{
  "id": 339642,
  "title": "Starting to think it's not possible with this data",
  "url": "/competitions/mayo-clinic-strip-ai/discussion/339642",
  "author_name": "Iuryck Santos",
  "post_date": "2022-07-25T19:34:35.105000",
  "votes": 6,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Convolutional NNs, EfficientNet, ViTs, ViTs on top of ViTs, all of them stay arround the 50-60% accuracy mark when training<br>\nI've tried changing the data, black background, zoom in, zoom out, barely changes anything. At best the model guesses based on class imbalance.</p>\n<p>Everywhere I look on tasks equal to this, people use data with the image resolution we are using, but the blood sample is 90% of the image, making it possible to see deep detail. If we try to see the same level of detail on these images, you just \"run out of pixels\". If you extract a 1000x1000 image as a sample from these images and look to other examples you will see that for the same sample size that image is around 4kx4k at least. We have images that have 60k pixels in one axis but the actual blood sample has about 5k or 4k.</p>\n<p>Just look at this model: <a href=\"https://www.youtube.com/watch?v=cABkB1J-GTA\" target=\"_blank\">https://www.youtube.com/watch?v=cABkB1J-GTA</a></p>",
  "messages": [
    {
      "id": 1870736,
      "postDate": "2022-07-25T19:34:35.107Z",
      "content": "<p>Convolutional NNs, EfficientNet, ViTs, ViTs on top of ViTs, all of them stay arround the 50-60% accuracy mark when training<br>\nI've tried changing the data, black background, zoom in, zoom out, barely changes anything. At best the model guesses based on class imbalance.</p>\n<p>Everywhere I look on tasks equal to this, people use data with the image resolution we are using, but the blood sample is 90% of the image, making it possible to see deep detail. If we try to see the same level of detail on these images, you just \"run out of pixels\". If you extract a 1000x1000 image as a sample from these images and look to other examples you will see that for the same sample size that image is around 4kx4k at least. We have images that have 60k pixels in one axis but the actual blood sample has about 5k or 4k.</p>\n<p>Just look at this model: <a href=\"https://www.youtube.com/watch?v=cABkB1J-GTA\" target=\"_blank\">https://www.youtube.com/watch?v=cABkB1J-GTA</a></p>",
      "rawMarkdown": "Convolutional NNs, EfficientNet, ViTs, ViTs on top of ViTs, all of them stay arround the 50-60% accuracy mark when training\nI've tried changing the data, black background, zoom in, zoom out, barely changes anything. At best the model guesses based on class imbalance.\n\nEverywhere I look on tasks equal to this, people use data with the image resolution we are using, but the blood sample is 90% of the image, making it possible to see deep detail. If we try to see the same level of detail on these images, you just \"run out of pixels\". If you extract a 1000x1000 image as a sample from these images and look to other examples you will see that for the same sample size that image is around 4kx4k at least. We have images that have 60k pixels in one axis but the actual blood sample has about 5k or 4k.\n\nJust look at this model: https://www.youtube.com/watch?v=cABkB1J-GTA",
      "votes": 6
    },
    {
      "id": 1872361,
      "postDate": "2022-07-26T21:31:17.613Z",
      "content": "<p>As far as I tried, classification with a single image or patch did not work. I also saw similar patterns as you said. </p>",
      "rawMarkdown": "As far as I tried, classification with a single image or patch did not work. I also saw similar patterns as you said. ",
      "votes": 1
    },
    {
      "id": 1905915,
      "postDate": "2022-08-19T12:33:13.707Z",
      "content": "<p>Good point.  Competitions like this one sometimes also suffer from Survivorship Bias - if we only have data from people who survived the stroke we can miss important info- not exactly the same point tho</p>",
      "rawMarkdown": "Good point.  Competitions like this one sometimes also suffer from Survivorship Bias - if we only have data from people who survived the stroke we can miss important info- not exactly the same point tho",
      "votes": 2
    },
    {
      "id": 1918450,
      "postDate": "2022-08-29T15:30:00.727Z",
      "content": "<p>Thanks for the link to HIPT. Just wondering , did you do pretraining (like the author mentioned) on the \"others\" data we get from this competition ?</p>",
      "rawMarkdown": "Thanks for the link to HIPT. Just wondering , did you do pretraining (like the author mentioned) on the \"others\" data we get from this competition ?"
    }
  ],
  "comments": [
    {
      "id": 1872361,
      "author_name": "hjunlee941",
      "author_url": "",
      "post_date": "2022-07-26T21:31:17.613000",
      "content": "<p>As far as I tried, classification with a single image or patch did not work. I also saw similar patterns as you said. </p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1905915,
      "author_name": "mbmlearner",
      "author_url": "",
      "post_date": "2022-08-19T12:33:13.707000",
      "content": "<p>Good point.  Competitions like this one sometimes also suffer from Survivorship Bias - if we only have data from people who survived the stroke we can miss important info- not exactly the same point tho</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 1918450,
      "author_name": "yukiya",
      "author_url": "",
      "post_date": "2022-08-29T15:30:00.727000",
      "content": "<p>Thanks for the link to HIPT. Just wondering , did you do pretraining (like the author mentioned) on the \"others\" data we get from this competition ?</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1870736": "Convolutional NNs, EfficientNet, ViTs, ViTs on top of ViTs, all of them stay arround the 50-60% accuracy mark when training\nI've tried changing the data, black background, zoom in, zoom out, barely changes anything. At best the model guesses based on class imbalance.\n\nEverywhere I look on tasks equal to this, people use data with the image resolution we are using, but the blood sample is 90% of the image, making it possible to see deep detail. If we try to see the same level of detail on these images, you just \"run out of pixels\". If you extract a 1000x1000 image as a sample from these images and look to other examples you will see that for the same sample size that image is around 4kx4k at least. We have images that have 60k pixels in one axis but the actual blood sample has about 5k or 4k.\n\nJust look at this model: https://www.youtube.com/watch?v=cABkB1J-GTA",
    "1872361": "As far as I tried, classification with a single image or patch did not work. I also saw similar patterns as you said. ",
    "1905915": "Good point.  Competitions like this one sometimes also suffer from Survivorship Bias - if we only have data from people who survived the stroke we can miss important info- not exactly the same point tho",
    "1918450": "Thanks for the link to HIPT. Just wondering , did you do pretraining (like the author mentioned) on the \"others\" data we get from this competition ?"
  }
}