{
  "id": 361861,
  "title": "Any better ways than weighted average aggregation?",
  "url": "/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/361861",
  "author_name": "JunHyeonKwon",
  "post_date": "2022-10-24T07:59:36.044000",
  "votes": 6,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Since training with a 3D stack of axial slices is costly, I guess many teams prefer to detect fractures from each axial slice.   <br>\nIn this case, I saw many discussions and codes, namely <a href=\"https://www.kaggle.com/code/vslaykovsky/train-pytorch-effnetv2-baseline-cv-0-49\" target=\"_blank\">this one</a>, which uses the weighted average to aggregate predictions per slices into one prediction.</p>\n<p>The weighted average here is:<br>\n$$ {(\\text{fracture probability})*(\\text{probability for a certain spine to be in img}) \\over sum(\\text{probability for a certain spine to be in img})} $$  </p>\n<p>But intuitively, I think maximum aggregation makes sense because if one slice out of 20 C1 slices has a fracture, p_C1 should be 1. But of course, this method returns 10+ score since almost every prediction is near 1</p>\n<p>It's already near the deadline but can you guys share any better methods?  <br>\nThank you.</p>",
  "messages": [
    {
      "id": 2001651,
      "postDate": "2022-10-24T07:59:36.043Z",
      "content": "<p>Since training with a 3D stack of axial slices is costly, I guess many teams prefer to detect fractures from each axial slice.   <br>\nIn this case, I saw many discussions and codes, namely <a href=\"https://www.kaggle.com/code/vslaykovsky/train-pytorch-effnetv2-baseline-cv-0-49\" target=\"_blank\">this one</a>, which uses the weighted average to aggregate predictions per slices into one prediction.</p>\n<p>The weighted average here is:<br>\n$$ {(\\text{fracture probability})*(\\text{probability for a certain spine to be in img}) \\over sum(\\text{probability for a certain spine to be in img})} $$  </p>\n<p>But intuitively, I think maximum aggregation makes sense because if one slice out of 20 C1 slices has a fracture, p_C1 should be 1. But of course, this method returns 10+ score since almost every prediction is near 1</p>\n<p>It's already near the deadline but can you guys share any better methods?  <br>\nThank you.</p>",
      "rawMarkdown": "Since training with a 3D stack of axial slices is costly, I guess many teams prefer to detect fractures from each axial slice.   \nIn this case, I saw many discussions and codes, namely [this one](https://www.kaggle.com/code/vslaykovsky/train-pytorch-effnetv2-baseline-cv-0-49), which uses the weighted average to aggregate predictions per slices into one prediction.\n\nThe weighted average here is:\n$$ {(\\text{fracture probability})*(\\text{probability for a certain spine to be in img}) \\over sum(\\text{probability for a certain spine to be in img})} $$  \n\nBut intuitively, I think maximum aggregation makes sense because if one slice out of 20 C1 slices has a fracture, p_C1 should be 1. But of course, this method returns 10+ score since almost every prediction is near 1\n\nIt's already near the deadline but can you guys share any better methods?  \nThank you.",
      "votes": 6
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2001651": "Since training with a 3D stack of axial slices is costly, I guess many teams prefer to detect fractures from each axial slice.   \nIn this case, I saw many discussions and codes, namely [this one](https://www.kaggle.com/code/vslaykovsky/train-pytorch-effnetv2-baseline-cv-0-49), which uses the weighted average to aggregate predictions per slices into one prediction.\n\nThe weighted average here is:\n$$ {(\\text{fracture probability})*(\\text{probability for a certain spine to be in img}) \\over sum(\\text{probability for a certain spine to be in img})} $$  \n\nBut intuitively, I think maximum aggregation makes sense because if one slice out of 20 C1 slices has a fracture, p_C1 should be 1. But of course, this method returns 10+ score since almost every prediction is near 1\n\nIt's already near the deadline but can you guys share any better methods?  \nThank you."
  }
}