{
  "id": 356477,
  "title": "Just a theory..",
  "url": "/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/356477",
  "author_name": "Harshit Sheoran",
  "post_date": "2022-09-30T17:49:34.747000",
  "votes": 5,
  "comment_count": 1,
  "views": 0,
  "content": "<p>So, I was thinking… what if we just take all the slices which are in the allegedly fractured vetebrae and run a classifier on that…. but Harshit, that would be very controversial for the model, on one side we are giving it some slices correctly but not every slice in every vertebrae is fractured, how do you solve that?</p>\n<p>My theory is, if we do take all the slices from a vertebrae and label them fractured although a lot of them are not fractured… we can then run it on the validation set and choose slices where the model is highly confident and probably correct, hence, probably able to generalize…</p>\n<p>With these chosen slices we can make a new dataset which will probably be a lot more stable.</p>\n<p>this is just a theory, nothing to take serious, it might or might not work.</p>",
  "messages": [
    {
      "id": 1964342,
      "postDate": "2022-09-30T17:49:34.747Z",
      "content": "<p>So, I was thinking… what if we just take all the slices which are in the allegedly fractured vetebrae and run a classifier on that…. but Harshit, that would be very controversial for the model, on one side we are giving it some slices correctly but not every slice in every vertebrae is fractured, how do you solve that?</p>\n<p>My theory is, if we do take all the slices from a vertebrae and label them fractured although a lot of them are not fractured… we can then run it on the validation set and choose slices where the model is highly confident and probably correct, hence, probably able to generalize…</p>\n<p>With these chosen slices we can make a new dataset which will probably be a lot more stable.</p>\n<p>this is just a theory, nothing to take serious, it might or might not work.</p>",
      "rawMarkdown": "So, I was thinking... what if we just take all the slices which are in the allegedly fractured vetebrae and run a classifier on that.... but Harshit, that would be very controversial for the model, on one side we are giving it some slices correctly but not every slice in every vertebrae is fractured, how do you solve that?\n\nMy theory is, if we do take all the slices from a vertebrae and label them fractured although a lot of them are not fractured... we can then run it on the validation set and choose slices where the model is highly confident and probably correct, hence, probably able to generalize...\n\nWith these chosen slices we can make a new dataset which will probably be a lot more stable.\n\nthis is just a theory, nothing to take serious, it might or might not work.",
      "votes": 5
    },
    {
      "id": 1979802,
      "postDate": "2022-10-09T18:04:08.747Z",
      "content": "<p>This might be good idea, afaik all the images that have bounding boxes have some kind of damage in there that was marked by expert team, so that can be used for getting valid images.</p>",
      "rawMarkdown": "This might be good idea, afaik all the images that have bounding boxes have some kind of damage in there that was marked by expert team, so that can be used for getting valid images."
    }
  ],
  "comments": [
    {
      "id": 1979802,
      "author_name": "Paweł Chłąd",
      "author_url": "",
      "post_date": "2022-10-09T18:04:08.747000",
      "content": "<p>This might be good idea, afaik all the images that have bounding boxes have some kind of damage in there that was marked by expert team, so that can be used for getting valid images.</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1964342": "So, I was thinking... what if we just take all the slices which are in the allegedly fractured vetebrae and run a classifier on that.... but Harshit, that would be very controversial for the model, on one side we are giving it some slices correctly but not every slice in every vertebrae is fractured, how do you solve that?\n\nMy theory is, if we do take all the slices from a vertebrae and label them fractured although a lot of them are not fractured... we can then run it on the validation set and choose slices where the model is highly confident and probably correct, hence, probably able to generalize...\n\nWith these chosen slices we can make a new dataset which will probably be a lot more stable.\n\nthis is just a theory, nothing to take serious, it might or might not work.",
    "1979802": "This might be good idea, afaik all the images that have bounding boxes have some kind of damage in there that was marked by expert team, so that can be used for getting valid images."
  }
}