{
  "id": 110507,
  "title": "Approach for Better  Accuracy",
  "url": "/competitions/rsna-intracranial-hemorrhage-detection/discussion/110507",
  "author_name": "Vinay Turpati",
  "post_date": "2019-09-28T14:30:40.699000",
  "votes": 8,
  "comment_count": 0,
  "views": 0,
  "content": "<ol>\n<li><p>Search all data which you can use to train the model from different competitions.</p></li>\n<li><p>Look for similar problem statements and what standard procedures have been implemented to solve them. I think for most of medical datasets I would go for Chexnet.</p></li>\n<li><p>By visualising the output label distribution, check the presence of unbalanced classes. Generate more images for classes, where the number of examples was lower than a given threshold.</p></li>\n<li><p>Maintain your work logs: The most important thing during the whole competition is to maintain a file with all input parameters and the corresponding output values.</p></li>\n<li><p>Size of the image: Having big images increases the neural network complexity, but we don’t want to lose important details. - Pay attention to the relation between your image size and model performance. Try different sizes , choose the best one that works for you.</p></li>\n<li><p>Transfer learning to create model as this could be more efficient.</p></li>\n<li><p>Instead of working on the whole dataset, work on a sample of the whole dataset to check if your model is behaving predictably.</p></li>\n</ol>",
  "messages": [
    {
      "id": 635964,
      "postDate": "2019-09-28T14:30:40.700Z",
      "content": "<ol>\n<li><p>Search all data which you can use to train the model from different competitions.</p></li>\n<li><p>Look for similar problem statements and what standard procedures have been implemented to solve them. I think for most of medical datasets I would go for Chexnet.</p></li>\n<li><p>By visualising the output label distribution, check the presence of unbalanced classes. Generate more images for classes, where the number of examples was lower than a given threshold.</p></li>\n<li><p>Maintain your work logs: The most important thing during the whole competition is to maintain a file with all input parameters and the corresponding output values.</p></li>\n<li><p>Size of the image: Having big images increases the neural network complexity, but we don’t want to lose important details. - Pay attention to the relation between your image size and model performance. Try different sizes , choose the best one that works for you.</p></li>\n<li><p>Transfer learning to create model as this could be more efficient.</p></li>\n<li><p>Instead of working on the whole dataset, work on a sample of the whole dataset to check if your model is behaving predictably.</p></li>\n</ol>",
      "rawMarkdown": "1. Search all data which you can use to train the model from different competitions.\n\n2. Look for similar problem statements and what standard procedures have been implemented to solve them. I think for most of medical datasets I would go for Chexnet.\n\n3. By visualising the output label distribution, check the presence of unbalanced classes. Generate more images for classes, where the number of examples was lower than a given threshold.\n\n6. Maintain your work logs: The most important thing during the whole competition is to maintain a file with all input parameters and the corresponding output values.\n\n7. Size of the image: Having big images increases the neural network complexity, but we don’t want to lose important details. - Pay attention to the relation between your image size and model performance. Try different sizes , choose the best one that works for you.\n\n8. Transfer learning to create model as this could be more efficient.\n\n9.  Instead of working on the whole dataset, work on a sample of the whole dataset to check if your model is behaving predictably.\n",
      "votes": 8
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "635964": "1. Search all data which you can use to train the model from different competitions.\n\n2. Look for similar problem statements and what standard procedures have been implemented to solve them. I think for most of medical datasets I would go for Chexnet.\n\n3. By visualising the output label distribution, check the presence of unbalanced classes. Generate more images for classes, where the number of examples was lower than a given threshold.\n\n6. Maintain your work logs: The most important thing during the whole competition is to maintain a file with all input parameters and the corresponding output values.\n\n7. Size of the image: Having big images increases the neural network complexity, but we don’t want to lose important details. - Pay attention to the relation between your image size and model performance. Try different sizes , choose the best one that works for you.\n\n8. Transfer learning to create model as this could be more efficient.\n\n9.  Instead of working on the whole dataset, work on a sample of the whole dataset to check if your model is behaving predictably.\n"
  }
}