{
  "id": 362633,
  "title": "Hunting for 3D Classification Solutions",
  "url": "/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/362633",
  "author_name": "Qishen Ha",
  "post_date": "2022-10-28T08:33:05.164000",
  "votes": 6,
  "comment_count": 7,
  "views": 0,
  "content": "<p>I, as well as many others, reported that could not get a good score using 3D classification.</p>\n<p>Whether to train directly, or to train after 3D semantic segmentation (as stage2).</p>\n<p>That' s pretty weird. Not really intuitive.</p>\n<p>So I was wondering if anyone used 3D classification in the competition and got a good score?</p>",
  "messages": [
    {
      "id": 2007468,
      "postDate": "2022-10-28T08:42:27.370Z",
      "content": "<p>I have used a pure two stage 3D solution (3d unet + 3d classifier) . <br>\nDue to a small ensemble (1 model 4 folds) and log loss metric it did not survive shakeup though. </p>",
      "rawMarkdown": "I have used a pure two stage 3D solution (3d unet + 3d classifier) . \nDue to a small ensemble (1 model 4 folds) and log loss metric it did not survive shakeup though. ",
      "votes": 5,
      "replies": [
        {
          "id": 2007482,
          "postDate": "2022-10-28T09:04:58.363Z",
          "content": "<p>Is it pretrained 3d model? Will you share the code of your model?</p>",
          "rawMarkdown": "Is it pretrained 3d model? Will you share the code of your model?"
        },
        {
          "id": 2007503,
          "postDate": "2022-10-28T09:24:31.847Z",
          "content": "<p>Great! Waiting for your solution summary 👍</p>",
          "rawMarkdown": "Great! Waiting for your solution summary 👍"
        }
      ]
    },
    {
      "id": 2007460,
      "postDate": "2022-10-28T08:33:05.163Z",
      "content": "<p>I, as well as many others, reported that could not get a good score using 3D classification.</p>\n<p>Whether to train directly, or to train after 3D semantic segmentation (as stage2).</p>\n<p>That' s pretty weird. Not really intuitive.</p>\n<p>So I was wondering if anyone used 3D classification in the competition and got a good score?</p>",
      "rawMarkdown": "I, as well as many others, reported that could not get a good score using 3D classification.\n\nWhether to train directly, or to train after 3D semantic segmentation (as stage2).\n\nThat' s pretty weird. Not really intuitive.\n\nSo I was wondering if anyone used 3D classification in the competition and got a good score?",
      "votes": 6
    },
    {
      "id": 2007594,
      "postDate": "2022-10-28T10:49:24.380Z",
      "content": "<p>I used <a href=\"https://pytorchvideo.readthedocs.io/en/latest/api/models/x3d.html\" target=\"_blank\">https://pytorchvideo.readthedocs.io/en/latest/api/models/x3d.html</a>. </p>\n<p>I found that this model was very effective for 3D fracture classification of individual vertebra. </p>",
      "rawMarkdown": "I used https://pytorchvideo.readthedocs.io/en/latest/api/models/x3d.html. \n\nI found that this model was very effective for 3D fracture classification of individual vertebra. ",
      "votes": 4,
      "replies": [
        {
          "id": 2007662,
          "postDate": "2022-10-28T12:04:55.177Z",
          "content": "<p>Congrats and thanks for sharing!</p>",
          "rawMarkdown": "Congrats and thanks for sharing!"
        }
      ]
    },
    {
      "id": 2007934,
      "postDate": "2022-10-28T15:27:29.937Z",
      "content": "<p>I also used 3D classifiers from MONAI and they worked pretty well. Although because they were randomly initialized, they took much longer to train (&gt;100 epochs).</p>",
      "rawMarkdown": "I also used 3D classifiers from MONAI and they worked pretty well. Although because they were randomly initialized, they took much longer to train (>100 epochs).",
      "votes": 1,
      "replies": [
        {
          "id": 2007981,
          "postDate": "2022-10-28T16:13:37.893Z",
          "content": "<p>Yah I see your solution, will have a try later ;)</p>",
          "rawMarkdown": "Yah I see your solution, will have a try later ;)"
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 2007468,
      "author_name": "Selim Seferbekov",
      "author_url": "",
      "post_date": "2022-10-28T08:42:27.370000",
      "content": "<p>I have used a pure two stage 3D solution (3d unet + 3d classifier) . <br>\nDue to a small ensemble (1 model 4 folds) and log loss metric it did not survive shakeup though. </p>",
      "votes": 5,
      "replies": [
        {
          "id": 2007482,
          "author_name": "Victor Durnov",
          "author_url": "",
          "post_date": "2022-10-28T09:04:58.363000",
          "content": "<p>Is it pretrained 3d model? Will you share the code of your model?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 2007503,
          "author_name": "Qishen Ha",
          "author_url": "",
          "post_date": "2022-10-28T09:24:31.847000",
          "content": "<p>Great! Waiting for your solution summary 👍</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 2007594,
      "author_name": "Ian Pan",
      "author_url": "",
      "post_date": "2022-10-28T10:49:24.380000",
      "content": "<p>I used <a href=\"https://pytorchvideo.readthedocs.io/en/latest/api/models/x3d.html\" target=\"_blank\">https://pytorchvideo.readthedocs.io/en/latest/api/models/x3d.html</a>. </p>\n<p>I found that this model was very effective for 3D fracture classification of individual vertebra. </p>",
      "votes": 4,
      "replies": [
        {
          "id": 2007662,
          "author_name": "Qishen Ha",
          "author_url": "",
          "post_date": "2022-10-28T12:04:55.177000",
          "content": "<p>Congrats and thanks for sharing!</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 2007934,
      "author_name": "Bardia Khosravi",
      "author_url": "",
      "post_date": "2022-10-28T15:27:29.937000",
      "content": "<p>I also used 3D classifiers from MONAI and they worked pretty well. Although because they were randomly initialized, they took much longer to train (&gt;100 epochs).</p>",
      "votes": 1,
      "replies": [
        {
          "id": 2007981,
          "author_name": "Qishen Ha",
          "author_url": "",
          "post_date": "2022-10-28T16:13:37.893000",
          "content": "<p>Yah I see your solution, will have a try later ;)</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2007468": "I have used a pure two stage 3D solution (3d unet + 3d classifier) . \nDue to a small ensemble (1 model 4 folds) and log loss metric it did not survive shakeup though. ",
    "2007460": "I, as well as many others, reported that could not get a good score using 3D classification.\n\nWhether to train directly, or to train after 3D semantic segmentation (as stage2).\n\nThat' s pretty weird. Not really intuitive.\n\nSo I was wondering if anyone used 3D classification in the competition and got a good score?",
    "2007594": "I used https://pytorchvideo.readthedocs.io/en/latest/api/models/x3d.html. \n\nI found that this model was very effective for 3D fracture classification of individual vertebra. ",
    "2007934": "I also used 3D classifiers from MONAI and they worked pretty well. Although because they were randomly initialized, they took much longer to train (>100 epochs)."
  }
}