{
  "id": 208745,
  "title": "[Idea] Transfer learning from Chexpert?",
  "url": "/competitions/vinbigdata-chest-xray-abnormalities-detection/discussion/208745",
  "author_name": "xhlulu",
  "post_date": "2021-01-04T21:05:42.077000",
  "votes": 6,
  "comment_count": 2,
  "views": 0,
  "content": "<p>This is just an idea that came to mind and I haven't tested, but I'm wondering how well it would work.</p>\n<p>Since models like <code>FasterRCNN</code> support custom backbones (see <a href=\"https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html\" target=\"_blank\">this tutorial</a>), would it be helpful to first train the backbone on x-ray classification (e.g. on the Chexpert dataset) before training on the task of object detection for this competition?</p>\n<p>Happy to hear your thoughts on this.</p>",
  "messages": [
    {
      "id": 1138658,
      "postDate": "2021-01-04T21:05:42.077Z",
      "content": "<p>This is just an idea that came to mind and I haven't tested, but I'm wondering how well it would work.</p>\n<p>Since models like <code>FasterRCNN</code> support custom backbones (see <a href=\"https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html\" target=\"_blank\">this tutorial</a>), would it be helpful to first train the backbone on x-ray classification (e.g. on the Chexpert dataset) before training on the task of object detection for this competition?</p>\n<p>Happy to hear your thoughts on this.</p>",
      "rawMarkdown": "This is just an idea that came to mind and I haven't tested, but I'm wondering how well it would work.\n\nSince models like `FasterRCNN` support custom backbones (see [this tutorial](https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html)), would it be helpful to first train the backbone on x-ray classification (e.g. on the Chexpert dataset) before training on the task of object detection for this competition?\n\nHappy to hear your thoughts on this.",
      "votes": 6
    },
    {
      "id": 1144847,
      "postDate": "2021-01-08T17:50:52.323Z",
      "content": "<p>I was also thinking of this. There already models that give heatmaps for some of these diseases. This <a href=\"https://github.com/jfhealthcare/Chexpert\" target=\"_blank\">one</a> looks really good and might help.</p>",
      "rawMarkdown": "I was also thinking of this. There already models that give heatmaps for some of these diseases. This [one](https://github.com/jfhealthcare/Chexpert) looks really good and might help.",
      "votes": 1
    },
    {
      "id": 3230818,
      "postDate": "2025-06-23T13:24:56.963Z",
      "content": "<p>Hello friends.<br>\nHave you tried this? Have you had any success?<br>\nHow did you do it?</p>",
      "rawMarkdown": "Hello friends.\nHave you tried this? Have you had any success?\nHow did you do it?"
    }
  ],
  "comments": [
    {
      "id": 1144847,
      "author_name": "user",
      "author_url": "",
      "post_date": "2021-01-08T17:50:52.323000",
      "content": "<p>I was also thinking of this. There already models that give heatmaps for some of these diseases. This <a href=\"https://github.com/jfhealthcare/Chexpert\" target=\"_blank\">one</a> looks really good and might help.</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 3230818,
      "author_name": "Vitor Dalto",
      "author_url": "",
      "post_date": "2025-06-23T13:24:56.963000",
      "content": "<p>Hello friends.<br>\nHave you tried this? Have you had any success?<br>\nHow did you do it?</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1138658": "This is just an idea that came to mind and I haven't tested, but I'm wondering how well it would work.\n\nSince models like `FasterRCNN` support custom backbones (see [this tutorial](https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html)), would it be helpful to first train the backbone on x-ray classification (e.g. on the Chexpert dataset) before training on the task of object detection for this competition?\n\nHappy to hear your thoughts on this.",
    "1144847": "I was also thinking of this. There already models that give heatmaps for some of these diseases. This [one](https://github.com/jfhealthcare/Chexpert) looks really good and might help.",
    "3230818": "Hello friends.\nHave you tried this? Have you had any success?\nHow did you do it?"
  }
}