{
  "id": 226762,
  "title": "Medical domain knowledge help needed for worst performing classes ( Other lesion,  Calcification,  ILD)",
  "url": "/competitions/vinbigdata-chest-xray-abnormalities-detection/discussion/226762",
  "author_name": "NakedKoala",
  "post_date": "2021-03-17T16:26:22.635000",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<ul>\n<li><p>Other lesion</p></li>\n<li><p>Calcification</p></li>\n<li><p>ILD</p></li>\n</ul>\n<p>These 3 classes have the worst mAP for my model.  I am trying to improve the model performance on theses classes by leveraging hierarchical relationship between classes.</p>\n<p>Can any medical expert / radiologist tell me if the 3 classes above have any parent - child hierarchical relationship with other class label ( such that presence of child symptom implies the presence of parent symptom ). </p>",
  "messages": [
    {
      "id": 1242951,
      "postDate": "2021-03-17T23:19:06.773Z",
      "content": "<p>No significant hierarchical relation for these labels in my opinion for the training dataset.</p>\n<p><strong>Other lesion</strong> is a very variable label not related to any other label.</p>\n<p><strong>Calcification</strong> is usually related to a focal higher density (whiter) for the size of the anomaly. Some calcifications are spatially related to aortic enlargement in the dataset.</p>\n<p>Technically, <strong>ILD</strong> can be a non specific sign of pulmonary fibrosis. But based on the train annotations there is not much significant overlap between these two labels.</p>",
      "rawMarkdown": "No significant hierarchical relation for these labels in my opinion for the training dataset.\n\n**Other lesion** is a very variable label not related to any other label.\n\n**Calcification** is usually related to a focal higher density (whiter) for the size of the anomaly. Some calcifications are spatially related to aortic enlargement in the dataset.\n\nTechnically, **ILD** can be a non specific sign of pulmonary fibrosis. But based on the train annotations there is not much significant overlap between these two labels.\n",
      "votes": 5
    },
    {
      "id": 1242473,
      "postDate": "2021-03-17T16:26:22.637Z",
      "content": "<ul>\n<li><p>Other lesion</p></li>\n<li><p>Calcification</p></li>\n<li><p>ILD</p></li>\n</ul>\n<p>These 3 classes have the worst mAP for my model.  I am trying to improve the model performance on theses classes by leveraging hierarchical relationship between classes.</p>\n<p>Can any medical expert / radiologist tell me if the 3 classes above have any parent - child hierarchical relationship with other class label ( such that presence of child symptom implies the presence of parent symptom ). </p>",
      "rawMarkdown": "* Other lesion\n \n* Calcification\n \n* ILD\n\nThese 3 classes have the worst mAP for my model.  I am trying to improve the model performance on theses classes by leveraging hierarchical relationship between classes.\n\nCan any medical expert / radiologist tell me if the 3 classes above have any parent - child hierarchical relationship with other class label ( such that presence of child symptom implies the presence of parent symptom ). \n\n\n\n\n",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 1242951,
      "author_name": "Alexandre Cadrin-Chênevert",
      "author_url": "",
      "post_date": "2021-03-17T23:19:06.773000",
      "content": "<p>No significant hierarchical relation for these labels in my opinion for the training dataset.</p>\n<p><strong>Other lesion</strong> is a very variable label not related to any other label.</p>\n<p><strong>Calcification</strong> is usually related to a focal higher density (whiter) for the size of the anomaly. Some calcifications are spatially related to aortic enlargement in the dataset.</p>\n<p>Technically, <strong>ILD</strong> can be a non specific sign of pulmonary fibrosis. But based on the train annotations there is not much significant overlap between these two labels.</p>",
      "votes": 5,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1242951": "No significant hierarchical relation for these labels in my opinion for the training dataset.\n\n**Other lesion** is a very variable label not related to any other label.\n\n**Calcification** is usually related to a focal higher density (whiter) for the size of the anomaly. Some calcifications are spatially related to aortic enlargement in the dataset.\n\nTechnically, **ILD** can be a non specific sign of pulmonary fibrosis. But based on the train annotations there is not much significant overlap between these two labels.\n",
    "1242473": "* Other lesion\n \n* Calcification\n \n* ILD\n\nThese 3 classes have the worst mAP for my model.  I am trying to improve the model performance on theses classes by leveraging hierarchical relationship between classes.\n\nCan any medical expert / radiologist tell me if the 3 classes above have any parent - child hierarchical relationship with other class label ( such that presence of child symptom implies the presence of parent symptom ). \n\n\n\n\n"
  }
}