{
  "id": 345930,
  "title": "How do we proceed?? with so much missing data?",
  "url": "/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/345930",
  "author_name": "Krish78",
  "post_date": "2022-08-17T06:47:05.435000",
  "votes": -1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>2019 - case results are available, <br>\n787 CT 3D images are available, <br>\n87 segmentations<br>\n235 bounding boxes.<br>\nat the best we have 787 data * 7  ie if there is fracture in Cz of patient X,<br>\nSo are we suppose to CNN over the entire 3D image.-- and brood force fit it with results 1 3D image to 7 results??. Most Probably thats not going to give us good results.</p>\n<p>What do we do with bounding boxes and segmentations?</p>\n<p>if the bounding box area is the ROI of C1,C2…etc  i thought it could be useful but thats not the case. so don't know how we could use that info?<br>\nSo, Correct me if i am wrong, some kind of ROI for each of C1,C2….C7 might be needed. else we are processing too much of non-sense.?  how are we suppose go about this, if not manually?? for creating BBs?<br>\nThe segementation is available for two few images ie 87 of 787 images ie 11%, so are we suppose to do the segementation for the rest, use some kind of transfer learning + 87 and lets we do that have it.</p>\n<p>then again do we use to brood force  3D image to 7 results per image, or get ROI-BB of C1-7 and then try to fit </p>",
  "messages": [
    {
      "id": 1903202,
      "postDate": "2022-08-17T07:29:05.337Z",
      "content": "<p>I am a lot curious, Where do you get the idea that only 787 CT 3D Images are available, every case have 3D…</p>",
      "rawMarkdown": "I am a lot curious, Where do you get the idea that only 787 CT 3D Images are available, every case have 3D...",
      "votes": 2,
      "replies": [
        {
          "id": 1903292,
          "postDate": "2022-08-17T09:25:38.550Z",
          "content": "<p>Train CSV has 2019 unique study uid, where as in the downloaded data set had 787 folders in the train images folder, each folder corresponds to one Case or something went wrong in my downloading or extraction.</p>",
          "rawMarkdown": "Train CSV has 2019 unique study uid, where as in the downloaded data set had 787 folders in the train images folder, each folder corresponds to one Case or something went wrong in my downloading or extraction."
        },
        {
          "id": 1903293,
          "postDate": "2022-08-17T09:27:32.837Z",
          "content": "<p>Your downloaded data is corrupted then, because I have 2019 sub-folders in the folder 'train_images', where each subfolder corresponds to one case having 100's of dicom files</p>",
          "rawMarkdown": "Your downloaded data is corrupted then, because I have 2019 sub-folders in the folder 'train_images', where each subfolder corresponds to one case having 100's of dicom files",
          "votes": 1
        }
      ]
    },
    {
      "id": 1903160,
      "postDate": "2022-08-17T06:47:05.437Z",
      "content": "<p>2019 - case results are available, <br>\n787 CT 3D images are available, <br>\n87 segmentations<br>\n235 bounding boxes.<br>\nat the best we have 787 data * 7  ie if there is fracture in Cz of patient X,<br>\nSo are we suppose to CNN over the entire 3D image.-- and brood force fit it with results 1 3D image to 7 results??. Most Probably thats not going to give us good results.</p>\n<p>What do we do with bounding boxes and segmentations?</p>\n<p>if the bounding box area is the ROI of C1,C2…etc  i thought it could be useful but thats not the case. so don't know how we could use that info?<br>\nSo, Correct me if i am wrong, some kind of ROI for each of C1,C2….C7 might be needed. else we are processing too much of non-sense.?  how are we suppose go about this, if not manually?? for creating BBs?<br>\nThe segementation is available for two few images ie 87 of 787 images ie 11%, so are we suppose to do the segementation for the rest, use some kind of transfer learning + 87 and lets we do that have it.</p>\n<p>then again do we use to brood force  3D image to 7 results per image, or get ROI-BB of C1-7 and then try to fit </p>",
      "rawMarkdown": "2019 - case results are available, \n787 CT 3D images are available, \n87 segmentations\n235 bounding boxes.\nat the best we have 787 data * 7  ie if there is fracture in Cz of patient X,\nSo are we suppose to CNN over the entire 3D image.-- and brood force fit it with results 1 3D image to 7 results??. Most Probably thats not going to give us good results.\n\nWhat do we do with bounding boxes and segmentations?\n\nif the bounding box area is the ROI of C1,C2...etc  i thought it could be useful but thats not the case. so don't know how we could use that info?\nSo, Correct me if i am wrong, some kind of ROI for each of C1,C2....C7 might be needed. else we are processing too much of non-sense.?  how are we suppose go about this, if not manually?? for creating BBs?\nThe segementation is available for two few images ie 87 of 787 images ie 11%, so are we suppose to do the segementation for the rest, use some kind of transfer learning + 87 and lets we do that have it.\n\nthen again do we use to brood force  3D image to 7 results per image, or get ROI-BB of C1-7 and then try to fit ",
      "votes": -1
    }
  ],
  "comments": [
    {
      "id": 1903202,
      "author_name": "Harshit Sheoran",
      "author_url": "",
      "post_date": "2022-08-17T07:29:05.337000",
      "content": "<p>I am a lot curious, Where do you get the idea that only 787 CT 3D Images are available, every case have 3D…</p>",
      "votes": 2,
      "replies": [
        {
          "id": 1903292,
          "author_name": "Krish78",
          "author_url": "",
          "post_date": "2022-08-17T09:25:38.550000",
          "content": "<p>Train CSV has 2019 unique study uid, where as in the downloaded data set had 787 folders in the train images folder, each folder corresponds to one Case or something went wrong in my downloading or extraction.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1903293,
          "author_name": "Harshit Sheoran",
          "author_url": "",
          "post_date": "2022-08-17T09:27:32.837000",
          "content": "<p>Your downloaded data is corrupted then, because I have 2019 sub-folders in the folder 'train_images', where each subfolder corresponds to one case having 100's of dicom files</p>",
          "votes": 1,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1903202": "I am a lot curious, Where do you get the idea that only 787 CT 3D Images are available, every case have 3D...",
    "1903160": "2019 - case results are available, \n787 CT 3D images are available, \n87 segmentations\n235 bounding boxes.\nat the best we have 787 data * 7  ie if there is fracture in Cz of patient X,\nSo are we suppose to CNN over the entire 3D image.-- and brood force fit it with results 1 3D image to 7 results??. Most Probably thats not going to give us good results.\n\nWhat do we do with bounding boxes and segmentations?\n\nif the bounding box area is the ROI of C1,C2...etc  i thought it could be useful but thats not the case. so don't know how we could use that info?\nSo, Correct me if i am wrong, some kind of ROI for each of C1,C2....C7 might be needed. else we are processing too much of non-sense.?  how are we suppose go about this, if not manually?? for creating BBs?\nThe segementation is available for two few images ie 87 of 787 images ie 11%, so are we suppose to do the segementation for the rest, use some kind of transfer learning + 87 and lets we do that have it.\n\nthen again do we use to brood force  3D image to 7 results per image, or get ROI-BB of C1-7 and then try to fit "
  }
}