{
  "id": 597912,
  "title": "How Do Radiologists Actually Detect Aneurysms?",
  "url": "/competitions/rsna-intracranial-aneurysm-detection/discussion/597912",
  "author_name": "MT",
  "post_date": "2025-08-07T22:15:38.326000",
  "votes": 24,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi everyone! </p>\n<p>I've been staring at these DICOM images, and I'm honestly puzzled. When I look at the labelled aneurysm locations, they look identical to other bright spots in the same image, and  I'm struggling to understand what makes one white blob an aneurysm. I read around that radiologists need the full 3D context to scroll through multiple slices to track vessel continuity. However, when I check the training data:</p>\n<pre><code>.groupby('SeriesInstanceUID').size().value_counts()\n    \n     \n      \n      \n       \n</code></pre>\n<p>The majority of series have only one slice annotated with an aneurysm.</p>\n<p>Does this mean the aneurysm is only visible on that exact slice?<br>\nOr is this the \"best\" slice, but the aneurysm spans multiple slices?</p>\n<p>Thanks!</p>",
  "messages": [
    {
      "id": 3265735,
      "postDate": "2025-08-07T22:15:38.327Z",
      "content": "<p>Hi everyone! </p>\n<p>I've been staring at these DICOM images, and I'm honestly puzzled. When I look at the labelled aneurysm locations, they look identical to other bright spots in the same image, and  I'm struggling to understand what makes one white blob an aneurysm. I read around that radiologists need the full 3D context to scroll through multiple slices to track vessel continuity. However, when I check the training data:</p>\n<pre><code>.groupby('SeriesInstanceUID').size().value_counts()\n    \n     \n      \n      \n       \n</code></pre>\n<p>The majority of series have only one slice annotated with an aneurysm.</p>\n<p>Does this mean the aneurysm is only visible on that exact slice?<br>\nOr is this the \"best\" slice, but the aneurysm spans multiple slices?</p>\n<p>Thanks!</p>",
      "rawMarkdown": "Hi everyone! \n\nI've been staring at these DICOM images, and I'm honestly puzzled. When I look at the labelled aneurysm locations, they look identical to other bright spots in the same image, and  I'm struggling to understand what makes one white blob an aneurysm. I read around that radiologists need the full 3D context to scroll through multiple slices to track vessel continuity. However, when I check the training data:\n\n```\ntrain_loc.groupby('SeriesInstanceUID').size().value_counts()\n1    1592\n2     223\n3      57\n4      13\n5       5\n```\nThe majority of series have only one slice annotated with an aneurysm.\n\nDoes this mean the aneurysm is only visible on that exact slice?\nOr is this the \"best\" slice, but the aneurysm spans multiple slices?\n\nThanks!",
      "votes": 24
    },
    {
      "id": 3265749,
      "postDate": "2025-08-07T22:46:20.473Z",
      "content": "<p>Great questions!</p>\n<p>When we look for aneurysms in real clinical cases, we scroll through the series in multiple orthogonal planes and look for vessels that have saccular outpouchings that do not continue on to normal vessel (i.e. an aneurysm). We may also look at 3D renderings or maximum intensity projections of the vessels to aid in this purpose. When we see a sac or stump (and not a tube that gradually tapers) then we know its an aneurysm or potentially some other vascular pathologies.</p>\n<p>Regarding the aneurysm location annotations, these are just a single point in (roughly) the center of the aneurysm. We opted not to do bounding boxes or aneurysm segmentations to reduce annotator burden (which was extreme for this dataset). So you can think of this as the \"center\" or \"best\" single slice to see the aneurysm.</p>",
      "rawMarkdown": "Great questions!\n\nWhen we look for aneurysms in real clinical cases, we scroll through the series in multiple orthogonal planes and look for vessels that have saccular outpouchings that do not continue on to normal vessel (i.e. an aneurysm). We may also look at 3D renderings or maximum intensity projections of the vessels to aid in this purpose. When we see a sac or stump (and not a tube that gradually tapers) then we know its an aneurysm or potentially some other vascular pathologies.\n\nRegarding the aneurysm location annotations, these are just a single point in (roughly) the center of the aneurysm. We opted not to do bounding boxes or aneurysm segmentations to reduce annotator burden (which was extreme for this dataset). So you can think of this as the \"center\" or \"best\" single slice to see the aneurysm.",
      "votes": 17
    }
  ],
  "comments": [
    {
      "id": 3265749,
      "author_name": "Evan Calabrese",
      "author_url": "",
      "post_date": "2025-08-07T22:46:20.473000",
      "content": "<p>Great questions!</p>\n<p>When we look for aneurysms in real clinical cases, we scroll through the series in multiple orthogonal planes and look for vessels that have saccular outpouchings that do not continue on to normal vessel (i.e. an aneurysm). We may also look at 3D renderings or maximum intensity projections of the vessels to aid in this purpose. When we see a sac or stump (and not a tube that gradually tapers) then we know its an aneurysm or potentially some other vascular pathologies.</p>\n<p>Regarding the aneurysm location annotations, these are just a single point in (roughly) the center of the aneurysm. We opted not to do bounding boxes or aneurysm segmentations to reduce annotator burden (which was extreme for this dataset). So you can think of this as the \"center\" or \"best\" single slice to see the aneurysm.</p>",
      "votes": 17,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3265735": "Hi everyone! \n\nI've been staring at these DICOM images, and I'm honestly puzzled. When I look at the labelled aneurysm locations, they look identical to other bright spots in the same image, and  I'm struggling to understand what makes one white blob an aneurysm. I read around that radiologists need the full 3D context to scroll through multiple slices to track vessel continuity. However, when I check the training data:\n\n```\ntrain_loc.groupby('SeriesInstanceUID').size().value_counts()\n1    1592\n2     223\n3      57\n4      13\n5       5\n```\nThe majority of series have only one slice annotated with an aneurysm.\n\nDoes this mean the aneurysm is only visible on that exact slice?\nOr is this the \"best\" slice, but the aneurysm spans multiple slices?\n\nThanks!",
    "3265749": "Great questions!\n\nWhen we look for aneurysms in real clinical cases, we scroll through the series in multiple orthogonal planes and look for vessels that have saccular outpouchings that do not continue on to normal vessel (i.e. an aneurysm). We may also look at 3D renderings or maximum intensity projections of the vessels to aid in this purpose. When we see a sac or stump (and not a tube that gradually tapers) then we know its an aneurysm or potentially some other vascular pathologies.\n\nRegarding the aneurysm location annotations, these are just a single point in (roughly) the center of the aneurysm. We opted not to do bounding boxes or aneurysm segmentations to reduce annotator burden (which was extreme for this dataset). So you can think of this as the \"center\" or \"best\" single slice to see the aneurysm."
  }
}