{
  "id": 557018,
  "title": "Labeling Errors in Multiple Images in RSNA Dataset",
  "url": "/competitions/rsna-intracranial-hemorrhage-detection/discussion/557018",
  "author_name": "Miguel Herencia",
  "post_date": "2025-01-16T11:15:04.379000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hello everyone,</p>\n<p>While exploring the RSNA Intracranial Hemorrhage Detection dataset, I have identified multiple instances of potential labeling errors. For example, the images with IDs <strong>ID_2b8bee2d0</strong> and <strong>ID_657ec537e</strong> appear to clearly show an intracranial hemorrhage, as confirmed by a radiologist, yet they are labeled as healthy</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F12384089%2F53a9a7e5122ce58c2090e4f9543c192a%2FID_2b8bee2d0.png?generation=1737025899481073&amp;alt=media\" alt=\"\"></p>\n<p>ID_2b8bee2d0_epidural,0<br>\nID_2b8bee2d0_intraparenchymal,0<br>\nID_2b8bee2d0_intraventricular,0<br>\nID_2b8bee2d0_subarachnoid,0<br>\nID_2b8bee2d0_subdural,0<br>\nID_2b8bee2d0_any,0</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F12384089%2F2e6841e41e4a6020fb02047206c24e50%2FID_657ec537e.png?generation=1737025979775242&amp;alt=media\" alt=\"\"></p>\n<p>ID_657ec537e_epidural,0<br>\nID_657ec537e_intraparenchymal,0<br>\nID_657ec537e_intraventricular,0<br>\nID_657ec537e_subarachnoid,0<br>\nID_657ec537e_subdural,0<br>\nID_657ec537e_any,0</p>\n<p>This raises concerns about <strong>other potential misannotations throughout the dataset</strong>, <strong>which could negatively impact model performance and research outcomes</strong></p>\n<p>Has anyone else observed similar inconsistencies? Are there any known efforts to audit and correct these labels, or any best practices to handle such errors during training? A collaborative review or discussion with the dataset organizers might help improve the dataset's reliability</p>",
  "messages": [
    {
      "id": 3098324,
      "postDate": "2025-01-16T11:15:04.380Z",
      "content": "<p>Hello everyone,</p>\n<p>While exploring the RSNA Intracranial Hemorrhage Detection dataset, I have identified multiple instances of potential labeling errors. For example, the images with IDs <strong>ID_2b8bee2d0</strong> and <strong>ID_657ec537e</strong> appear to clearly show an intracranial hemorrhage, as confirmed by a radiologist, yet they are labeled as healthy</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F12384089%2F53a9a7e5122ce58c2090e4f9543c192a%2FID_2b8bee2d0.png?generation=1737025899481073&amp;alt=media\" alt=\"\"></p>\n<p>ID_2b8bee2d0_epidural,0<br>\nID_2b8bee2d0_intraparenchymal,0<br>\nID_2b8bee2d0_intraventricular,0<br>\nID_2b8bee2d0_subarachnoid,0<br>\nID_2b8bee2d0_subdural,0<br>\nID_2b8bee2d0_any,0</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F12384089%2F2e6841e41e4a6020fb02047206c24e50%2FID_657ec537e.png?generation=1737025979775242&amp;alt=media\" alt=\"\"></p>\n<p>ID_657ec537e_epidural,0<br>\nID_657ec537e_intraparenchymal,0<br>\nID_657ec537e_intraventricular,0<br>\nID_657ec537e_subarachnoid,0<br>\nID_657ec537e_subdural,0<br>\nID_657ec537e_any,0</p>\n<p>This raises concerns about <strong>other potential misannotations throughout the dataset</strong>, <strong>which could negatively impact model performance and research outcomes</strong></p>\n<p>Has anyone else observed similar inconsistencies? Are there any known efforts to audit and correct these labels, or any best practices to handle such errors during training? A collaborative review or discussion with the dataset organizers might help improve the dataset's reliability</p>",
      "rawMarkdown": "Hello everyone,\n\nWhile exploring the RSNA Intracranial Hemorrhage Detection dataset, I have identified multiple instances of potential labeling errors. For example, the images with IDs **ID_2b8bee2d0** and **ID_657ec537e** appear to clearly show an intracranial hemorrhage, as confirmed by a radiologist, yet they are labeled as healthy\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F12384089%2F53a9a7e5122ce58c2090e4f9543c192a%2FID_2b8bee2d0.png?generation=1737025899481073&alt=media)\n\nID_2b8bee2d0_epidural,0\nID_2b8bee2d0_intraparenchymal,0\nID_2b8bee2d0_intraventricular,0\nID_2b8bee2d0_subarachnoid,0\nID_2b8bee2d0_subdural,0\nID_2b8bee2d0_any,0\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F12384089%2F2e6841e41e4a6020fb02047206c24e50%2FID_657ec537e.png?generation=1737025979775242&alt=media)\n\nID_657ec537e_epidural,0\nID_657ec537e_intraparenchymal,0\nID_657ec537e_intraventricular,0\nID_657ec537e_subarachnoid,0\nID_657ec537e_subdural,0\nID_657ec537e_any,0\n\nThis raises concerns about **other potential misannotations throughout the dataset**, **which could negatively impact model performance and research outcomes**\n\nHas anyone else observed similar inconsistencies? Are there any known efforts to audit and correct these labels, or any best practices to handle such errors during training? A collaborative review or discussion with the dataset organizers might help improve the dataset's reliability",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3098324": "Hello everyone,\n\nWhile exploring the RSNA Intracranial Hemorrhage Detection dataset, I have identified multiple instances of potential labeling errors. For example, the images with IDs **ID_2b8bee2d0** and **ID_657ec537e** appear to clearly show an intracranial hemorrhage, as confirmed by a radiologist, yet they are labeled as healthy\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F12384089%2F53a9a7e5122ce58c2090e4f9543c192a%2FID_2b8bee2d0.png?generation=1737025899481073&alt=media)\n\nID_2b8bee2d0_epidural,0\nID_2b8bee2d0_intraparenchymal,0\nID_2b8bee2d0_intraventricular,0\nID_2b8bee2d0_subarachnoid,0\nID_2b8bee2d0_subdural,0\nID_2b8bee2d0_any,0\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F12384089%2F2e6841e41e4a6020fb02047206c24e50%2FID_657ec537e.png?generation=1737025979775242&alt=media)\n\nID_657ec537e_epidural,0\nID_657ec537e_intraparenchymal,0\nID_657ec537e_intraventricular,0\nID_657ec537e_subarachnoid,0\nID_657ec537e_subdural,0\nID_657ec537e_any,0\n\nThis raises concerns about **other potential misannotations throughout the dataset**, **which could negatively impact model performance and research outcomes**\n\nHas anyone else observed similar inconsistencies? Are there any known efforts to audit and correct these labels, or any best practices to handle such errors during training? A collaborative review or discussion with the dataset organizers might help improve the dataset's reliability"
  }
}