{
  "id": 208175,
  "title": "Medical considerations ",
  "url": "/competitions/vinbigdata-chest-xray-abnormalities-detection/discussion/208175",
  "author_name": "Tanguy Perennec",
  "post_date": "2021-01-02T10:34:56.056000",
  "votes": 9,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi everyone,</p>\n<p>When looking at the annotated images with a medical background, it seems important to integrate some informations into our models.</p>\n<p>First of all, the \"other lesions\" are very different from each other and often concern regions outside the thorax (thyroid or breast). Given the number of \"other lesions\" boxes, it is very likely that the competition will be played on this. But do you think our model will be able to perform on this particular category ?</p>\n<p>Then, some specific lesions make the existence of other lesions more likely, unlikely or ever impossible. For example, the existence of cardiomegaly makes the existence of an interstitial syndrome more likely and the presence of a pneumothorax less likely. The presence of a pneumothorax makes the presence of most other lesions impossible in the same box because the lung is repelled. Finally, the presence of atelectasis or pneumothorax distorts the assessment of the cardiothoracic index to assess the presence of cardiomegaly (since one lung lacks air) and the model may not have enough examples to learn this. All this information and many others are, in my opinion, are not enough in number for learning the model. Maybe it could be useful  to add it in post-processing?</p>\n<p>What do you think ?</p>",
  "messages": [
    {
      "id": 1135545,
      "postDate": "2021-01-02T10:34:56.057Z",
      "content": "<p>Hi everyone,</p>\n<p>When looking at the annotated images with a medical background, it seems important to integrate some informations into our models.</p>\n<p>First of all, the \"other lesions\" are very different from each other and often concern regions outside the thorax (thyroid or breast). Given the number of \"other lesions\" boxes, it is very likely that the competition will be played on this. But do you think our model will be able to perform on this particular category ?</p>\n<p>Then, some specific lesions make the existence of other lesions more likely, unlikely or ever impossible. For example, the existence of cardiomegaly makes the existence of an interstitial syndrome more likely and the presence of a pneumothorax less likely. The presence of a pneumothorax makes the presence of most other lesions impossible in the same box because the lung is repelled. Finally, the presence of atelectasis or pneumothorax distorts the assessment of the cardiothoracic index to assess the presence of cardiomegaly (since one lung lacks air) and the model may not have enough examples to learn this. All this information and many others are, in my opinion, are not enough in number for learning the model. Maybe it could be useful  to add it in post-processing?</p>\n<p>What do you think ?</p>",
      "rawMarkdown": "Hi everyone,\n\nWhen looking at the annotated images with a medical background, it seems important to integrate some informations into our models.\n \nFirst of all, the \"other lesions\" are very different from each other and often concern regions outside the thorax (thyroid or breast). Given the number of \"other lesions\" boxes, it is very likely that the competition will be played on this. But do you think our model will be able to perform on this particular category ?\n\nThen, some specific lesions make the existence of other lesions more likely, unlikely or ever impossible. For example, the existence of cardiomegaly makes the existence of an interstitial syndrome more likely and the presence of a pneumothorax less likely. The presence of a pneumothorax makes the presence of most other lesions impossible in the same box because the lung is repelled. Finally, the presence of atelectasis or pneumothorax distorts the assessment of the cardiothoracic index to assess the presence of cardiomegaly (since one lung lacks air) and the model may not have enough examples to learn this. All this information and many others are, in my opinion, are not enough in number for learning the model. Maybe it could be useful  to add it in post-processing?\n\nWhat do you think ?",
      "votes": 9
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1135545": "Hi everyone,\n\nWhen looking at the annotated images with a medical background, it seems important to integrate some informations into our models.\n \nFirst of all, the \"other lesions\" are very different from each other and often concern regions outside the thorax (thyroid or breast). Given the number of \"other lesions\" boxes, it is very likely that the competition will be played on this. But do you think our model will be able to perform on this particular category ?\n\nThen, some specific lesions make the existence of other lesions more likely, unlikely or ever impossible. For example, the existence of cardiomegaly makes the existence of an interstitial syndrome more likely and the presence of a pneumothorax less likely. The presence of a pneumothorax makes the presence of most other lesions impossible in the same box because the lung is repelled. Finally, the presence of atelectasis or pneumothorax distorts the assessment of the cardiothoracic index to assess the presence of cardiomegaly (since one lung lacks air) and the model may not have enough examples to learn this. All this information and many others are, in my opinion, are not enough in number for learning the model. Maybe it could be useful  to add it in post-processing?\n\nWhat do you think ?"
  }
}