{
  "id": 444461,
  "title": "Imbalanced dataset",
  "url": "/competitions/rsna-2023-abdominal-trauma-detection/discussion/444461",
  "author_name": "Julien Genzling",
  "post_date": "2023-10-02T06:28:01.758000",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>When training a 2.5D model on cropped scans for kidney, liver or spleen, the model quickly ends up predicting only healthy labels, even with a weighted loss. I tried downsampling but the results weren’t good because there were too few samples.<br>\nWhat are you solutions to solve this problem ?</p>",
  "messages": [
    {
      "id": 2464308,
      "postDate": "2023-10-02T06:28:01.757Z",
      "content": "<p>When training a 2.5D model on cropped scans for kidney, liver or spleen, the model quickly ends up predicting only healthy labels, even with a weighted loss. I tried downsampling but the results weren’t good because there were too few samples.<br>\nWhat are you solutions to solve this problem ?</p>",
      "rawMarkdown": "When training a 2.5D model on cropped scans for kidney, liver or spleen, the model quickly ends up predicting only healthy labels, even with a weighted loss. I tried downsampling but the results weren’t good because there were too few samples.\nWhat are you solutions to solve this problem ?",
      "votes": 1
    },
    {
      "id": 2464506,
      "postDate": "2023-10-02T10:18:32.173Z",
      "content": "<p><a href=\"https://www.kaggle.com/juliengenzling\" target=\"_blank\">@juliengenzling</a> Can you describe the model in detail? I have the same problem</p>",
      "rawMarkdown": "@juliengenzling Can you describe the model in detail? I have the same problem"
    }
  ],
  "comments": [
    {
      "id": 2464506,
      "author_name": "Halley",
      "author_url": "",
      "post_date": "2023-10-02T10:18:32.173000",
      "content": "<p><a href=\"https://www.kaggle.com/juliengenzling\" target=\"_blank\">@juliengenzling</a> Can you describe the model in detail? I have the same problem</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2464308": "When training a 2.5D model on cropped scans for kidney, liver or spleen, the model quickly ends up predicting only healthy labels, even with a weighted loss. I tried downsampling but the results weren’t good because there were too few samples.\nWhat are you solutions to solve this problem ?",
    "2464506": "@juliengenzling Can you describe the model in detail? I have the same problem"
  }
}