{
  "id": 440537,
  "title": "Has anyone added meta data to the model?",
  "url": "/competitions/rsna-2023-abdominal-trauma-detection/discussion/440537",
  "author_name": "Andrij",
  "post_date": "2023-09-15T09:33:42.456000",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hello everybody! <br>\nI added metadata to the model, other things being equal, but my val loss got worse. <br>\nHas anyone else tried adding meta data to the model as a separate input?<br>\nIf so, what results did you get?<br>\nHere is the list of columns that I added:<br>\n ['BitsStored',<br>\n 'HighBit',<br>\n 'KVP',<br>\n 'PatientPosition',<br>\n 'PixelRepresentation',<br>\n 'RescaleIntercept',<br>\n 'SliceThickness',<br>\n 'ImageOrientationPatient0',<br>\n 'ImageOrientationPatient1',<br>\n 'ImageOrientationPatient2',<br>\n 'ImageOrientationPatient3',<br>\n 'ImageOrientationPatient4',<br>\n 'ImageOrientationPatient5',<br>\n 'ImagePositionPatient0',<br>\n 'ImagePositionPatient1',<br>\n 'ImagePositionPatient2',<br>\n 'PixelSpacing0',<br>\n 'PixelSpacing1',<br>\n 'aortic_hu']</p>",
  "messages": [
    {
      "id": 2441340,
      "postDate": "2023-09-16T06:44:01.803Z",
      "content": "<p>Not sure if you are combining train_series_meta and train_dicom_tags ?  Perhaps not useful for including in the model, but for preprocessing data and selection of series and slices for training -</p>\n<p>In terms of train_series_meta, column aortic_hu could be useful <br>\nfrom the host see this post:<br>\n<a href=\"https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/427226\" target=\"_blank\">https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/427226</a></p>\n<blockquote>\n  <p>One way to determine phases in a multiphasic CT is to measure the density of blood in the abdominal aorta in Hounsfield units. This is provided in the [train/test]_series_meta.csv file in the aortic_hu column. For a multiphasic CT scan, the higher value typically indicates the late arterial phase. </p>\n</blockquote>\n<p>and </p>\n<blockquote>\n  <p>Late arterial phase: Images are taken approximately 35 seconds after injecting IV contrast. The phase helps to evaluate the blood vessels and find areas of active bleeding. </p>\n</blockquote>\n<p>also the incomplete organ column</p>\n<blockquote>\n  <p>The portal venous phase and split bolus CTs in the challenge dataset include the entire abdomen and pelvis, and potentially parts of the chest. The late arterial phase images have variable coverage of the abdomen and pelvis. We decided to keep late arterial phase images even if they only partially covered the solid abdominal organs (i.e. liver, spleen, and kidney) and highlighted them by the incomplete_organ column in the CSV file. The incomplete_organ label is only provided for CT scans in the training set.</p>\n</blockquote>",
      "rawMarkdown": "Not sure if you are combining train_series_meta and train_dicom_tags ?  Perhaps not useful for including in the model, but for preprocessing data and selection of series and slices for training -\n\nIn terms of train_series_meta, column aortic_hu could be useful \nfrom the host see this post:\nhttps://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/427226\n\n>One way to determine phases in a multiphasic CT is to measure the density of blood in the abdominal aorta in Hounsfield units. This is provided in the [train/test]_series_meta.csv file in the aortic_hu column. For a multiphasic CT scan, the higher value typically indicates the late arterial phase. \n\n\nand \n>Late arterial phase: Images are taken approximately 35 seconds after injecting IV contrast. The phase helps to evaluate the blood vessels and find areas of active bleeding. \n\nalso the incomplete organ column\n\n>The portal venous phase and split bolus CTs in the challenge dataset include the entire abdomen and pelvis, and potentially parts of the chest. The late arterial phase images have variable coverage of the abdomen and pelvis. We decided to keep late arterial phase images even if they only partially covered the solid abdominal organs (i.e. liver, spleen, and kidney) and highlighted them by the incomplete_organ column in the CSV file. The incomplete_organ label is only provided for CT scans in the training set.",
      "votes": 1
    },
    {
      "id": 2440127,
      "postDate": "2023-09-15T09:33:42.457Z",
      "content": "<p>Hello everybody! <br>\nI added metadata to the model, other things being equal, but my val loss got worse. <br>\nHas anyone else tried adding meta data to the model as a separate input?<br>\nIf so, what results did you get?<br>\nHere is the list of columns that I added:<br>\n ['BitsStored',<br>\n 'HighBit',<br>\n 'KVP',<br>\n 'PatientPosition',<br>\n 'PixelRepresentation',<br>\n 'RescaleIntercept',<br>\n 'SliceThickness',<br>\n 'ImageOrientationPatient0',<br>\n 'ImageOrientationPatient1',<br>\n 'ImageOrientationPatient2',<br>\n 'ImageOrientationPatient3',<br>\n 'ImageOrientationPatient4',<br>\n 'ImageOrientationPatient5',<br>\n 'ImagePositionPatient0',<br>\n 'ImagePositionPatient1',<br>\n 'ImagePositionPatient2',<br>\n 'PixelSpacing0',<br>\n 'PixelSpacing1',<br>\n 'aortic_hu']</p>",
      "rawMarkdown": "Hello everybody! \nI added metadata to the model, other things being equal, but my val loss got worse. \nHas anyone else tried adding meta data to the model as a separate input?\nIf so, what results did you get?\nHere is the list of columns that I added:\n ['BitsStored',\n 'HighBit',\n 'KVP',\n 'PatientPosition',\n 'PixelRepresentation',\n 'RescaleIntercept',\n 'SliceThickness',\n 'ImageOrientationPatient0',\n 'ImageOrientationPatient1',\n 'ImageOrientationPatient2',\n 'ImageOrientationPatient3',\n 'ImageOrientationPatient4',\n 'ImageOrientationPatient5',\n 'ImagePositionPatient0',\n 'ImagePositionPatient1',\n 'ImagePositionPatient2',\n 'PixelSpacing0',\n 'PixelSpacing1',\n 'aortic_hu']",
      "votes": 1
    },
    {
      "id": 2440334,
      "postDate": "2023-09-15T12:47:35.453Z",
      "content": "<p>I have not used the metadata, but I don't think it would contain any kind of signal useful to this competition. Almost all of those are scan parameters that come from a pre-programmed protocol and would be the same for each patient/scan.</p>",
      "rawMarkdown": "I have not used the metadata, but I don't think it would contain any kind of signal useful to this competition. Almost all of those are scan parameters that come from a pre-programmed protocol and would be the same for each patient/scan.",
      "votes": 2
    }
  ],
  "comments": [
    {
      "id": 2441340,
      "author_name": "something4kag",
      "author_url": "",
      "post_date": "2023-09-16T06:44:01.803000",
      "content": "<p>Not sure if you are combining train_series_meta and train_dicom_tags ?  Perhaps not useful for including in the model, but for preprocessing data and selection of series and slices for training -</p>\n<p>In terms of train_series_meta, column aortic_hu could be useful <br>\nfrom the host see this post:<br>\n<a href=\"https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/427226\" target=\"_blank\">https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/427226</a></p>\n<blockquote>\n  <p>One way to determine phases in a multiphasic CT is to measure the density of blood in the abdominal aorta in Hounsfield units. This is provided in the [train/test]_series_meta.csv file in the aortic_hu column. For a multiphasic CT scan, the higher value typically indicates the late arterial phase. </p>\n</blockquote>\n<p>and </p>\n<blockquote>\n  <p>Late arterial phase: Images are taken approximately 35 seconds after injecting IV contrast. The phase helps to evaluate the blood vessels and find areas of active bleeding. </p>\n</blockquote>\n<p>also the incomplete organ column</p>\n<blockquote>\n  <p>The portal venous phase and split bolus CTs in the challenge dataset include the entire abdomen and pelvis, and potentially parts of the chest. The late arterial phase images have variable coverage of the abdomen and pelvis. We decided to keep late arterial phase images even if they only partially covered the solid abdominal organs (i.e. liver, spleen, and kidney) and highlighted them by the incomplete_organ column in the CSV file. The incomplete_organ label is only provided for CT scans in the training set.</p>\n</blockquote>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 2440334,
      "author_name": "David Roberts",
      "author_url": "",
      "post_date": "2023-09-15T12:47:35.453000",
      "content": "<p>I have not used the metadata, but I don't think it would contain any kind of signal useful to this competition. Almost all of those are scan parameters that come from a pre-programmed protocol and would be the same for each patient/scan.</p>",
      "votes": 2,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2441340": "Not sure if you are combining train_series_meta and train_dicom_tags ?  Perhaps not useful for including in the model, but for preprocessing data and selection of series and slices for training -\n\nIn terms of train_series_meta, column aortic_hu could be useful \nfrom the host see this post:\nhttps://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/427226\n\n>One way to determine phases in a multiphasic CT is to measure the density of blood in the abdominal aorta in Hounsfield units. This is provided in the [train/test]_series_meta.csv file in the aortic_hu column. For a multiphasic CT scan, the higher value typically indicates the late arterial phase. \n\n\nand \n>Late arterial phase: Images are taken approximately 35 seconds after injecting IV contrast. The phase helps to evaluate the blood vessels and find areas of active bleeding. \n\nalso the incomplete organ column\n\n>The portal venous phase and split bolus CTs in the challenge dataset include the entire abdomen and pelvis, and potentially parts of the chest. The late arterial phase images have variable coverage of the abdomen and pelvis. We decided to keep late arterial phase images even if they only partially covered the solid abdominal organs (i.e. liver, spleen, and kidney) and highlighted them by the incomplete_organ column in the CSV file. The incomplete_organ label is only provided for CT scans in the training set.",
    "2440127": "Hello everybody! \nI added metadata to the model, other things being equal, but my val loss got worse. \nHas anyone else tried adding meta data to the model as a separate input?\nIf so, what results did you get?\nHere is the list of columns that I added:\n ['BitsStored',\n 'HighBit',\n 'KVP',\n 'PatientPosition',\n 'PixelRepresentation',\n 'RescaleIntercept',\n 'SliceThickness',\n 'ImageOrientationPatient0',\n 'ImageOrientationPatient1',\n 'ImageOrientationPatient2',\n 'ImageOrientationPatient3',\n 'ImageOrientationPatient4',\n 'ImageOrientationPatient5',\n 'ImagePositionPatient0',\n 'ImagePositionPatient1',\n 'ImagePositionPatient2',\n 'PixelSpacing0',\n 'PixelSpacing1',\n 'aortic_hu']",
    "2440334": "I have not used the metadata, but I don't think it would contain any kind of signal useful to this competition. Almost all of those are scan parameters that come from a pre-programmed protocol and would be the same for each patient/scan."
  }
}