{
  "id": 432217,
  "title": "2D and 2.5D experiments",
  "url": "/competitions/rsna-2023-abdominal-trauma-detection/discussion/432217",
  "author_name": "Gunes Evitan",
  "post_date": "2023-08-16T13:42:06.693000",
  "votes": 12,
  "comment_count": 5,
  "views": 0,
  "content": "<p>I started my experiments with 2D (middle slice) then I tried stacking more slices on the channel dimension</p>\n<ul>\n<li>3 channel 45th + middle slice + 55th percentile slices</li>\n<li>3 channel 40th + middle slice + 60th percentile slices</li>\n<li>3 channel 25th + middle slice + 75th percentile slices</li>\n<li>5 channel 20th + 40th + middle slice + 60th + 80th percentile slices</li>\n<li>9 channel 10th + … + 90th percentile slices</li>\n</ul>\n<p>and I can say that there wasn't any significant improvement over using only the middle slice. My best OOF score is between 0.53 and 0.54. I think we should do 3D approach.</p>",
  "messages": [
    {
      "id": 2393741,
      "postDate": "2023-08-16T13:42:06.693Z",
      "content": "<p>I started my experiments with 2D (middle slice) then I tried stacking more slices on the channel dimension</p>\n<ul>\n<li>3 channel 45th + middle slice + 55th percentile slices</li>\n<li>3 channel 40th + middle slice + 60th percentile slices</li>\n<li>3 channel 25th + middle slice + 75th percentile slices</li>\n<li>5 channel 20th + 40th + middle slice + 60th + 80th percentile slices</li>\n<li>9 channel 10th + … + 90th percentile slices</li>\n</ul>\n<p>and I can say that there wasn't any significant improvement over using only the middle slice. My best OOF score is between 0.53 and 0.54. I think we should do 3D approach.</p>",
      "rawMarkdown": "I started my experiments with 2D (middle slice) then I tried stacking more slices on the channel dimension\n* 3 channel 45th + middle slice + 55th percentile slices\n* 3 channel 40th + middle slice + 60th percentile slices\n* 3 channel 25th + middle slice + 75th percentile slices\n* 5 channel 20th + 40th + middle slice + 60th + 80th percentile slices\n* 9 channel 10th + ... + 90th percentile slices\n\nand I can say that there wasn't any significant improvement over using only the middle slice. My best OOF score is between 0.53 and 0.54. I think we should do 3D approach.",
      "votes": 10
    },
    {
      "id": 2397926,
      "postDate": "2023-08-19T10:28:38.217Z",
      "content": "<p>I agree, I think a winning solution would be 2 stage prediction, the first one segmentation model to get organ location, and the second one is a 3d model.</p>",
      "rawMarkdown": "I agree, I think a winning solution would be 2 stage prediction, the first one segmentation model to get organ location, and the second one is a 3d model.",
      "votes": 5
    },
    {
      "id": 2399096,
      "postDate": "2023-08-20T06:28:28.063Z",
      "content": "<p>Worked on a previous competition using a 3D model.  I agree that 3D takes care of many of the issues in this data set, but I consumed more compute than available on kaggle on that past attempt.  </p>\n<p>Very hopeful to see a 3D model shared before/after end of the competition.</p>",
      "rawMarkdown": "Worked on a previous competition using a 3D model.  I agree that 3D takes care of many of the issues in this data set, but I consumed more compute than available on kaggle on that past attempt.  \n\nVery hopeful to see a 3D model shared before/after end of the competition.",
      "votes": 2,
      "replies": [
        {
          "id": 2407551,
          "postDate": "2023-08-25T06:06:54.137Z",
          "content": "<p>You are right but you can also resize/interpolate on 3D volume. Check this <a href=\"https://stackoverflow.com/a/47785863/7975962\" target=\"_blank\">https://stackoverflow.com/a/47785863/7975962</a></p>",
          "rawMarkdown": "You are right but you can also resize/interpolate on 3D volume. Check this https://stackoverflow.com/a/47785863/7975962"
        }
      ]
    },
    {
      "id": 2395170,
      "postDate": "2023-08-17T10:36:39.193Z",
      "content": "<p>Do you mean, we can use the 3D model method for the middle slice to get similar and good results?</p>",
      "rawMarkdown": "Do you mean, we can use the 3D model method for the middle slice to get similar and good results?",
      "isDeleted": true,
      "replies": [
        {
          "id": 2407554,
          "postDate": "2023-08-25T06:07:54.143Z",
          "content": "<p>3D approach is using all of the slices and stacking them on the z axis.</p>",
          "rawMarkdown": "3D approach is using all of the slices and stacking them on the z axis."
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 2397926,
      "author_name": "AnatoliK",
      "author_url": "",
      "post_date": "2023-08-19T10:28:38.217000",
      "content": "<p>I agree, I think a winning solution would be 2 stage prediction, the first one segmentation model to get organ location, and the second one is a 3d model.</p>",
      "votes": 5,
      "replies": []
    },
    {
      "id": 2399096,
      "author_name": "PC Jimmmy",
      "author_url": "",
      "post_date": "2023-08-20T06:28:28.063000",
      "content": "<p>Worked on a previous competition using a 3D model.  I agree that 3D takes care of many of the issues in this data set, but I consumed more compute than available on kaggle on that past attempt.  </p>\n<p>Very hopeful to see a 3D model shared before/after end of the competition.</p>",
      "votes": 2,
      "replies": [
        {
          "id": 2407551,
          "author_name": "Gunes Evitan",
          "author_url": "",
          "post_date": "2023-08-25T06:06:54.137000",
          "content": "<p>You are right but you can also resize/interpolate on 3D volume. Check this <a href=\"https://stackoverflow.com/a/47785863/7975962\" target=\"_blank\">https://stackoverflow.com/a/47785863/7975962</a></p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 2395170,
      "author_name": "",
      "author_url": "",
      "post_date": "2023-08-17T10:36:39.193000",
      "content": "<p>Do you mean, we can use the 3D model method for the middle slice to get similar and good results?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 2407554,
          "author_name": "Gunes Evitan",
          "author_url": "",
          "post_date": "2023-08-25T06:07:54.143000",
          "content": "<p>3D approach is using all of the slices and stacking them on the z axis.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2393741": "I started my experiments with 2D (middle slice) then I tried stacking more slices on the channel dimension\n* 3 channel 45th + middle slice + 55th percentile slices\n* 3 channel 40th + middle slice + 60th percentile slices\n* 3 channel 25th + middle slice + 75th percentile slices\n* 5 channel 20th + 40th + middle slice + 60th + 80th percentile slices\n* 9 channel 10th + ... + 90th percentile slices\n\nand I can say that there wasn't any significant improvement over using only the middle slice. My best OOF score is between 0.53 and 0.54. I think we should do 3D approach.",
    "2397926": "I agree, I think a winning solution would be 2 stage prediction, the first one segmentation model to get organ location, and the second one is a 3d model.",
    "2399096": "Worked on a previous competition using a 3D model.  I agree that 3D takes care of many of the issues in this data set, but I consumed more compute than available on kaggle on that past attempt.  \n\nVery hopeful to see a 3D model shared before/after end of the competition.",
    "2395170": "Do you mean, we can use the 3D model method for the middle slice to get similar and good results?"
  }
}