{
  "id": 276898,
  "title": "Paper: Transformer for 3D Image Segmentation of MRI scans",
  "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/276898",
  "author_name": "Max Baugh",
  "post_date": "2021-10-06T21:43:11.798000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I found this paper: <a href=\"https://arxiv.org/pdf/2103.10504.pdf\" target=\"_blank\">https://arxiv.org/pdf/2103.10504.pdf</a><br>\nAnd the affiliated PyTorch code: <a href=\"https://github.com/Project-MONAI/MONAI/blob/1651f1b003b0ffae8b615d191952ad65ad091277/monai/networks/nets/unetr.py\" target=\"_blank\">https://github.com/Project-MONAI/MONAI/blob/1651f1b003b0ffae8b615d191952ad65ad091277/monai/networks/nets/unetr.py</a> </p>\n<p>The task they are solving is different but closely related: they are doing full 3D image segmentation, classifying each voxel according to whatever class it belongs to.  A version of this could be used to localize the tumor, though admittedly I'm not sure <em>how</em> to do that.  It seems plausible that the transformer half of their model (it's an encoder-decoder model, using a transformer as the encoder) can be broken off &amp; used as a classifier for this problem directly.   </p>",
  "messages": [
    {
      "id": 1536634,
      "postDate": "2021-10-06T21:43:11.800Z",
      "content": "<p>I found this paper: <a href=\"https://arxiv.org/pdf/2103.10504.pdf\" target=\"_blank\">https://arxiv.org/pdf/2103.10504.pdf</a><br>\nAnd the affiliated PyTorch code: <a href=\"https://github.com/Project-MONAI/MONAI/blob/1651f1b003b0ffae8b615d191952ad65ad091277/monai/networks/nets/unetr.py\" target=\"_blank\">https://github.com/Project-MONAI/MONAI/blob/1651f1b003b0ffae8b615d191952ad65ad091277/monai/networks/nets/unetr.py</a> </p>\n<p>The task they are solving is different but closely related: they are doing full 3D image segmentation, classifying each voxel according to whatever class it belongs to.  A version of this could be used to localize the tumor, though admittedly I'm not sure <em>how</em> to do that.  It seems plausible that the transformer half of their model (it's an encoder-decoder model, using a transformer as the encoder) can be broken off &amp; used as a classifier for this problem directly.   </p>",
      "rawMarkdown": "I found this paper: https://arxiv.org/pdf/2103.10504.pdf\nAnd the affiliated PyTorch code: https://github.com/Project-MONAI/MONAI/blob/1651f1b003b0ffae8b615d191952ad65ad091277/monai/networks/nets/unetr.py \n\nThe task they are solving is different but closely related: they are doing full 3D image segmentation, classifying each voxel according to whatever class it belongs to.  A version of this could be used to localize the tumor, though admittedly I'm not sure *how* to do that.  It seems plausible that the transformer half of their model (it's an encoder-decoder model, using a transformer as the encoder) can be broken off & used as a classifier for this problem directly.   ",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1536634": "I found this paper: https://arxiv.org/pdf/2103.10504.pdf\nAnd the affiliated PyTorch code: https://github.com/Project-MONAI/MONAI/blob/1651f1b003b0ffae8b615d191952ad65ad091277/monai/networks/nets/unetr.py \n\nThe task they are solving is different but closely related: they are doing full 3D image segmentation, classifying each voxel according to whatever class it belongs to.  A version of this could be used to localize the tumor, though admittedly I'm not sure *how* to do that.  It seems plausible that the transformer half of their model (it's an encoder-decoder model, using a transformer as the encoder) can be broken off & used as a classifier for this problem directly.   "
  }
}