{
  "id": 117502,
  "title": "How success of our model performance in a real-world basis",
  "url": "/competitions/rsna-intracranial-hemorrhage-detection/discussion/117502",
  "author_name": "Neuron Engineer",
  "post_date": "2019-11-15T23:23:45.173000",
  "votes": 4,
  "comment_count": 0,
  "views": 0,
  "content": "<p><a href=\"/lechuck0\">@lechuck0</a> <a href=\"/philculliton\">@philculliton</a> Thanks again for this wonderful competition and congratulate all participants, we have improved ourselves a lot !</p>\n\n<p>Having worked on this project for months, besides knowledge learning, I would love to know <em>how promising there is</em> for our top solutions models applying to the real world basis.</p>\n\n<p>I calculated F1 of my baselines for each Hemorrhage Types and get around 70% - 90% range on Intraparenchymal , intraventricularm subarachnoid, and subdural cases. But only F1 of 20%-30% for epidural cases (fewest data, and most difficult?) -- </p>\n\n<p>(This F1 is before post-processing, with post-processing as log loss is 10% improved, perhaps around the same improvement for F1)</p>\n\n<p>So how about this performance, is it satisfying? Or which direction should we improve further (especially the epidural cases)? </p>\n\n<p>(CC: <a href=\"/dcstang\">@dcstang</a> perhaps you can give us great opinions on this subject again :D )</p>",
  "messages": [
    {
      "id": 674131,
      "postDate": "2019-11-15T23:23:45.173Z",
      "content": "<p><a href=\"/lechuck0\">@lechuck0</a> <a href=\"/philculliton\">@philculliton</a> Thanks again for this wonderful competition and congratulate all participants, we have improved ourselves a lot !</p>\n\n<p>Having worked on this project for months, besides knowledge learning, I would love to know <em>how promising there is</em> for our top solutions models applying to the real world basis.</p>\n\n<p>I calculated F1 of my baselines for each Hemorrhage Types and get around 70% - 90% range on Intraparenchymal , intraventricularm subarachnoid, and subdural cases. But only F1 of 20%-30% for epidural cases (fewest data, and most difficult?) -- </p>\n\n<p>(This F1 is before post-processing, with post-processing as log loss is 10% improved, perhaps around the same improvement for F1)</p>\n\n<p>So how about this performance, is it satisfying? Or which direction should we improve further (especially the epidural cases)? </p>\n\n<p>(CC: <a href=\"/dcstang\">@dcstang</a> perhaps you can give us great opinions on this subject again :D )</p>",
      "rawMarkdown": "@lechuck0 @philculliton Thanks again for this wonderful competition and congratulate all participants, we have improved ourselves a lot !\n\nHaving worked on this project for months, besides knowledge learning, I would love to know *how promising there is* for our top solutions models applying to the real world basis.\n\nI calculated F1 of my baselines for each Hemorrhage Types and get around 70% - 90% range on Intraparenchymal , intraventricularm subarachnoid, and subdural cases. But only F1 of 20%-30% for epidural cases (fewest data, and most difficult?) -- \n\n(This F1 is before post-processing, with post-processing as log loss is 10% improved, perhaps around the same improvement for F1)\n\nSo how about this performance, is it satisfying? Or which direction should we improve further (especially the epidural cases)? \n\n (CC: @dcstang perhaps you can give us great opinions on this subject again :D )",
      "votes": 4
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "674131": "@lechuck0 @philculliton Thanks again for this wonderful competition and congratulate all participants, we have improved ourselves a lot !\n\nHaving worked on this project for months, besides knowledge learning, I would love to know *how promising there is* for our top solutions models applying to the real world basis.\n\nI calculated F1 of my baselines for each Hemorrhage Types and get around 70% - 90% range on Intraparenchymal , intraventricularm subarachnoid, and subdural cases. But only F1 of 20%-30% for epidural cases (fewest data, and most difficult?) -- \n\n(This F1 is before post-processing, with post-processing as log loss is 10% improved, perhaps around the same improvement for F1)\n\nSo how about this performance, is it satisfying? Or which direction should we improve further (especially the epidural cases)? \n\n (CC: @dcstang perhaps you can give us great opinions on this subject again :D )"
  }
}