{
  "id": 255280,
  "title": "How to merge nifti format files and modify them for input in DL model?",
  "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/255280",
  "author_name": "",
  "post_date": "2021-07-26T17:37:33.938806900Z",
  "votes": 1,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Please assist me.</p>\n<p>At first I converted all the dicom files to nifti format and resized the nifti files in 240<em>240</em>155.</p>\n<p>For instance, we have four modalities (T2,T1,T1w and FLAIR) for each patient. I was thinking of merging the modalities (nifti files) into one and modify it to give it as a input to trained DL models?</p>\n<p>Any idea will be highly appreciated.</p>",
  "messages": [
    {
      "id": "1400896",
      "postDate": "07/26/2021 17:37:33",
      "content": "<p>Please assist me.</p>\n<p>At first I converted all the dicom files to nifti format and resized the nifti files in 240<em>240</em>155.</p>\n<p>For instance, we have four modalities (T2,T1,T1w and FLAIR) for each patient. I was thinking of merging the modalities (nifti files) into one and modify it to give it as a input to trained DL models?</p>\n<p>Any idea will be highly appreciated.</p>",
      "rawMarkdown": "Please assist me.\n\nAt first I converted all the dicom files to nifti format and resized the nifti files in 240*240*155.\n\nFor instance, we have four modalities (T2,T1,T1w and FLAIR) for each patient. I was thinking of merging the modalities (nifti files) into one and modify it to give it as a input to trained DL models?\n\nAny idea will be highly appreciated.",
      "votes": null
    },
    {
      "id": "1401045",
      "postDate": "07/26/2021 21:08:57",
      "content": "<p>My guess is that you can read the different files and average the data from each modality, obtain the NumPy matrix and average all 4 modalities. However, you will get a 5th element Frankensteinian MRI.  Nibabel allows you to get the nifti matrices nimpy. </p>",
      "rawMarkdown": "My guess is that you can read the different files and average the data from each modality, obtain the NumPy matrix and average all 4 modalities. However, you will get a 5th element Frankensteinian MRI.  Nibabel allows you to get the nifti matrices nimpy.",
      "votes": null
    },
    {
      "id": "1401063",
      "postDate": "07/26/2021 21:46:47",
      "content": "<p>Hi, thanks for your reply.<br>\nUsing niable_concat_ image function, we can do so.<br>\nAfter the concatenate,  the dimension is like 4* 240<em>240</em>155 which means 4 channels for 4 modalities and each nifti image has dimension of 240<em>240</em>155 which implies 155 slices with dimension 240*240.</p>\n<p>Do you have any idea what pre-trained 3D deep learning model can be used for feature extraction from 3D images?</p>",
      "rawMarkdown": "Hi, thanks for your reply.\nUsing niable_concat_ image function, we can do so.\nAfter the concatenate,  the dimension is like 4* 240*240*155 which means 4 channels for 4 modalities and each nifti image has dimension of 240*240*155 which implies 155 slices with dimension 240*240.\n\nDo you have any idea what pre-trained 3D deep learning model can be used for feature extraction from 3D images?",
      "votes": null
    },
    {
      "id": "1402084",
      "postDate": "07/27/2021 20:39:39",
      "content": "<p>3D-UNet might be of help, by changing the upsampling part of the network…</p>\n<p>However, don't we have the problem, that the four modalities have different depths?</p>\n<p>I was thinking of potentially creating 4 parallel networks, which finally combine into a single output.</p>",
      "rawMarkdown": "3D-UNet might be of help, by changing the upsampling part of the network...\n\nHowever, don't we have the problem, that the four modalities have different depths?\n\nI was thinking of potentially creating 4 parallel networks, which finally combine into a single output.",
      "votes": null
    },
    {
      "id": "1402167",
      "postDate": "07/27/2021 23:45:51",
      "content": "<p>Hello, can you please explain more regarding creating four parallel network? </p>\n<p>I was considering using DL models for feature extraction but did not yet figure out how to do that.</p>",
      "rawMarkdown": "Hello, can you please explain more regarding creating four parallel network? \n\nI was considering using DL models for feature extraction but did not yet figure out how to do that.",
      "votes": null
    },
    {
      "id": "1402552",
      "postDate": "07/28/2021 09:38:02",
      "content": "<p>You can have a look at the literature for multimodal input neural networks, such as [<a href=\"https://link.springer.com/article/10.1007/s00371-021-02166-7]\" target=\"_blank\">https://link.springer.com/article/10.1007/s00371-021-02166-7]</a>.</p>\n<p>As for  DL models, how did you do your first submission?</p>\n<p>You can first try to take a pretrained model, such as an ResNet50 or EfficientNet and feed through the slices of the MRI scans. This is obviously not a winning solution, but it could give you a start in with deep learning</p>",
      "rawMarkdown": "You can have a look at the literature for multimodal input neural networks, such as [https://link.springer.com/article/10.1007/s00371-021-02166-7].\n\nAs for  DL models, how did you do your first submission?\n\nYou can first try to take a pretrained model, such as an ResNet50 or EfficientNet and feed through the slices of the MRI scans. This is obviously not a winning solution, but it could give you a start in with deep learning",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1401045,
      "author_name": "ranafago",
      "author_url": "",
      "post_date": "07/26/2021 21:08:57",
      "content": "<p>My guess is that you can read the different files and average the data from each modality, obtain the NumPy matrix and average all 4 modalities. However, you will get a 5th element Frankensteinian MRI.  Nibabel allows you to get the nifti matrices nimpy. </p>",
      "votes": null,
      "replies": [
        {
          "id": 1401063,
          "author_name": "wfarzana",
          "author_url": "",
          "post_date": "07/26/2021 21:46:47",
          "content": "<p>Hi, thanks for your reply.<br>\nUsing niable_concat_ image function, we can do so.<br>\nAfter the concatenate,  the dimension is like 4* 240<em>240</em>155 which means 4 channels for 4 modalities and each nifti image has dimension of 240<em>240</em>155 which implies 155 slices with dimension 240*240.</p>\n<p>Do you have any idea what pre-trained 3D deep learning model can be used for feature extraction from 3D images?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1402084,
          "author_name": "fanconic",
          "author_url": "",
          "post_date": "07/27/2021 20:39:39",
          "content": "<p>3D-UNet might be of help, by changing the upsampling part of the network…</p>\n<p>However, don't we have the problem, that the four modalities have different depths?</p>\n<p>I was thinking of potentially creating 4 parallel networks, which finally combine into a single output.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1402167,
          "author_name": "wfarzana",
          "author_url": "",
          "post_date": "07/27/2021 23:45:51",
          "content": "<p>Hello, can you please explain more regarding creating four parallel network? </p>\n<p>I was considering using DL models for feature extraction but did not yet figure out how to do that.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1402552,
          "author_name": "fanconic",
          "author_url": "",
          "post_date": "07/28/2021 09:38:02",
          "content": "<p>You can have a look at the literature for multimodal input neural networks, such as [<a href=\"https://link.springer.com/article/10.1007/s00371-021-02166-7]\" target=\"_blank\">https://link.springer.com/article/10.1007/s00371-021-02166-7]</a>.</p>\n<p>As for  DL models, how did you do your first submission?</p>\n<p>You can first try to take a pretrained model, such as an ResNet50 or EfficientNet and feed through the slices of the MRI scans. This is obviously not a winning solution, but it could give you a start in with deep learning</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1400896": "Please assist me.\n\nAt first I converted all the dicom files to nifti format and resized the nifti files in 240*240*155.\n\nFor instance, we have four modalities (T2,T1,T1w and FLAIR) for each patient. I was thinking of merging the modalities (nifti files) into one and modify it to give it as a input to trained DL models?\n\nAny idea will be highly appreciated.",
    "1401045": "My guess is that you can read the different files and average the data from each modality, obtain the NumPy matrix and average all 4 modalities. However, you will get a 5th element Frankensteinian MRI.  Nibabel allows you to get the nifti matrices nimpy.",
    "1401063": "Hi, thanks for your reply.\nUsing niable_concat_ image function, we can do so.\nAfter the concatenate,  the dimension is like 4* 240*240*155 which means 4 channels for 4 modalities and each nifti image has dimension of 240*240*155 which implies 155 slices with dimension 240*240.\n\nDo you have any idea what pre-trained 3D deep learning model can be used for feature extraction from 3D images?",
    "1402084": "3D-UNet might be of help, by changing the upsampling part of the network...\n\nHowever, don't we have the problem, that the four modalities have different depths?\n\nI was thinking of potentially creating 4 parallel networks, which finally combine into a single output.",
    "1402167": "Hello, can you please explain more regarding creating four parallel network? \n\nI was considering using DL models for feature extraction but did not yet figure out how to do that.",
    "1402552": "You can have a look at the literature for multimodal input neural networks, such as [https://link.springer.com/article/10.1007/s00371-021-02166-7].\n\nAs for  DL models, how did you do your first submission?\n\nYou can first try to take a pretrained model, such as an ResNet50 or EfficientNet and feed through the slices of the MRI scans. This is obviously not a winning solution, but it could give you a start in with deep learning"
  },
  "source": "meta"
}