 in ground truth data (I'm assuming, that scoring will work in the same way), i.e.:<ul>\n<li>14018, 1 out of 199</li>\n<li>15472, 2 out of 233</li>\n<li>1675, 1 out of 665</li>\n<li>19249, 1 out of 513</li>\n<li>22755, 1 out of 222</li>\n<li>26324, 1 out of 105</li>\n<li>36686, 1 out of 326<br>\nand many more with less than 10 out of N.</li></ul></li>\n</ul>\n<p>To sum up, the amount of Bowel or Extravasation is very low among all patients and I think it's very easy to predict a false positive, while 1 positive slice is enough for positive classification expectation.</p>\n<p>For me, both Bowel and Extravasation are perfect candidates for hardcoded predictions. </p>\n<p>Have you proceeded with any experiments in this field already? I would like to hear your opinions.</p>",
      "rawMarkdown": "I wanted to start building my pipeline from a 2D pipeline for Bowel and Extravasation classification.\nI'm assuming, that all scans containing positive samples are listed in `image_level_label.csv`, while the rest are negative samples. \nHowever, I'm afraid of some dataset characteristics, namely:\n* 246 / 3147 patients have bowel, extravasation, or both injuries,\n* 0 of 246 patients have all scan slices positive in that context,\n* hard cases, where only 1 positive slice out of all slices is 1.0 in ground truth data (I'm assuming, that scoring will work in the same way), i.e.:\n     - 14018, 1 out of 199\n     - 15472, 2 out of 233\n     - 1675, 1 out of 665\n     - 19249, 1 out of 513\n     - 22755, 1 out of 222\n     - 26324, 1 out of 105\n     - 36686, 1 out of 326\nand many more with less than 10 out of N.\n\nTo sum up, the amount of Bowel or Extravasation is very low among all patients and I think it's very easy to predict a false positive, while 1 positive slice is enough for positive classification expectation.\n\nFor me, both Bowel and Extravasation are perfect candidates for hardcoded predictions. \n\nHave you proceeded with any experiments in this field already? I would like to hear your opinions.\n",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 2392051,
      "author_name": "Gunes Evitan",
      "author_url": "",
      "post_date": "2023-08-15T12:46:43.953000",
      "content": "<p>Bowel was easier for me to predict while there wasn't much signal for extrasavation injuries. I'll try to make my first submission today to see if my model is working or not.</p>",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2392051": "Bowel was easier for me to predict while there wasn't much signal for extrasavation injuries. I'll try to make my first submission today to see if my model is working or not.",
    "2391658": "I wanted to start building my pipeline from a 2D pipeline for Bowel and Extravasation classification.\nI'm assuming, that all scans containing positive samples are listed in `image_level_label.csv`, while the rest are negative samples. \nHowever, I'm afraid of some dataset characteristics, namely:\n* 246 / 3147 patients have bowel, extravasation, or both injuries,\n* 0 of 246 patients have all scan slices positive in that context,\n* hard cases, where only 1 positive slice out of all slices is 1.0 in ground truth data (I'm assuming, that scoring will work in the same way), i.e.:\n     - 14018, 1 out of 199\n     - 15472, 2 out of 233\n     - 1675, 1 out of 665\n     - 19249, 1 out of 513\n     - 22755, 1 out of 222\n     - 26324, 1 out of 105\n     - 36686, 1 out of 326\nand many more with less than 10 out of N.\n\nTo sum up, the amount of Bowel or Extravasation is very low among all patients and I think it's very easy to predict a false positive, while 1 positive slice is enough for positive classification expectation.\n\nFor me, both Bowel and Extravasation are perfect candidates for hardcoded predictions. \n\nHave you proceeded with any experiments in this field already? I would like to hear your opinions.\n"
  }
}