{
  "id": 527010,
  "title": "Label for segmentation and Classification Models After Segmentation",
  "url": "/competitions/rsna-2024-lumbar-spine-degenerative-classification/discussion/527010",
  "author_name": "",
  "post_date": "2024-08-09T14:11:13.922795200Z",
  "votes": 1,
  "comment_count": 5,
  "views": 0,
  "content": "<p>I have two questions:</p>\n<p>1)\"The labels' positions are only given as x and y center coordinates. Do you create masks by, for example, enclosing them in a 10-unit box, or do you manually annotate each image?\"</p>\n<p>2)\"After segmentation, the resulting images will be fed into a model again. Is it better to use one model or three models? For instance, after segmenting the spinal canal stenosis at L1-L2, should we use separate models to determine if it is moderate or severe for spinal canal stenosis, neural foraminal narrowing, and subarticular stenosis, or is a single classification model sufficient?\"</p>",
  "messages": [
    {
      "id": "2954257",
      "postDate": "08/09/2024 14:11:13",
      "content": "<p>I have two questions:</p>\n<p>1)\"The labels' positions are only given as x and y center coordinates. Do you create masks by, for example, enclosing them in a 10-unit box, or do you manually annotate each image?\"</p>\n<p>2)\"After segmentation, the resulting images will be fed into a model again. Is it better to use one model or three models? For instance, after segmenting the spinal canal stenosis at L1-L2, should we use separate models to determine if it is moderate or severe for spinal canal stenosis, neural foraminal narrowing, and subarticular stenosis, or is a single classification model sufficient?\"</p>",
      "rawMarkdown": "I have two questions:\n\n1)\"The labels' positions are only given as x and y center coordinates. Do you create masks by, for example, enclosing them in a 10-unit box, or do you manually annotate each image?\"\n\n2)\"After segmentation, the resulting images will be fed into a model again. Is it better to use one model or three models? For instance, after segmenting the spinal canal stenosis at L1-L2, should we use separate models to determine if it is moderate or severe for spinal canal stenosis, neural foraminal narrowing, and subarticular stenosis, or is a single classification model sufficient?\"",
      "votes": null
    },
    {
      "id": "2954735",
      "postDate": "08/10/2024 04:55:45",
      "content": "<p>For the first question, I just guessed that the size of the bounding boxes is between 20~40 mm.<br>\nFor the second question, I actually tried 25 separate models ([spinal canal stenosis, neural foraminal narrowing, subarticular stenosis] x [L1/L2, …, L5/S1]). However, the performance dropped significantly. I also tried using one model with 25 classification heads, which performed well. But I haven't tried using three models yet. I think too many models may cause overfitting, so maybe three models could perform well.</p>",
      "rawMarkdown": "For the first question, I just guessed that the size of the bounding boxes is between 20~40 mm.\nFor the second question, I actually tried 25 separate models ([spinal canal stenosis, neural foraminal narrowing, subarticular stenosis] x [L1/L2, ..., L5/S1]). However, the performance dropped significantly. I also tried using one model with 25 classification heads, which performed well. But I haven't tried using three models yet. I think too many models may cause overfitting, so maybe three models could perform well.",
      "votes": null
    },
    {
      "id": "2954941",
      "postDate": "08/10/2024 11:58:47",
      "content": "<p>There are no absolutely correct answers for both questions.</p>\n<ol>\n<li>Previous competitions and literature used various segmentation approaches. box size can be hyperparameter that you can tune.</li>\n<li>3-separate models will have fewer parameters to train, may converge faster, but can overfit.<br>\n1-single model will have more parameters to train,but will take longer to converge and may still underfit.<br>\nAgain, you are the best judge for your choices.</li>\n</ol>",
      "rawMarkdown": "There are no absolutely correct answers for both questions.\n1. Previous competitions and literature used various segmentation approaches. box size can be hyperparameter that you can tune.\n2. 3-separate models will have fewer parameters to train, may converge faster, but can overfit.\n1-single model will have more parameters to train,but will take longer to converge and may still underfit.\nAgain, you are the best judge for your choices.",
      "votes": null
    },
    {
      "id": "2954946",
      "postDate": "08/10/2024 12:20:21",
      "content": "<p>You can refer discussion <a href=\"https://www.kaggle.com/competitions/rsna-2024-lumbar-spine-degenerative-classification/discussion/519628\" target=\"_blank\">here</a> by leaderboard topper in this competition.</p>\n<p>you can find many high quality insights in the discussion.</p>",
      "rawMarkdown": "You can refer discussion [here](https://www.kaggle.com/competitions/rsna-2024-lumbar-spine-degenerative-classification/discussion/519628) by leaderboard topper in this competition.\n\nyou can find many high quality insights in the discussion.",
      "votes": null
    },
    {
      "id": "2956174",
      "postDate": "08/11/2024 20:46:06",
      "content": "<p>Ok . I have one more question.Should the model receive 3D input or 2D? Can you also suggest a source for this? I couldn't find much for this type of dicom files.</p>",
      "rawMarkdown": "Ok . I have one more question.Should the model receive 3D input or 2D? Can you also suggest a source for this? I couldn't find much for this type of dicom files.",
      "votes": null
    },
    {
      "id": "2958974",
      "postDate": "08/14/2024 14:52:50",
      "content": "<p>I haven't tried 3D input yet. I've only tried 2D and 2.5D input. 2.5D is like stacking different slices on RGB dimension. However, 2.5D doesn't show much improvement compared to 2D in my case. Still, it might enable the model to distinguish between left subarticular stenosis and right subarticular stenosis.</p>\n<p>2.5D: <a href=\"https://www.kaggle.com/competitions/uw-madison-gi-tract-image-segmentation/discussion/322549\" target=\"_blank\">https://www.kaggle.com/competitions/uw-madison-gi-tract-image-segmentation/discussion/322549</a></p>",
      "rawMarkdown": "I haven't tried 3D input yet. I've only tried 2D and 2.5D input. 2.5D is like stacking different slices on RGB dimension. However, 2.5D doesn't show much improvement compared to 2D in my case. Still, it might enable the model to distinguish between left subarticular stenosis and right subarticular stenosis.\n\n2.5D: https://www.kaggle.com/competitions/uw-madison-gi-tract-image-segmentation/discussion/322549",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2954735,
      "author_name": "followjohn",
      "author_url": "",
      "post_date": "08/10/2024 04:55:45",
      "content": "<p>For the first question, I just guessed that the size of the bounding boxes is between 20~40 mm.<br>\nFor the second question, I actually tried 25 separate models ([spinal canal stenosis, neural foraminal narrowing, subarticular stenosis] x [L1/L2, …, L5/S1]). However, the performance dropped significantly. I also tried using one model with 25 classification heads, which performed well. But I haven't tried using three models yet. I think too many models may cause overfitting, so maybe three models could perform well.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2956174,
          "author_name": "ahmetkoraysonal",
          "author_url": "",
          "post_date": "08/11/2024 20:46:06",
          "content": "<p>Ok . I have one more question.Should the model receive 3D input or 2D? Can you also suggest a source for this? I couldn't find much for this type of dicom files.</p>",
          "votes": null,
          "replies": [
            {
              "id": 2958974,
              "author_name": "followjohn",
              "author_url": "",
              "post_date": "08/14/2024 14:52:50",
              "content": "<p>I haven't tried 3D input yet. I've only tried 2D and 2.5D input. 2.5D is like stacking different slices on RGB dimension. However, 2.5D doesn't show much improvement compared to 2D in my case. Still, it might enable the model to distinguish between left subarticular stenosis and right subarticular stenosis.</p>\n<p>2.5D: <a href=\"https://www.kaggle.com/competitions/uw-madison-gi-tract-image-segmentation/discussion/322549\" target=\"_blank\">https://www.kaggle.com/competitions/uw-madison-gi-tract-image-segmentation/discussion/322549</a></p>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 2954941,
      "author_name": "rohitchaudhari25",
      "author_url": "",
      "post_date": "08/10/2024 11:58:47",
      "content": "<p>There are no absolutely correct answers for both questions.</p>\n<ol>\n<li>Previous competitions and literature used various segmentation approaches. box size can be hyperparameter that you can tune.</li>\n<li>3-separate models will have fewer parameters to train, may converge faster, but can overfit.<br>\n1-single model will have more parameters to train,but will take longer to converge and may still underfit.<br>\nAgain, you are the best judge for your choices.</li>\n</ol>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2954946,
      "author_name": "rohitchaudhari25",
      "author_url": "",
      "post_date": "08/10/2024 12:20:21",
      "content": "<p>You can refer discussion <a href=\"https://www.kaggle.com/competitions/rsna-2024-lumbar-spine-degenerative-classification/discussion/519628\" target=\"_blank\">here</a> by leaderboard topper in this competition.</p>\n<p>you can find many high quality insights in the discussion.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2954257": "I have two questions:\n\n1)\"The labels' positions are only given as x and y center coordinates. Do you create masks by, for example, enclosing them in a 10-unit box, or do you manually annotate each image?\"\n\n2)\"After segmentation, the resulting images will be fed into a model again. Is it better to use one model or three models? For instance, after segmenting the spinal canal stenosis at L1-L2, should we use separate models to determine if it is moderate or severe for spinal canal stenosis, neural foraminal narrowing, and subarticular stenosis, or is a single classification model sufficient?\"",
    "2954735": "For the first question, I just guessed that the size of the bounding boxes is between 20~40 mm.\nFor the second question, I actually tried 25 separate models ([spinal canal stenosis, neural foraminal narrowing, subarticular stenosis] x [L1/L2, ..., L5/S1]). However, the performance dropped significantly. I also tried using one model with 25 classification heads, which performed well. But I haven't tried using three models yet. I think too many models may cause overfitting, so maybe three models could perform well.",
    "2954941": "There are no absolutely correct answers for both questions.\n1. Previous competitions and literature used various segmentation approaches. box size can be hyperparameter that you can tune.\n2. 3-separate models will have fewer parameters to train, may converge faster, but can overfit.\n1-single model will have more parameters to train,but will take longer to converge and may still underfit.\nAgain, you are the best judge for your choices.",
    "2954946": "You can refer discussion [here](https://www.kaggle.com/competitions/rsna-2024-lumbar-spine-degenerative-classification/discussion/519628) by leaderboard topper in this competition.\n\nyou can find many high quality insights in the discussion.",
    "2956174": "Ok . I have one more question.Should the model receive 3D input or 2D? Can you also suggest a source for this? I couldn't find much for this type of dicom files.",
    "2958974": "I haven't tried 3D input yet. I've only tried 2D and 2.5D input. 2.5D is like stacking different slices on RGB dimension. However, 2.5D doesn't show much improvement compared to 2D in my case. Still, it might enable the model to distinguish between left subarticular stenosis and right subarticular stenosis.\n\n2.5D: https://www.kaggle.com/competitions/uw-madison-gi-tract-image-segmentation/discussion/322549"
  },
  "source": "meta"
}