{
  "id": 458710,
  "title": "To make patches or not to?",
  "url": "/competitions/UBC-OCEAN/discussion/458710",
  "author_name": "Aditya Parikh",
  "post_date": "2023-12-01T08:06:13.383000",
  "votes": 2,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I see many approaches used tiling or simple making patches of the larger image and then train each instance. What is your opinion regarding this?</p>",
  "messages": [
    {
      "id": 2549807,
      "postDate": "2023-12-05T14:32:30.513Z",
      "content": "<p>I think they are also critical as the cancer type is visible mainly in detail <a href=\"https://www.kaggle.com/code/jirkaborovec/cancer-subtype-eda-load-wsi-segmentation\" target=\"_blank\">see illustration from litearture</a>.<br>\nIn fact, I got a boost with tiles/patches more then 0.15</p>",
      "rawMarkdown": "I think they are also critical as the cancer type is visible mainly in detail [see illustration from litearture](https://www.kaggle.com/code/jirkaborovec/cancer-subtype-eda-load-wsi-segmentation).\nIn fact, I got a boost with tiles/patches more then 0.15",
      "votes": 1
    },
    {
      "id": 2545027,
      "postDate": "2023-12-01T08:39:39.253Z",
      "content": "<p>I have achieved better results with the patches method. But still the difference between the validation score (&gt;70% bal. acc.) and the leaderboard score (34%) is huge. </p>",
      "rawMarkdown": "I have achieved better results with the patches method. But still the difference between the validation score (>70% bal. acc.) and the leaderboard score (34%) is huge. ",
      "votes": 2
    },
    {
      "id": 2544968,
      "postDate": "2023-12-01T08:06:13.383Z",
      "content": "<p>I see many approaches used tiling or simple making patches of the larger image and then train each instance. What is your opinion regarding this?</p>",
      "rawMarkdown": "I see many approaches used tiling or simple making patches of the larger image and then train each instance. What is your opinion regarding this?",
      "votes": 2
    },
    {
      "id": 2548811,
      "postDate": "2023-12-04T18:00:11.360Z",
      "content": "<p>Patches are absolutely necessary for this project. The thumbnails just don't have enough resolution to be able to have the model learn visual features that would translate to the TMA images, which comprise more than half of the test dataset.</p>",
      "rawMarkdown": "Patches are absolutely necessary for this project. The thumbnails just don't have enough resolution to be able to have the model learn visual features that would translate to the TMA images, which comprise more than half of the test dataset."
    },
    {
      "id": 2547065,
      "postDate": "2023-12-03T06:26:43.700Z",
      "content": "<p>I also adapted patches. We have to predict the type of TMA images using WSI images.<br>\nBy making patches, we can make the difference between them lower.</p>",
      "rawMarkdown": "I also adapted patches. We have to predict the type of TMA images using WSI images.\nBy making patches, we can make the difference between them lower."
    }
  ],
  "comments": [
    {
      "id": 2549807,
      "author_name": "Jirka",
      "author_url": "",
      "post_date": "2023-12-05T14:32:30.513000",
      "content": "<p>I think they are also critical as the cancer type is visible mainly in detail <a href=\"https://www.kaggle.com/code/jirkaborovec/cancer-subtype-eda-load-wsi-segmentation\" target=\"_blank\">see illustration from litearture</a>.<br>\nIn fact, I got a boost with tiles/patches more then 0.15</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 2545027,
      "author_name": "Manuel K",
      "author_url": "",
      "post_date": "2023-12-01T08:39:39.253000",
      "content": "<p>I have achieved better results with the patches method. But still the difference between the validation score (&gt;70% bal. acc.) and the leaderboard score (34%) is huge. </p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 2548811,
      "author_name": "romstefano",
      "author_url": "",
      "post_date": "2023-12-04T18:00:11.360000",
      "content": "<p>Patches are absolutely necessary for this project. The thumbnails just don't have enough resolution to be able to have the model learn visual features that would translate to the TMA images, which comprise more than half of the test dataset.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 2547065,
      "author_name": "Taro Kuroda",
      "author_url": "",
      "post_date": "2023-12-03T06:26:43.700000",
      "content": "<p>I also adapted patches. We have to predict the type of TMA images using WSI images.<br>\nBy making patches, we can make the difference between them lower.</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2549807": "I think they are also critical as the cancer type is visible mainly in detail [see illustration from litearture](https://www.kaggle.com/code/jirkaborovec/cancer-subtype-eda-load-wsi-segmentation).\nIn fact, I got a boost with tiles/patches more then 0.15",
    "2545027": "I have achieved better results with the patches method. But still the difference between the validation score (>70% bal. acc.) and the leaderboard score (34%) is huge. ",
    "2544968": "I see many approaches used tiling or simple making patches of the larger image and then train each instance. What is your opinion regarding this?",
    "2548811": "Patches are absolutely necessary for this project. The thumbnails just don't have enough resolution to be able to have the model learn visual features that would translate to the TMA images, which comprise more than half of the test dataset.",
    "2547065": "I also adapted patches. We have to predict the type of TMA images using WSI images.\nBy making patches, we can make the difference between them lower."
  }
}