{
  "id": 265950,
  "title": "Submissions part is not quite clear",
  "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/265950",
  "author_name": "",
  "post_date": "2021-08-17T13:15:29.117245500Z",
  "votes": 4,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi, all. This is what I understood by my Test &amp; Trials:</p>\n<ul>\n<li>We shall get all the test data directly from the /test folder (which might be populated with more samples later). If so, then preprocessing and working on the previously created tfrecords is impossible. We need to process the .dcm files in real-time. </li>\n<li>Since the Internet shall be disabled and we are supposed to process files in real-time, how to use transfer learning (EfficientNet or InceptionV3)?</li>\n<li>The possible values can be different than 1.0 and 0.0 (understood it intuitively, from the evaluation scores). In medicine, usually, the prediction is given as yes or no.</li>\n</ul>",
  "messages": [
    {
      "id": "1477400",
      "postDate": "08/17/2021 13:15:29",
      "content": "<p>Hi, all. This is what I understood by my Test &amp; Trials:</p>\n<ul>\n<li>We shall get all the test data directly from the /test folder (which might be populated with more samples later). If so, then preprocessing and working on the previously created tfrecords is impossible. We need to process the .dcm files in real-time. </li>\n<li>Since the Internet shall be disabled and we are supposed to process files in real-time, how to use transfer learning (EfficientNet or InceptionV3)?</li>\n<li>The possible values can be different than 1.0 and 0.0 (understood it intuitively, from the evaluation scores). In medicine, usually, the prediction is given as yes or no.</li>\n</ul>",
      "rawMarkdown": "Hi, all. This is what I understood by my Test & Trials:\n- We shall get all the test data directly from the /test folder (which might be populated with more samples later). If so, then preprocessing and working on the previously created tfrecords is impossible. We need to process the .dcm files in real-time. \n- Since the Internet shall be disabled and we are supposed to process files in real-time, how to use transfer learning (EfficientNet or InceptionV3)?\n- The possible values can be different than 1.0 and 0.0 (understood it intuitively, from the evaluation scores). In medicine, usually, the prediction is given as yes or no.",
      "votes": null
    },
    {
      "id": "1480253",
      "postDate": "08/18/2021 22:32:16",
      "content": "<ol>\n<li><p>We cannot see the private test data, so you are correct, we must process the DICOM files directly.</p></li>\n<li><p>Load pretrained models as a dataset and include in your notebook.</p></li>\n<li><p>In the AI world, I think most things are predictions. Never absolute. The same can be said for diagnostic radiology in cases that involve more than a visual interpretation of an image to make a diagnosis. In the absence of correlating data, a  radiologist might say \"there IS a tumor that is suspicious for malignancy\" .. but you will likely never see a diagnosis of \"malignant tumor\" from a single MR study.</p></li>\n</ol>",
      "rawMarkdown": "1. We cannot see the private test data, so you are correct, we must process the DICOM files directly.\n\n2. Load pretrained models as a dataset and include in your notebook.\n\n3. In the AI world, I think most things are predictions. Never absolute. The same can be said for diagnostic radiology in cases that involve more than a visual interpretation of an image to make a diagnosis. In the absence of correlating data, a  radiologist might say \"there IS a tumor that is suspicious for malignancy\" .. but you will likely never see a diagnosis of \"malignant tumor\" from a single MR study.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1480253,
      "author_name": "davidbroberts",
      "author_url": "",
      "post_date": "08/18/2021 22:32:16",
      "content": "<ol>\n<li><p>We cannot see the private test data, so you are correct, we must process the DICOM files directly.</p></li>\n<li><p>Load pretrained models as a dataset and include in your notebook.</p></li>\n<li><p>In the AI world, I think most things are predictions. Never absolute. The same can be said for diagnostic radiology in cases that involve more than a visual interpretation of an image to make a diagnosis. In the absence of correlating data, a  radiologist might say \"there IS a tumor that is suspicious for malignancy\" .. but you will likely never see a diagnosis of \"malignant tumor\" from a single MR study.</p></li>\n</ol>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1477400": "Hi, all. This is what I understood by my Test & Trials:\n- We shall get all the test data directly from the /test folder (which might be populated with more samples later). If so, then preprocessing and working on the previously created tfrecords is impossible. We need to process the .dcm files in real-time. \n- Since the Internet shall be disabled and we are supposed to process files in real-time, how to use transfer learning (EfficientNet or InceptionV3)?\n- The possible values can be different than 1.0 and 0.0 (understood it intuitively, from the evaluation scores). In medicine, usually, the prediction is given as yes or no.",
    "1480253": "1. We cannot see the private test data, so you are correct, we must process the DICOM files directly.\n\n2. Load pretrained models as a dataset and include in your notebook.\n\n3. In the AI world, I think most things are predictions. Never absolute. The same can be said for diagnostic radiology in cases that involve more than a visual interpretation of an image to make a diagnosis. In the absence of correlating data, a  radiologist might say \"there IS a tumor that is suspicious for malignancy\" .. but you will likely never see a diagnosis of \"malignant tumor\" from a single MR study."
  },
  "source": "meta"
}