{
  "id": 48317,
  "title": "how to improve your unet",
  "url": "/competitions/data-science-bowl-2018/discussion/48317",
  "author_name": "hengck23",
  "post_date": "2018-01-26T05:56:44.479000",
  "votes": 22,
  "comment_count": 6,
  "views": 0,
  "content": "<p>see attached pptx. All results are on LB test set</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/274242/8376/multi.png\" alt=\"enter image description here\"></p>",
  "messages": [
    {
      "id": 274242,
      "postDate": "2018-01-26T05:56:44.480Z",
      "content": "<p>see attached pptx. All results are on LB test set</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/274242/8376/multi.png\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "see attached pptx. All results are on LB test set\n\n ![enter image description here][1]\n\n\n  [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/274242/8376/multi.png",
      "votes": 22
    },
    {
      "id": 276102,
      "postDate": "2018-01-30T17:42:52.953Z",
      "content": "<p>Interesting solution, great job! If I understand correctly we split masks by nuclei properties detected (color, shape, size) on training data preparation step and train on multiclass UNet? I'm stuck at 0.338, but one major challenge remains, how to split nuclei masks (I'm thinking about watershed algorithm) but I believe that your solution will also help as nuclei with different shapes/sizes will give different output classes. </p>",
      "rawMarkdown": "Interesting solution, great job! If I understand correctly we split masks by nuclei properties detected (color, shape, size) on training data preparation step and train on multiclass UNet? I'm stuck at 0.338, but one major challenge remains, how to split nuclei masks (I'm thinking about watershed algorithm) but I believe that your solution will also help as nuclei with different shapes/sizes will give different output classes. ",
      "replies": [
        {
          "id": 276343,
          "postDate": "2018-01-31T09:12:49.637Z",
          "content": "<p>I have tested your solution with 3 classes (image modality). I can confirm that this increases score by approx 10% (local mean IoU). I'll test it with 3 more classes for nuclei size (small, medium, large) and report back. Thank you for sharing this idea, this way we can implement more robust solution for future (stage2) image modalities by simple unsupervised color classification of nuclei and background.</p>",
          "rawMarkdown": "I have tested your solution with 3 classes (image modality). I can confirm that this increases score by approx 10% (local mean IoU). I'll test it with 3 more classes for nuclei size (small, medium, large) and report back. Thank you for sharing this idea, this way we can implement more robust solution for future (stage2) image modalities by simple unsupervised color classification of nuclei and background."
        },
        {
          "id": 276849,
          "postDate": "2018-02-01T17:11:19.093Z",
          "content": "<p>@Bostjan Mrak: what do you mean by 3 classes (image modality) pls? as I guess, you split masks into 3 different categories for nuclei color on training data preparation? </p>",
          "rawMarkdown": "@Bostjan Mrak: what do you mean by 3 classes (image modality) pls? as I guess, you split masks into 3 different categories for nuclei color on training data preparation? ",
          "votes": 1
        }
      ]
    },
    {
      "id": 276247,
      "postDate": "2018-01-31T03:02:09.850Z",
      "rawMarkdown": "",
      "isDeleted": true
    },
    {
      "id": 274579,
      "postDate": "2018-01-26T20:17:02.387Z",
      "content": "<p>thanks!</p>",
      "rawMarkdown": "thanks!"
    },
    {
      "id": 274428,
      "postDate": "2018-01-26T15:38:44.840Z",
      "content": "<p>Thanks for sharing.</p>",
      "rawMarkdown": "Thanks for sharing."
    }
  ],
  "comments": [
    {
      "id": 276102,
      "author_name": "Bostjan Mrak",
      "author_url": "",
      "post_date": "2018-01-30T17:42:52.953000",
      "content": "<p>Interesting solution, great job! If I understand correctly we split masks by nuclei properties detected (color, shape, size) on training data preparation step and train on multiclass UNet? I'm stuck at 0.338, but one major challenge remains, how to split nuclei masks (I'm thinking about watershed algorithm) but I believe that your solution will also help as nuclei with different shapes/sizes will give different output classes. </p>",
      "votes": 0,
      "replies": [
        {
          "id": 276343,
          "author_name": "Bostjan Mrak",
          "author_url": "",
          "post_date": "2018-01-31T09:12:49.637000",
          "content": "<p>I have tested your solution with 3 classes (image modality). I can confirm that this increases score by approx 10% (local mean IoU). I'll test it with 3 more classes for nuclei size (small, medium, large) and report back. Thank you for sharing this idea, this way we can implement more robust solution for future (stage2) image modalities by simple unsupervised color classification of nuclei and background.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 276849,
          "author_name": "Tu Nguyen",
          "author_url": "",
          "post_date": "2018-02-01T17:11:19.093000",
          "content": "<p>@Bostjan Mrak: what do you mean by 3 classes (image modality) pls? as I guess, you split masks into 3 different categories for nuclei color on training data preparation? </p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 276247,
      "author_name": "",
      "author_url": "",
      "post_date": "2018-01-31T03:02:09.850000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 274579,
      "author_name": "Zhijian Li",
      "author_url": "",
      "post_date": "2018-01-26T20:17:02.387000",
      "content": "<p>thanks!</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 274428,
      "author_name": "Josan",
      "author_url": "",
      "post_date": "2018-01-26T15:38:44.840000",
      "content": "<p>Thanks for sharing.</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "274242": "see attached pptx. All results are on LB test set\n\n ![enter image description here][1]\n\n\n  [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/274242/8376/multi.png",
    "276102": "Interesting solution, great job! If I understand correctly we split masks by nuclei properties detected (color, shape, size) on training data preparation step and train on multiclass UNet? I'm stuck at 0.338, but one major challenge remains, how to split nuclei masks (I'm thinking about watershed algorithm) but I believe that your solution will also help as nuclei with different shapes/sizes will give different output classes. ",
    "276247": "",
    "274579": "thanks!",
    "274428": "Thanks for sharing."
  }
}