{
  "id": 65197,
  "title": "Has anyone tried lung segmentation?",
  "url": "/competitions/rsna-pneumonia-detection-challenge/discussion/65197",
  "author_name": "",
  "post_date": "2018-09-07T09:01:00.925538100Z",
  "votes": 3,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Training a model that segments the lung first and then fill the rest of the image with 0s. Would it be beneficial since only the necessary information is passed to the actual bounding-box prediction network? </p>",
  "messages": [
    {
      "id": "382880",
      "postDate": "09/07/2018 09:01:00",
      "content": "<p>Training a model that segments the lung first and then fill the rest of the image with 0s. Would it be beneficial since only the necessary information is passed to the actual bounding-box prediction network? </p>",
      "rawMarkdown": "Training a model that segments the lung first and then fill the rest of the image with 0s. Would it be beneficial since only the necessary information is passed to the actual bounding-box prediction network?",
      "votes": null
    },
    {
      "id": "382895",
      "postDate": "09/07/2018 09:23:34",
      "content": "<p>I tried some, but the gt usually extends beyond the segmentation. I am going to try and train on the clipped bboxes in the near future... </p>",
      "rawMarkdown": "I tried some, but the gt usually extends beyond the segmentation. I am going to try and train on the clipped bboxes in the near future...",
      "votes": null
    },
    {
      "id": "383119",
      "postDate": "09/07/2018 18:14:23",
      "content": "<p>I have this idea too, but there is a reason, why we with my associate reject this idea:</p>\n\n<ul>\n<li>That's difficult, cause we have picture with different scale, we have rolling picture and so on. </li>\n<li>If we split our challenge on two part - lung detection and then bounded boxes detection - then we have additional error and if we have, for instance, 0.8 score in each task, final score maybe becomes 0.8*0,8=0,64.</li>\n<li>Let's take a look on data. There is a patiant, that have correct boxes, that consist not only lung, but heart too. So, we can't consider only lung, cause boxes can overlap,  for example, heart. </li>\n<li>Finally , and this is bigger problem, if we leave only lung, then all pictures will have different size, and we will have trouble with neural net input. </li>\n</ul>\n\n<p>We can solve all this problem, but you will have this problem on your way with big confidence. Good luck) </p>",
      "rawMarkdown": "I have this idea too, but there is a reason, why we with my associate reject this idea:\n\n - That's difficult, cause we have picture with different scale, we have rolling picture and so on. \n - If we split our challenge on two part - lung detection and then bounded boxes detection - then we have additional error and if we have, for instance, 0.8 score in each task, final score maybe becomes 0.8*0,8=0,64.\n - Let's take a look on data. There is a patiant, that have correct boxes, that consist not only lung, but heart too. So, we can't consider only lung, cause boxes can overlap,  for example, heart. \n - Finally , and this is bigger problem, if we leave only lung, then all pictures will have different size, and we will have trouble with neural net input. \n \n \n\nWe can solve all this problem, but you will have this problem on your way with big confidence. Good luck)",
      "votes": null
    },
    {
      "id": "383514",
      "postDate": "09/08/2018 21:58:20",
      "content": "<p>I've started but haven't got too far with it yet.</p>",
      "rawMarkdown": "I've started but haven't got too far with it yet.",
      "votes": null
    },
    {
      "id": "383549",
      "postDate": "09/09/2018 00:31:56",
      "content": "<p>Here's a snapshot of some of my progress so far:</p>\n\n<p><a href=\"https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/65300\">https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/65300</a></p>",
      "rawMarkdown": "Here's a snapshot of some of my progress so far:\n\nhttps://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/65300",
      "votes": null
    },
    {
      "id": "383794",
      "postDate": "09/09/2018 16:18:43",
      "content": "<p>The Pathology we are looking for is overlayed on the image, although i think that we can create a dataset that would get better results, take a look:\n<a href=\"https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/64979\">https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/64979</a></p>",
      "rawMarkdown": "The Pathology we are looking for is overlayed on the image, although i think that we can create a dataset that would get better results, take a look:\nhttps://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/64979",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 382895,
      "author_name": "moshel",
      "author_url": "",
      "post_date": "09/07/2018 09:23:34",
      "content": "<p>I tried some, but the gt usually extends beyond the segmentation. I am going to try and train on the clipped bboxes in the near future... </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 383119,
      "author_name": "koza4ukdmitrij",
      "author_url": "",
      "post_date": "09/07/2018 18:14:23",
      "content": "<p>I have this idea too, but there is a reason, why we with my associate reject this idea:</p>\n\n<ul>\n<li>That's difficult, cause we have picture with different scale, we have rolling picture and so on. </li>\n<li>If we split our challenge on two part - lung detection and then bounded boxes detection - then we have additional error and if we have, for instance, 0.8 score in each task, final score maybe becomes 0.8*0,8=0,64.</li>\n<li>Let's take a look on data. There is a patiant, that have correct boxes, that consist not only lung, but heart too. So, we can't consider only lung, cause boxes can overlap,  for example, heart. </li>\n<li>Finally , and this is bigger problem, if we leave only lung, then all pictures will have different size, and we will have trouble with neural net input. </li>\n</ul>\n\n<p>We can solve all this problem, but you will have this problem on your way with big confidence. Good luck) </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 383514,
      "author_name": "puremath86",
      "author_url": "",
      "post_date": "09/08/2018 21:58:20",
      "content": "<p>I've started but haven't got too far with it yet.</p>",
      "votes": null,
      "replies": [
        {
          "id": 383549,
          "author_name": "puremath86",
          "author_url": "",
          "post_date": "09/09/2018 00:31:56",
          "content": "<p>Here's a snapshot of some of my progress so far:</p>\n\n<p><a href=\"https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/65300\">https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/65300</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 383794,
      "author_name": "sandorkonya",
      "author_url": "",
      "post_date": "09/09/2018 16:18:43",
      "content": "<p>The Pathology we are looking for is overlayed on the image, although i think that we can create a dataset that would get better results, take a look:\n<a href=\"https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/64979\">https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/64979</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "382880": "Training a model that segments the lung first and then fill the rest of the image with 0s. Would it be beneficial since only the necessary information is passed to the actual bounding-box prediction network?",
    "382895": "I tried some, but the gt usually extends beyond the segmentation. I am going to try and train on the clipped bboxes in the near future...",
    "383119": "I have this idea too, but there is a reason, why we with my associate reject this idea:\n\n - That's difficult, cause we have picture with different scale, we have rolling picture and so on. \n - If we split our challenge on two part - lung detection and then bounded boxes detection - then we have additional error and if we have, for instance, 0.8 score in each task, final score maybe becomes 0.8*0,8=0,64.\n - Let's take a look on data. There is a patiant, that have correct boxes, that consist not only lung, but heart too. So, we can't consider only lung, cause boxes can overlap,  for example, heart. \n - Finally , and this is bigger problem, if we leave only lung, then all pictures will have different size, and we will have trouble with neural net input. \n \n \n\nWe can solve all this problem, but you will have this problem on your way with big confidence. Good luck)",
    "383514": "I've started but haven't got too far with it yet.",
    "383549": "Here's a snapshot of some of my progress so far:\n\nhttps://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/65300",
    "383794": "The Pathology we are looking for is overlayed on the image, although i think that we can create a dataset that would get better results, take a look:\nhttps://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/64979"
  },
  "source": "meta"
}