{
  "id": 271799,
  "title": "not 3dconv，should 2dconv",
  "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/271799",
  "author_name": "",
  "post_date": "2021-09-12T17:02:12.000394300Z",
  "votes": 8,
  "comment_count": 8,
  "views": 0,
  "content": "<p>Because of the different scanning distances, i think 3dconv is not suitable.so use conv2d for voxel</p>\n<p>i had check the data of scan distance ,it not same .</p>\n<p>0   data/heart/train/00999/T2w/Image-1.dcm  tmpno -107.229 -135.155 -60.1161<br>\n1   data/heart/train/00999/T2w/Image-2.dcm  tmpno -107.229 -135.155 -53.6161<br>\n2   data/heart/train/00999/T2w/Image-3.dcm  tmpno -107.229 -135.155 -47.1161<br>\n3   data/heart/train/00999/T2w/Image-4.dcm  tmpno -107.229 -135.155 -40.6161<br>\n4   data/heart/train/00999/T2w/Image-5.dcm  tmpno -107.229 -135.155 -34.1161<br>\n5   data/heart/train/00999/T2w/Image-6.dcm  tmpno -107.229 -135.155 -27.6161<br>\n6   data/heart/train/00999/T2w/Image-7.dcm  tmpno -107.229 -135.155 -21.1161<br>\n7   data/heart/train/00999/T2w/Image-8.dcm  tmpno -107.229 -135.155 -14.6161<br>\n8   data/heart/train/00999/T2w/Image-9.dcm  tmpno -107.229 -135.155  -8.1161</p>\n<p>distance 7</p>\n<p>0   data/heart/train/00691/FLAIR/Image-1.dcm  tmpno  75.6942 -163.296  123.048<br>\n1   data/heart/train/00691/FLAIR/Image-2.dcm  tmpno  74.5975 -163.237  122.987<br>\n2   data/heart/train/00691/FLAIR/Image-3.dcm  tmpno  73.5008 -163.178  122.926<br>\n3   data/heart/train/00691/FLAIR/Image-4.dcm  tmpno  72.4041 -163.118  122.865<br>\n4   data/heart/train/00691/FLAIR/Image-5.dcm  tmpno  71.3074 -163.059  122.804<br>\n5   data/heart/train/00691/FLAIR/Image-6.dcm  tmpno  70.2107 -163.000  122.742<br>\n6   data/heart/train/00691/FLAIR/Image-7.dcm  tmpno  69.1140 -162.940  122.681<br>\n7   data/heart/train/00691/FLAIR/Image-8.dcm  tmpno  68.0173 -162.881  122.620<br>\n8   data/heart/train/00691/FLAIR/Image-9.dcm  tmpno  66.9206 -162.822  122.559<br>\n9  data/heart/train/00691/FLAIR/Image-10.dcm  tmpno  65.8239 -162.762  122.498:</p>\n<p>distance 0.06</p>\n<p>0   data/heart/train/00691/T1w/Image-1.dcm  tmpno -129.382 -155.674 -104.728<br>\n1   data/heart/train/00691/T1w/Image-2.dcm  tmpno -129.418 -155.674 -104.229<br>\n2   data/heart/train/00691/T1w/Image-3.dcm  tmpno -129.453 -155.674 -103.730<br>\n3   data/heart/train/00691/T1w/Image-4.dcm  tmpno -129.489 -155.674 -103.231<br>\n4   data/heart/train/00691/T1w/Image-5.dcm  tmpno -129.524 -155.674 -102.733<br>\n5   data/heart/train/00691/T1w/Image-6.dcm  tmpno -129.560 -155.674 -102.234<br>\n6   data/heart/train/00691/T1w/Image-7.dcm  tmpno -129.595 -155.674 -101.735<br>\n7   data/heart/train/00691/T1w/Image-8.dcm  tmpno -129.631 -155.674 -101.236<br>\n8   data/heart/train/00691/T1w/Image-9.dcm  tmpno -129.666 -155.674 -100.738<br>\n9  data/heart/train/00691/T1w/Image-10.dcm  tmpno -129.702 -155.674 -100.239:</p>\n<p>distance 0.5</p>",
  "messages": [
    {
      "id": "1510710",
      "postDate": "09/12/2021 17:02:12",
      "content": "<p>Because of the different scanning distances, i think 3dconv is not suitable.so use conv2d for voxel</p>\n<p>i had check the data of scan distance ,it not same .</p>\n<p>0   data/heart/train/00999/T2w/Image-1.dcm  tmpno -107.229 -135.155 -60.1161<br>\n1   data/heart/train/00999/T2w/Image-2.dcm  tmpno -107.229 -135.155 -53.6161<br>\n2   data/heart/train/00999/T2w/Image-3.dcm  tmpno -107.229 -135.155 -47.1161<br>\n3   data/heart/train/00999/T2w/Image-4.dcm  tmpno -107.229 -135.155 -40.6161<br>\n4   data/heart/train/00999/T2w/Image-5.dcm  tmpno -107.229 -135.155 -34.1161<br>\n5   data/heart/train/00999/T2w/Image-6.dcm  tmpno -107.229 -135.155 -27.6161<br>\n6   data/heart/train/00999/T2w/Image-7.dcm  tmpno -107.229 -135.155 -21.1161<br>\n7   data/heart/train/00999/T2w/Image-8.dcm  tmpno -107.229 -135.155 -14.6161<br>\n8   data/heart/train/00999/T2w/Image-9.dcm  tmpno -107.229 -135.155  -8.1161</p>\n<p>distance 7</p>\n<p>0   data/heart/train/00691/FLAIR/Image-1.dcm  tmpno  75.6942 -163.296  123.048<br>\n1   data/heart/train/00691/FLAIR/Image-2.dcm  tmpno  74.5975 -163.237  122.987<br>\n2   data/heart/train/00691/FLAIR/Image-3.dcm  tmpno  73.5008 -163.178  122.926<br>\n3   data/heart/train/00691/FLAIR/Image-4.dcm  tmpno  72.4041 -163.118  122.865<br>\n4   data/heart/train/00691/FLAIR/Image-5.dcm  tmpno  71.3074 -163.059  122.804<br>\n5   data/heart/train/00691/FLAIR/Image-6.dcm  tmpno  70.2107 -163.000  122.742<br>\n6   data/heart/train/00691/FLAIR/Image-7.dcm  tmpno  69.1140 -162.940  122.681<br>\n7   data/heart/train/00691/FLAIR/Image-8.dcm  tmpno  68.0173 -162.881  122.620<br>\n8   data/heart/train/00691/FLAIR/Image-9.dcm  tmpno  66.9206 -162.822  122.559<br>\n9  data/heart/train/00691/FLAIR/Image-10.dcm  tmpno  65.8239 -162.762  122.498:</p>\n<p>distance 0.06</p>\n<p>0   data/heart/train/00691/T1w/Image-1.dcm  tmpno -129.382 -155.674 -104.728<br>\n1   data/heart/train/00691/T1w/Image-2.dcm  tmpno -129.418 -155.674 -104.229<br>\n2   data/heart/train/00691/T1w/Image-3.dcm  tmpno -129.453 -155.674 -103.730<br>\n3   data/heart/train/00691/T1w/Image-4.dcm  tmpno -129.489 -155.674 -103.231<br>\n4   data/heart/train/00691/T1w/Image-5.dcm  tmpno -129.524 -155.674 -102.733<br>\n5   data/heart/train/00691/T1w/Image-6.dcm  tmpno -129.560 -155.674 -102.234<br>\n6   data/heart/train/00691/T1w/Image-7.dcm  tmpno -129.595 -155.674 -101.735<br>\n7   data/heart/train/00691/T1w/Image-8.dcm  tmpno -129.631 -155.674 -101.236<br>\n8   data/heart/train/00691/T1w/Image-9.dcm  tmpno -129.666 -155.674 -100.738<br>\n9  data/heart/train/00691/T1w/Image-10.dcm  tmpno -129.702 -155.674 -100.239:</p>\n<p>distance 0.5</p>",
      "rawMarkdown": "Because of the different scanning distances, i think 3dconv is not suitable.so use conv2d for voxel\n\ni had check the data of scan distance ,it not same .\n\n\n0   data/heart/train/00999/T2w/Image-1.dcm  tmpno -107.229 -135.155 -60.1161\n1   data/heart/train/00999/T2w/Image-2.dcm  tmpno -107.229 -135.155 -53.6161\n2   data/heart/train/00999/T2w/Image-3.dcm  tmpno -107.229 -135.155 -47.1161\n3   data/heart/train/00999/T2w/Image-4.dcm  tmpno -107.229 -135.155 -40.6161\n4   data/heart/train/00999/T2w/Image-5.dcm  tmpno -107.229 -135.155 -34.1161\n5   data/heart/train/00999/T2w/Image-6.dcm  tmpno -107.229 -135.155 -27.6161\n6   data/heart/train/00999/T2w/Image-7.dcm  tmpno -107.229 -135.155 -21.1161\n7   data/heart/train/00999/T2w/Image-8.dcm  tmpno -107.229 -135.155 -14.6161\n8   data/heart/train/00999/T2w/Image-9.dcm  tmpno -107.229 -135.155  -8.1161\n\ndistance 7\n\n0   data/heart/train/00691/FLAIR/Image-1.dcm  tmpno  75.6942 -163.296  123.048\n1   data/heart/train/00691/FLAIR/Image-2.dcm  tmpno  74.5975 -163.237  122.987\n2   data/heart/train/00691/FLAIR/Image-3.dcm  tmpno  73.5008 -163.178  122.926\n3   data/heart/train/00691/FLAIR/Image-4.dcm  tmpno  72.4041 -163.118  122.865\n4   data/heart/train/00691/FLAIR/Image-5.dcm  tmpno  71.3074 -163.059  122.804\n5   data/heart/train/00691/FLAIR/Image-6.dcm  tmpno  70.2107 -163.000  122.742\n6   data/heart/train/00691/FLAIR/Image-7.dcm  tmpno  69.1140 -162.940  122.681\n7   data/heart/train/00691/FLAIR/Image-8.dcm  tmpno  68.0173 -162.881  122.620\n8   data/heart/train/00691/FLAIR/Image-9.dcm  tmpno  66.9206 -162.822  122.559\n9  data/heart/train/00691/FLAIR/Image-10.dcm  tmpno  65.8239 -162.762  122.498:\n\ndistance 0.06\n\n0   data/heart/train/00691/T1w/Image-1.dcm  tmpno -129.382 -155.674 -104.728\n1   data/heart/train/00691/T1w/Image-2.dcm  tmpno -129.418 -155.674 -104.229\n2   data/heart/train/00691/T1w/Image-3.dcm  tmpno -129.453 -155.674 -103.730\n3   data/heart/train/00691/T1w/Image-4.dcm  tmpno -129.489 -155.674 -103.231\n4   data/heart/train/00691/T1w/Image-5.dcm  tmpno -129.524 -155.674 -102.733\n5   data/heart/train/00691/T1w/Image-6.dcm  tmpno -129.560 -155.674 -102.234\n6   data/heart/train/00691/T1w/Image-7.dcm  tmpno -129.595 -155.674 -101.735\n7   data/heart/train/00691/T1w/Image-8.dcm  tmpno -129.631 -155.674 -101.236\n8   data/heart/train/00691/T1w/Image-9.dcm  tmpno -129.666 -155.674 -100.738\n9  data/heart/train/00691/T1w/Image-10.dcm  tmpno -129.702 -155.674 -100.239:\n\ndistance 0.5",
      "votes": null
    },
    {
      "id": "1512403",
      "postDate": "09/14/2021 08:23:15",
      "content": "<p>I agree that this data is not perfect for 3D conv, however with 2D conv we encounter another problem: Single image may not contain all the information that is needed for classification. <br>\nYou could theoretically use another network to select these 2d images that actually contain cancer and train only on these, but the problem might still be present in this case.<br>\nAnother solution would be to run 2D conv on every slice and make it output feature maps instead of final predictions and then use another network on top of that, however from my experience it is just a more complicated version of 3d conv net in terms of performance.</p>",
      "rawMarkdown": "I agree that this data is not perfect for 3D conv, however with 2D conv we encounter another problem: Single image may not contain all the information that is needed for classification. \nYou could theoretically use another network to select these 2d images that actually contain cancer and train only on these, but the problem might still be present in this case.\nAnother solution would be to run 2D conv on every slice and make it output feature maps instead of final predictions and then use another network on top of that, however from my experience it is just a more complicated version of 3d conv net in terms of performance.",
      "votes": null
    },
    {
      "id": "1512970",
      "postDate": "09/14/2021 17:58:08",
      "content": "<p>I didn't understood your point can you pls elaborate it…</p>",
      "rawMarkdown": "I didn't understood your point can you pls elaborate it...",
      "votes": null
    },
    {
      "id": "1513007",
      "postDate": "09/14/2021 18:17:23",
      "content": "<p>There are indeed some notebooks <a href=\"https://www.kaggle.com/aimind\" target=\"_blank\">@aimind</a> using 2Dconv. Can you explain your point? It would be really helpful!</p>",
      "rawMarkdown": "There are indeed some notebooks @aimind using 2Dconv. Can you explain your point? It would be really helpful!",
      "votes": null
    },
    {
      "id": "1513249",
      "postDate": "09/15/2021 02:29:51",
      "content": "<p><a href=\"https://www.kaggle.com/aimind/3d-is-not-suitable-because-scan-intervals-vary\" target=\"_blank\">https://www.kaggle.com/aimind/3d-is-not-suitable-because-scan-intervals-vary</a></p>",
      "rawMarkdown": "https://www.kaggle.com/aimind/3d-is-not-suitable-because-scan-intervals-vary",
      "votes": null
    },
    {
      "id": "1513250",
      "postDate": "09/15/2021 02:29:57",
      "content": "<p><a href=\"https://www.kaggle.com/aimind/3d-is-not-suitable-because-scan-intervals-vary\" target=\"_blank\">https://www.kaggle.com/aimind/3d-is-not-suitable-because-scan-intervals-vary</a></p>",
      "rawMarkdown": "https://www.kaggle.com/aimind/3d-is-not-suitable-because-scan-intervals-vary",
      "votes": null
    },
    {
      "id": "1513255",
      "postDate": "09/15/2021 02:32:02",
      "content": "<p><a href=\"https://github.com/Ma-Lab-Berkeley/ReduNet\" target=\"_blank\">https://github.com/Ma-Lab-Berkeley/ReduNet</a></p>",
      "rawMarkdown": "https://github.com/Ma-Lab-Berkeley/ReduNet",
      "votes": null
    },
    {
      "id": "1513259",
      "postDate": "09/15/2021 02:36:37",
      "content": "<p><a href=\"https://www.kaggle.com/aimind/whitebox-voxel-96-conv2d-mcr2\" target=\"_blank\">https://www.kaggle.com/aimind/whitebox-voxel-96-conv2d-mcr2</a>  i try using mcr2loss ,but not good. local train it can reach 074 acc,but kaggle train only max 066 and not stable;  I am not sure whether this problem can be solved by deep learning, or even whether it can be solved.</p>",
      "rawMarkdown": "https://www.kaggle.com/aimind/whitebox-voxel-96-conv2d-mcr2  i try using mcr2loss ,but not good. local train it can reach 074 acc,but kaggle train only max 066 and not stable;  I am not sure whether this problem can be solved by deep learning, or even whether it can be solved.",
      "votes": null
    },
    {
      "id": "1517409",
      "postDate": "09/19/2021 15:09:14",
      "content": "<p>I talked about it <a href=\"https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/271902\" target=\"_blank\">https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/271902</a> here . You are right , the 3d mri is just like a 3d image of the brain ,just like 2d images in a 3d form , you classify them using a 3d cnn not a lstm not a 2d cnn . what people are calling 2d images are just slices of 3d images in that cause you can also use 1d convolutional neural network on 1D slices of a 2d image . But that will not work as you will miss info about the 2d image</p>",
      "rawMarkdown": "I talked about it https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/271902 here . You are right , the 3d mri is just like a 3d image of the brain ,just like 2d images in a 3d form , you classify them using a 3d cnn not a lstm not a 2d cnn . what people are calling 2d images are just slices of 3d images in that cause you can also use 1d convolutional neural network on 1D slices of a 2d image . But that will not work as you will miss info about the 2d image",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1512403,
      "author_name": "kuba386",
      "author_url": "",
      "post_date": "09/14/2021 08:23:15",
      "content": "<p>I agree that this data is not perfect for 3D conv, however with 2D conv we encounter another problem: Single image may not contain all the information that is needed for classification. <br>\nYou could theoretically use another network to select these 2d images that actually contain cancer and train only on these, but the problem might still be present in this case.<br>\nAnother solution would be to run 2D conv on every slice and make it output feature maps instead of final predictions and then use another network on top of that, however from my experience it is just a more complicated version of 3d conv net in terms of performance.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1517409,
          "author_name": "swaralipibose",
          "author_url": "",
          "post_date": "09/19/2021 15:09:14",
          "content": "<p>I talked about it <a href=\"https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/271902\" target=\"_blank\">https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/271902</a> here . You are right , the 3d mri is just like a 3d image of the brain ,just like 2d images in a 3d form , you classify them using a 3d cnn not a lstm not a 2d cnn . what people are calling 2d images are just slices of 3d images in that cause you can also use 1d convolutional neural network on 1D slices of a 2d image . But that will not work as you will miss info about the 2d image</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1512970,
      "author_name": "sj161199",
      "author_url": "",
      "post_date": "09/14/2021 17:58:08",
      "content": "<p>I didn't understood your point can you pls elaborate it…</p>",
      "votes": null,
      "replies": [
        {
          "id": 1513250,
          "author_name": "aimind",
          "author_url": "",
          "post_date": "09/15/2021 02:29:57",
          "content": "<p><a href=\"https://www.kaggle.com/aimind/3d-is-not-suitable-because-scan-intervals-vary\" target=\"_blank\">https://www.kaggle.com/aimind/3d-is-not-suitable-because-scan-intervals-vary</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1513007,
      "author_name": "conandoyle",
      "author_url": "",
      "post_date": "09/14/2021 18:17:23",
      "content": "<p>There are indeed some notebooks <a href=\"https://www.kaggle.com/aimind\" target=\"_blank\">@aimind</a> using 2Dconv. Can you explain your point? It would be really helpful!</p>",
      "votes": null,
      "replies": [
        {
          "id": 1513249,
          "author_name": "aimind",
          "author_url": "",
          "post_date": "09/15/2021 02:29:51",
          "content": "<p><a href=\"https://www.kaggle.com/aimind/3d-is-not-suitable-because-scan-intervals-vary\" target=\"_blank\">https://www.kaggle.com/aimind/3d-is-not-suitable-because-scan-intervals-vary</a></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1513259,
          "author_name": "aimind",
          "author_url": "",
          "post_date": "09/15/2021 02:36:37",
          "content": "<p><a href=\"https://www.kaggle.com/aimind/whitebox-voxel-96-conv2d-mcr2\" target=\"_blank\">https://www.kaggle.com/aimind/whitebox-voxel-96-conv2d-mcr2</a>  i try using mcr2loss ,but not good. local train it can reach 074 acc,but kaggle train only max 066 and not stable;  I am not sure whether this problem can be solved by deep learning, or even whether it can be solved.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1513255,
      "author_name": "aimind",
      "author_url": "",
      "post_date": "09/15/2021 02:32:02",
      "content": "<p><a href=\"https://github.com/Ma-Lab-Berkeley/ReduNet\" target=\"_blank\">https://github.com/Ma-Lab-Berkeley/ReduNet</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1510710": "Because of the different scanning distances, i think 3dconv is not suitable.so use conv2d for voxel\n\ni had check the data of scan distance ,it not same .\n\n\n0   data/heart/train/00999/T2w/Image-1.dcm  tmpno -107.229 -135.155 -60.1161\n1   data/heart/train/00999/T2w/Image-2.dcm  tmpno -107.229 -135.155 -53.6161\n2   data/heart/train/00999/T2w/Image-3.dcm  tmpno -107.229 -135.155 -47.1161\n3   data/heart/train/00999/T2w/Image-4.dcm  tmpno -107.229 -135.155 -40.6161\n4   data/heart/train/00999/T2w/Image-5.dcm  tmpno -107.229 -135.155 -34.1161\n5   data/heart/train/00999/T2w/Image-6.dcm  tmpno -107.229 -135.155 -27.6161\n6   data/heart/train/00999/T2w/Image-7.dcm  tmpno -107.229 -135.155 -21.1161\n7   data/heart/train/00999/T2w/Image-8.dcm  tmpno -107.229 -135.155 -14.6161\n8   data/heart/train/00999/T2w/Image-9.dcm  tmpno -107.229 -135.155  -8.1161\n\ndistance 7\n\n0   data/heart/train/00691/FLAIR/Image-1.dcm  tmpno  75.6942 -163.296  123.048\n1   data/heart/train/00691/FLAIR/Image-2.dcm  tmpno  74.5975 -163.237  122.987\n2   data/heart/train/00691/FLAIR/Image-3.dcm  tmpno  73.5008 -163.178  122.926\n3   data/heart/train/00691/FLAIR/Image-4.dcm  tmpno  72.4041 -163.118  122.865\n4   data/heart/train/00691/FLAIR/Image-5.dcm  tmpno  71.3074 -163.059  122.804\n5   data/heart/train/00691/FLAIR/Image-6.dcm  tmpno  70.2107 -163.000  122.742\n6   data/heart/train/00691/FLAIR/Image-7.dcm  tmpno  69.1140 -162.940  122.681\n7   data/heart/train/00691/FLAIR/Image-8.dcm  tmpno  68.0173 -162.881  122.620\n8   data/heart/train/00691/FLAIR/Image-9.dcm  tmpno  66.9206 -162.822  122.559\n9  data/heart/train/00691/FLAIR/Image-10.dcm  tmpno  65.8239 -162.762  122.498:\n\ndistance 0.06\n\n0   data/heart/train/00691/T1w/Image-1.dcm  tmpno -129.382 -155.674 -104.728\n1   data/heart/train/00691/T1w/Image-2.dcm  tmpno -129.418 -155.674 -104.229\n2   data/heart/train/00691/T1w/Image-3.dcm  tmpno -129.453 -155.674 -103.730\n3   data/heart/train/00691/T1w/Image-4.dcm  tmpno -129.489 -155.674 -103.231\n4   data/heart/train/00691/T1w/Image-5.dcm  tmpno -129.524 -155.674 -102.733\n5   data/heart/train/00691/T1w/Image-6.dcm  tmpno -129.560 -155.674 -102.234\n6   data/heart/train/00691/T1w/Image-7.dcm  tmpno -129.595 -155.674 -101.735\n7   data/heart/train/00691/T1w/Image-8.dcm  tmpno -129.631 -155.674 -101.236\n8   data/heart/train/00691/T1w/Image-9.dcm  tmpno -129.666 -155.674 -100.738\n9  data/heart/train/00691/T1w/Image-10.dcm  tmpno -129.702 -155.674 -100.239:\n\ndistance 0.5",
    "1512403": "I agree that this data is not perfect for 3D conv, however with 2D conv we encounter another problem: Single image may not contain all the information that is needed for classification. \nYou could theoretically use another network to select these 2d images that actually contain cancer and train only on these, but the problem might still be present in this case.\nAnother solution would be to run 2D conv on every slice and make it output feature maps instead of final predictions and then use another network on top of that, however from my experience it is just a more complicated version of 3d conv net in terms of performance.",
    "1512970": "I didn't understood your point can you pls elaborate it...",
    "1513007": "There are indeed some notebooks @aimind using 2Dconv. Can you explain your point? It would be really helpful!",
    "1513249": "https://www.kaggle.com/aimind/3d-is-not-suitable-because-scan-intervals-vary",
    "1513250": "https://www.kaggle.com/aimind/3d-is-not-suitable-because-scan-intervals-vary",
    "1513255": "https://github.com/Ma-Lab-Berkeley/ReduNet",
    "1513259": "https://www.kaggle.com/aimind/whitebox-voxel-96-conv2d-mcr2  i try using mcr2loss ,but not good. local train it can reach 074 acc,but kaggle train only max 066 and not stable;  I am not sure whether this problem can be solved by deep learning, or even whether it can be solved.",
    "1517409": "I talked about it https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/271902 here . You are right , the 3d mri is just like a 3d image of the brain ,just like 2d images in a 3d form , you classify them using a 3d cnn not a lstm not a 2d cnn . what people are calling 2d images are just slices of 3d images in that cause you can also use 1d convolutional neural network on 1D slices of a 2d image . But that will not work as you will miss info about the 2d image"
  },
  "source": "meta"
}