{
  "id": 374930,
  "title": "split the data by patient id or images",
  "url": "/competitions/rsna-breast-cancer-detection/discussion/374930",
  "author_name": "Lucario129",
  "post_date": "2022-12-29T14:22:28.105000",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>how do we need to split the data, by individual images or by patient, my approach was to stack all the images of each patient and train the model for it </p>",
  "messages": [
    {
      "id": 2079673,
      "postDate": "2022-12-29T14:22:28.107Z",
      "content": "<p>how do we need to split the data, by individual images or by patient, my approach was to stack all the images of each patient and train the model for it </p>",
      "rawMarkdown": "how do we need to split the data, by individual images or by patient, my approach was to stack all the images of each patient and train the model for it ",
      "votes": 1
    },
    {
      "id": 2080105,
      "postDate": "2022-12-29T21:01:05.190Z",
      "content": "<p>One rule of thumb for medical imaging is to always try to split by patient, in order to prevent data leakage between training, validation and test sets.</p>",
      "rawMarkdown": "One rule of thumb for medical imaging is to always try to split by patient, in order to prevent data leakage between training, validation and test sets."
    },
    {
      "id": 2079986,
      "postDate": "2022-12-29T19:04:40.750Z",
      "content": "<p>By patient reflects better the expected submission and seems to be the correct way. <br>\nI only tried by individual images so far though.<br>\nAs for stacking seems to be a good idea, only if you can support higher input sizes to mantain the details, plus to make sure you can do the same on test set without losing patient's samples.</p>",
      "rawMarkdown": "By patient reflects better the expected submission and seems to be the correct way. \nI only tried by individual images so far though.\nAs for stacking seems to be a good idea, only if you can support higher input sizes to mantain the details, plus to make sure you can do the same on test set without losing patient's samples."
    }
  ],
  "comments": [
    {
      "id": 2080105,
      "author_name": "JAbrantes",
      "author_url": "",
      "post_date": "2022-12-29T21:01:05.190000",
      "content": "<p>One rule of thumb for medical imaging is to always try to split by patient, in order to prevent data leakage between training, validation and test sets.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 2079986,
      "author_name": "Eleftherios Fanioudakis",
      "author_url": "",
      "post_date": "2022-12-29T19:04:40.750000",
      "content": "<p>By patient reflects better the expected submission and seems to be the correct way. <br>\nI only tried by individual images so far though.<br>\nAs for stacking seems to be a good idea, only if you can support higher input sizes to mantain the details, plus to make sure you can do the same on test set without losing patient's samples.</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2079673": "how do we need to split the data, by individual images or by patient, my approach was to stack all the images of each patient and train the model for it ",
    "2080105": "One rule of thumb for medical imaging is to always try to split by patient, in order to prevent data leakage between training, validation and test sets.",
    "2079986": "By patient reflects better the expected submission and seems to be the correct way. \nI only tried by individual images so far though.\nAs for stacking seems to be a good idea, only if you can support higher input sizes to mantain the details, plus to make sure you can do the same on test set without losing patient's samples."
  }
}