{
  "id": 253182,
  "title": "Good resources for 3D Convolutional models",
  "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/253182",
  "author_name": "",
  "post_date": "2021-07-15T09:05:33.165830Z",
  "votes": 19,
  "comment_count": 11,
  "views": 0,
  "content": "<p>Hello everyone, as we know that we are dealing with a problem that can also be solved by using 3D models. And training these models from scratch would be too expensive wrt computation.<br>\nSo are there any 3D pre-trained models that we can use? I hope you will share them<br>\n(by pretrained, I mean models that have been trained on free-to-use datasets like Imagenet, or use backbone networks like resnets,efficentnets etc that have been trained on Imagenet dataset)</p>\n<p>If you are confused with 3D Conv models, check out this video where a past competition like this was discussed by sentdex <a href=\"https://www.youtube.com/watch?v=ulq9DjCJPDU&amp;list=PLQVvvaa0QuDd5meH8cStO9cMi98tPT12_\" target=\"_blank\">https://www.youtube.com/watch?v=ulq9DjCJPDU&amp;list=PLQVvvaa0QuDd5meH8cStO9cMi98tPT12_</a></p>",
  "messages": [
    {
      "id": "1388853",
      "postDate": "07/15/2021 09:05:33",
      "content": "<p>Hello everyone, as we know that we are dealing with a problem that can also be solved by using 3D models. And training these models from scratch would be too expensive wrt computation.<br>\nSo are there any 3D pre-trained models that we can use? I hope you will share them<br>\n(by pretrained, I mean models that have been trained on free-to-use datasets like Imagenet, or use backbone networks like resnets,efficentnets etc that have been trained on Imagenet dataset)</p>\n<p>If you are confused with 3D Conv models, check out this video where a past competition like this was discussed by sentdex <a href=\"https://www.youtube.com/watch?v=ulq9DjCJPDU&amp;list=PLQVvvaa0QuDd5meH8cStO9cMi98tPT12_\" target=\"_blank\">https://www.youtube.com/watch?v=ulq9DjCJPDU&amp;list=PLQVvvaa0QuDd5meH8cStO9cMi98tPT12_</a></p>",
      "rawMarkdown": "Hello everyone, as we know that we are dealing with a problem that can also be solved by using 3D models. And training these models from scratch would be too expensive wrt computation.\nSo are there any 3D pre-trained models that we can use? I hope you will share them\n(by pretrained, I mean models that have been trained on free-to-use datasets like Imagenet, or use backbone networks like resnets,efficentnets etc that have been trained on Imagenet dataset)\n\nIf you are confused with 3D Conv models, check out this video where a past competition like this was discussed by sentdex https://www.youtube.com/watch?v=ulq9DjCJPDU&list=PLQVvvaa0QuDd5meH8cStO9cMi98tPT12_",
      "votes": null
    },
    {
      "id": "1388898",
      "postDate": "07/15/2021 09:52:52",
      "content": "<p><a href=\"https://www.kaggle.com/mrinath\" target=\"_blank\">@mrinath</a>, awesome post. It would be awesome to have a 3D pre-trained model indeed. My poor computer can't even handle training models that analyse pictues haha. I am following your topic. Hope someone might be able to help.</p>",
      "rawMarkdown": "mrinath, awesome post. It would be awesome to have a 3D pre-trained model indeed. My poor computer can't even handle training models that analyse pictues haha. I am following your topic. Hope someone might be able to help.",
      "votes": null
    },
    {
      "id": "1388908",
      "postDate": "07/15/2021 09:59:54",
      "content": "<p><a href=\"https://www.kaggle.com/mrinath\" target=\"_blank\">@mrinath</a> This is very relatable. I had to drop a lot of interesting projects because of the expense! This is an excellent share. Thanks a lot for this! Hopefully we get some help on this soon.</p>",
      "rawMarkdown": "mrinath This is very relatable. I had to drop a lot of interesting projects because of the expense! This is an excellent share. Thanks a lot for this! Hopefully we get some help on this soon.",
      "votes": null
    },
    {
      "id": "1388927",
      "postDate": "07/15/2021 10:09:42",
      "content": "<p>Same here, lol</p>",
      "rawMarkdown": "Same here, lol",
      "votes": null
    },
    {
      "id": "1389547",
      "postDate": "07/15/2021 20:06:15",
      "content": "<p>Yes, that would be great. I had a look at the MONAI library (an open-source python package for deep learning on medical data, <a href=\"https://docs.monai.io/en/latest/networks.html#nets\" target=\"_blank\">https://docs.monai.io/en/latest/networks.html#nets</a>), but they don't seem to have any pretrained 3D models. <br>\nI think a reliable 2D/3D model for tumor segmentation would be super helpful. Then it would be possible to cut out the tumors and train on a much smaller and hopefully more meaningful dataset.</p>",
      "rawMarkdown": "Yes, that would be great. I had a look at the MONAI library (an open-source python package for deep learning on medical data, [https://docs.monai.io/en/latest/networks.html#nets](https://docs.monai.io/en/latest/networks.html#nets)), but they don't seem to have any pretrained 3D models. \nI think a reliable 2D/3D model for tumor segmentation would be super helpful. Then it would be possible to cut out the tumors and train on a much smaller and hopefully more meaningful dataset.",
      "votes": null
    },
    {
      "id": "1389811",
      "postDate": "07/16/2021 05:20:46",
      "content": "<p>That's exactly what I was thinking of.</p>",
      "rawMarkdown": "That's exactly what I was thinking of.",
      "votes": null
    },
    {
      "id": "1391536",
      "postDate": "07/17/2021 17:28:53",
      "content": "<p>Great post!  I was actually just reading an article about Andrew Ng's CV course, yesterday, and there was a reference in there just around this, specifically, under the section on \"How to win CV competitions\" - :) --</p>\n<p>It recommended using \"an open source implementation and pretrained model to start and then fine-tune the parameters for your particular application.\"  However, not familiar to this subject so I am glad to see others posing the same question.  where would we find some pre-trained models for CVS? and which ones would be appropriate to use in this case?</p>\n<p>Thanks for sharing the video link, i will check it out.</p>",
      "rawMarkdown": "Great post!  I was actually just reading an article about Andrew Ng's CV course, yesterday, and there was a reference in there just around this, specifically, under the section on \"How to win CV competitions\" - :) --\n\nIt recommended using \"an open source implementation and pretrained model to start and then fine-tune the parameters for your particular application.\"  However, not familiar to this subject so I am glad to see others posing the same question.  where would we find some pre-trained models for CVS? and which ones would be appropriate to use in this case?\n\nThanks for sharing the video link, i will check it out.",
      "votes": null
    },
    {
      "id": "1392013",
      "postDate": "07/18/2021 09:42:24",
      "content": "<p>Take a look at <a href=\"https://github.com/Tencent/MedicalNet\" target=\"_blank\">https://github.com/Tencent/MedicalNet</a>, it seems very close to what you’re looking for </p>",
      "rawMarkdown": "Take a look at https://github.com/Tencent/MedicalNet, it seems very close to what you’re looking for",
      "votes": null
    },
    {
      "id": "1392048",
      "postDate": "07/18/2021 10:25:18",
      "content": "<p>Really thanks for this and yes it looks like what we were searching for</p>",
      "rawMarkdown": "Really thanks for this and yes it looks like what we were searching for",
      "votes": null
    },
    {
      "id": "1395274",
      "postDate": "07/21/2021 05:14:03",
      "content": "<p>I think the best way to find good models is to do a search on <a href=\"https://paperswithcode.com/sota\" target=\"_blank\">Paper with Code</a>(PwC).</p>\n<p>For example, here is a link to a dataset in the medical field with the search term \"brain\".<br>\n👉 <a href=\"https://paperswithcode.com/datasets?q=brain&amp;v=lst&amp;o=match&amp;mod=medical&amp;page=1\" target=\"_blank\">https://paperswithcode.com/datasets?q=brain&amp;v=lst&amp;o=match&amp;mod=medical&amp;page=1</a></p>\n<p><img src=\"https://f.easyuploader.app/20210721135737_73795730.JPG\" alt=\"\"></p>\n<p>For each dataset, you can select a model that is well implemented and seems to have high accuracy. The pre-trained weights may be important in our case. </p>\n<p>For example, <a href=\"https://paperswithcode.com/dataset/brats-2018-1\" target=\"_blank\">BraTS 2018</a>, which is a previous version of this competition, is also in the list, and if you follow the links from there, you can find, for example, MedicalZooPytorch.<br>\n👉 <a href=\"https://github.com/black0017/MedicalZooPytorch\" target=\"_blank\">https://github.com/black0017/MedicalZooPytorch</a></p>\n<p>MedicalNet, mentioned by <a href=\"https://www.kaggle.com/reubenschmidt\" target=\"_blank\">@reubenschmidt</a> in the comment below, is also registered with PwC.<br>\n<a href=\"https://paperswithcode.com/paper/med3d-transfer-learning-for-3d-medical-image\" target=\"_blank\">https://paperswithcode.com/paper/med3d-transfer-learning-for-3d-medical-image</a></p>\n<p>If you find something that looks good, we would be happy if you could share it here 😊</p>",
      "rawMarkdown": "I think the best way to find good models is to do a search on [Paper with Code](https://paperswithcode.com/sota)(PwC).\n\nFor example, here is a link to a dataset in the medical field with the search term \"brain\".\n👉 https://paperswithcode.com/datasets?q=brain&v=lst&o=match&mod=medical&page=1\n\n![](https://f.easyuploader.app/20210721135737_73795730.JPG)\n\nFor each dataset, you can select a model that is well implemented and seems to have high accuracy. The pre-trained weights may be important in our case. \n\nFor example, [BraTS 2018](https://paperswithcode.com/dataset/brats-2018-1), which is a previous version of this competition, is also in the list, and if you follow the links from there, you can find, for example, MedicalZooPytorch.\n👉 https://github.com/black0017/MedicalZooPytorch\n\nMedicalNet, mentioned by @reubenschmidt in the comment below, is also registered with PwC.\nhttps://paperswithcode.com/paper/med3d-transfer-learning-for-3d-medical-image\n\nIf you find something that looks good, we would be happy if you could share it here 😊",
      "votes": null
    },
    {
      "id": "1395939",
      "postDate": "07/21/2021 16:26:32",
      "content": "<p>thank you <a href=\"https://www.kaggle.com/maxwell110\" target=\"_blank\">@maxwell110</a>. very useful information.</p>",
      "rawMarkdown": "thank you @maxwell110. very useful information.",
      "votes": null
    },
    {
      "id": "1398259",
      "postDate": "07/23/2021 23:37:08",
      "content": "<p>For <code>tensorflow/keras</code>, one can use a pretrained 3d model of efficientnet b0-b7 and other classification models from this repo: </p>\n<ul>\n<li><a href=\"https://github.com/ZFTurbo/classification_models_3D\" target=\"_blank\">https://github.com/ZFTurbo/classification_models_3D</a></li>\n<li><a href=\"https://github.com/ZFTurbo/efficientnet_3D\" target=\"_blank\">https://github.com/ZFTurbo/efficientnet_3D</a></li>\n</ul>\n<p>FYI, in my public notebook, I showed how to use them. <a href=\"https://www.kaggle.com/ipythonx/tf-3d-2d-model-for-brain-tumor-classification#C.-3D-Model-:-Pre-trained-Weights\" target=\"_blank\">here</a>, in section C. 3D Model : Pre-trained Weights</p>",
      "rawMarkdown": "For `tensorflow/keras`, one can use a pretrained 3d model of efficientnet b0-b7 and other classification models from this repo: \n\n- https://github.com/ZFTurbo/classification_models_3D\n- https://github.com/ZFTurbo/efficientnet_3D\n\nFYI, in my public notebook, I showed how to use them. [here](https://www.kaggle.com/ipythonx/tf-3d-2d-model-for-brain-tumor-classification#C.-3D-Model-:-Pre-trained-Weights), in section C. 3D Model : Pre-trained Weights",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1388898,
      "author_name": "jonaslneri",
      "author_url": "",
      "post_date": "07/15/2021 09:52:52",
      "content": "<p><a href=\"https://www.kaggle.com/mrinath\" target=\"_blank\">@mrinath</a>, awesome post. It would be awesome to have a 3D pre-trained model indeed. My poor computer can't even handle training models that analyse pictues haha. I am following your topic. Hope someone might be able to help.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1388927,
          "author_name": "mrinath",
          "author_url": "",
          "post_date": "07/15/2021 10:09:42",
          "content": "<p>Same here, lol</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1388908,
      "author_name": "adhithia",
      "author_url": "",
      "post_date": "07/15/2021 09:59:54",
      "content": "<p><a href=\"https://www.kaggle.com/mrinath\" target=\"_blank\">@mrinath</a> This is very relatable. I had to drop a lot of interesting projects because of the expense! This is an excellent share. Thanks a lot for this! Hopefully we get some help on this soon.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1389547,
      "author_name": "lillik",
      "author_url": "",
      "post_date": "07/15/2021 20:06:15",
      "content": "<p>Yes, that would be great. I had a look at the MONAI library (an open-source python package for deep learning on medical data, <a href=\"https://docs.monai.io/en/latest/networks.html#nets\" target=\"_blank\">https://docs.monai.io/en/latest/networks.html#nets</a>), but they don't seem to have any pretrained 3D models. <br>\nI think a reliable 2D/3D model for tumor segmentation would be super helpful. Then it would be possible to cut out the tumors and train on a much smaller and hopefully more meaningful dataset.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1389811,
          "author_name": "mrinath",
          "author_url": "",
          "post_date": "07/16/2021 05:20:46",
          "content": "<p>That's exactly what I was thinking of.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1391536,
      "author_name": "elenaeb",
      "author_url": "",
      "post_date": "07/17/2021 17:28:53",
      "content": "<p>Great post!  I was actually just reading an article about Andrew Ng's CV course, yesterday, and there was a reference in there just around this, specifically, under the section on \"How to win CV competitions\" - :) --</p>\n<p>It recommended using \"an open source implementation and pretrained model to start and then fine-tune the parameters for your particular application.\"  However, not familiar to this subject so I am glad to see others posing the same question.  where would we find some pre-trained models for CVS? and which ones would be appropriate to use in this case?</p>\n<p>Thanks for sharing the video link, i will check it out.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1392013,
      "author_name": "reubenschmidt",
      "author_url": "",
      "post_date": "07/18/2021 09:42:24",
      "content": "<p>Take a look at <a href=\"https://github.com/Tencent/MedicalNet\" target=\"_blank\">https://github.com/Tencent/MedicalNet</a>, it seems very close to what you’re looking for </p>",
      "votes": null,
      "replies": [
        {
          "id": 1392048,
          "author_name": "mrinath",
          "author_url": "",
          "post_date": "07/18/2021 10:25:18",
          "content": "<p>Really thanks for this and yes it looks like what we were searching for</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1395274,
      "author_name": "maxwell110",
      "author_url": "",
      "post_date": "07/21/2021 05:14:03",
      "content": "<p>I think the best way to find good models is to do a search on <a href=\"https://paperswithcode.com/sota\" target=\"_blank\">Paper with Code</a>(PwC).</p>\n<p>For example, here is a link to a dataset in the medical field with the search term \"brain\".<br>\n👉 <a href=\"https://paperswithcode.com/datasets?q=brain&amp;v=lst&amp;o=match&amp;mod=medical&amp;page=1\" target=\"_blank\">https://paperswithcode.com/datasets?q=brain&amp;v=lst&amp;o=match&amp;mod=medical&amp;page=1</a></p>\n<p><img src=\"https://f.easyuploader.app/20210721135737_73795730.JPG\" alt=\"\"></p>\n<p>For each dataset, you can select a model that is well implemented and seems to have high accuracy. The pre-trained weights may be important in our case. </p>\n<p>For example, <a href=\"https://paperswithcode.com/dataset/brats-2018-1\" target=\"_blank\">BraTS 2018</a>, which is a previous version of this competition, is also in the list, and if you follow the links from there, you can find, for example, MedicalZooPytorch.<br>\n👉 <a href=\"https://github.com/black0017/MedicalZooPytorch\" target=\"_blank\">https://github.com/black0017/MedicalZooPytorch</a></p>\n<p>MedicalNet, mentioned by <a href=\"https://www.kaggle.com/reubenschmidt\" target=\"_blank\">@reubenschmidt</a> in the comment below, is also registered with PwC.<br>\n<a href=\"https://paperswithcode.com/paper/med3d-transfer-learning-for-3d-medical-image\" target=\"_blank\">https://paperswithcode.com/paper/med3d-transfer-learning-for-3d-medical-image</a></p>\n<p>If you find something that looks good, we would be happy if you could share it here 😊</p>",
      "votes": null,
      "replies": [
        {
          "id": 1395939,
          "author_name": "tharun2001",
          "author_url": "",
          "post_date": "07/21/2021 16:26:32",
          "content": "<p>thank you <a href=\"https://www.kaggle.com/maxwell110\" target=\"_blank\">@maxwell110</a>. very useful information.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1398259,
      "author_name": "ipythonx",
      "author_url": "",
      "post_date": "07/23/2021 23:37:08",
      "content": "<p>For <code>tensorflow/keras</code>, one can use a pretrained 3d model of efficientnet b0-b7 and other classification models from this repo: </p>\n<ul>\n<li><a href=\"https://github.com/ZFTurbo/classification_models_3D\" target=\"_blank\">https://github.com/ZFTurbo/classification_models_3D</a></li>\n<li><a href=\"https://github.com/ZFTurbo/efficientnet_3D\" target=\"_blank\">https://github.com/ZFTurbo/efficientnet_3D</a></li>\n</ul>\n<p>FYI, in my public notebook, I showed how to use them. <a href=\"https://www.kaggle.com/ipythonx/tf-3d-2d-model-for-brain-tumor-classification#C.-3D-Model-:-Pre-trained-Weights\" target=\"_blank\">here</a>, in section C. 3D Model : Pre-trained Weights</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1388853": "Hello everyone, as we know that we are dealing with a problem that can also be solved by using 3D models. And training these models from scratch would be too expensive wrt computation.\nSo are there any 3D pre-trained models that we can use? I hope you will share them\n(by pretrained, I mean models that have been trained on free-to-use datasets like Imagenet, or use backbone networks like resnets,efficentnets etc that have been trained on Imagenet dataset)\n\nIf you are confused with 3D Conv models, check out this video where a past competition like this was discussed by sentdex https://www.youtube.com/watch?v=ulq9DjCJPDU&list=PLQVvvaa0QuDd5meH8cStO9cMi98tPT12_",
    "1388898": "mrinath, awesome post. It would be awesome to have a 3D pre-trained model indeed. My poor computer can't even handle training models that analyse pictues haha. I am following your topic. Hope someone might be able to help.",
    "1388908": "mrinath This is very relatable. I had to drop a lot of interesting projects because of the expense! This is an excellent share. Thanks a lot for this! Hopefully we get some help on this soon.",
    "1388927": "Same here, lol",
    "1389547": "Yes, that would be great. I had a look at the MONAI library (an open-source python package for deep learning on medical data, [https://docs.monai.io/en/latest/networks.html#nets](https://docs.monai.io/en/latest/networks.html#nets)), but they don't seem to have any pretrained 3D models. \nI think a reliable 2D/3D model for tumor segmentation would be super helpful. Then it would be possible to cut out the tumors and train on a much smaller and hopefully more meaningful dataset.",
    "1389811": "That's exactly what I was thinking of.",
    "1391536": "Great post!  I was actually just reading an article about Andrew Ng's CV course, yesterday, and there was a reference in there just around this, specifically, under the section on \"How to win CV competitions\" - :) --\n\nIt recommended using \"an open source implementation and pretrained model to start and then fine-tune the parameters for your particular application.\"  However, not familiar to this subject so I am glad to see others posing the same question.  where would we find some pre-trained models for CVS? and which ones would be appropriate to use in this case?\n\nThanks for sharing the video link, i will check it out.",
    "1392013": "Take a look at https://github.com/Tencent/MedicalNet, it seems very close to what you’re looking for",
    "1392048": "Really thanks for this and yes it looks like what we were searching for",
    "1395274": "I think the best way to find good models is to do a search on [Paper with Code](https://paperswithcode.com/sota)(PwC).\n\nFor example, here is a link to a dataset in the medical field with the search term \"brain\".\n👉 https://paperswithcode.com/datasets?q=brain&v=lst&o=match&mod=medical&page=1\n\n![](https://f.easyuploader.app/20210721135737_73795730.JPG)\n\nFor each dataset, you can select a model that is well implemented and seems to have high accuracy. The pre-trained weights may be important in our case. \n\nFor example, [BraTS 2018](https://paperswithcode.com/dataset/brats-2018-1), which is a previous version of this competition, is also in the list, and if you follow the links from there, you can find, for example, MedicalZooPytorch.\n👉 https://github.com/black0017/MedicalZooPytorch\n\nMedicalNet, mentioned by @reubenschmidt in the comment below, is also registered with PwC.\nhttps://paperswithcode.com/paper/med3d-transfer-learning-for-3d-medical-image\n\nIf you find something that looks good, we would be happy if you could share it here 😊",
    "1395939": "thank you @maxwell110. very useful information.",
    "1398259": "For `tensorflow/keras`, one can use a pretrained 3d model of efficientnet b0-b7 and other classification models from this repo: \n\n- https://github.com/ZFTurbo/classification_models_3D\n- https://github.com/ZFTurbo/efficientnet_3D\n\nFYI, in my public notebook, I showed how to use them. [here](https://www.kaggle.com/ipythonx/tf-3d-2d-model-for-brain-tumor-classification#C.-3D-Model-:-Pre-trained-Weights), in section C. 3D Model : Pre-trained Weights"
  },
  "source": "meta"
}