{
  "id": 50095,
  "title": "concerns over training masks and scoring validity",
  "url": "/competitions/data-science-bowl-2018/discussion/50095",
  "author_name": "",
  "post_date": "2018-02-20T00:29:25.146212700Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Browsing through the training masks, I can see the mistakes that are being made are typical of a segmentation routine.  Specifically, I see many examples of false detection of small nuclei (likely from a local threshold algorithm).  If submissions are to be judged by how well they compare to the routine that created these masks then the winning entry will be no better.  What am I missing here?</p>\n\n<p>For those of you applying deep learning, be aware that the false positive rate in the training set is quite high while the false negative rate is much lower.</p>",
  "messages": [
    {
      "id": "285329",
      "postDate": "02/20/2018 00:29:25",
      "content": "<p>Browsing through the training masks, I can see the mistakes that are being made are typical of a segmentation routine.  Specifically, I see many examples of false detection of small nuclei (likely from a local threshold algorithm).  If submissions are to be judged by how well they compare to the routine that created these masks then the winning entry will be no better.  What am I missing here?</p>\n\n<p>For those of you applying deep learning, be aware that the false positive rate in the training set is quite high while the false negative rate is much lower.</p>",
      "rawMarkdown": "Browsing through the training masks, I can see the mistakes that are being made are typical of a segmentation routine.  Specifically, I see many examples of false detection of small nuclei (likely from a local threshold algorithm).  If submissions are to be judged by how well they compare to the routine that created these masks then the winning entry will be no better.  What am I missing here?\n\nFor those of you applying deep learning, be aware that the false positive rate in the training set is quite high while the false negative rate is much lower.",
      "votes": null
    },
    {
      "id": "285358",
      "postDate": "02/20/2018 03:22:59",
      "content": "<p>A thresholding method wasn’t used. Each of the annotations was annotated by hand.</p>",
      "rawMarkdown": "A thresholding method wasn’t used. Each of the annotations was annotated by hand.",
      "votes": null
    },
    {
      "id": "285647",
      "postDate": "02/20/2018 19:24:18",
      "content": "<p>Yeah. \nMichael, if you see some small objects they are likely real (often called micronuclei, which we decided to include in the competition). Sometimes the brightness of your screen can reveal that things that initially look like mistakes are actually real. Hope this helps, if you see anything that looks really off feel free to post it to the sticky thread about annotation errors.</p>",
      "rawMarkdown": "Yeah. \nMichael, if you see some small objects they are likely real (often called micronuclei, which we decided to include in the competition). Sometimes the brightness of your screen can reveal that things that initially look like mistakes are actually real. Hope this helps, if you see anything that looks really off feel free to post it to the sticky thread about annotation errors.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 285358,
      "author_name": "agoodman",
      "author_url": "",
      "post_date": "02/20/2018 03:22:59",
      "content": "<p>A thresholding method wasn’t used. Each of the annotations was annotated by hand.</p>",
      "votes": null,
      "replies": [
        {
          "id": 285647,
          "author_name": "drannecarpenter",
          "author_url": "",
          "post_date": "02/20/2018 19:24:18",
          "content": "<p>Yeah. \nMichael, if you see some small objects they are likely real (often called micronuclei, which we decided to include in the competition). Sometimes the brightness of your screen can reveal that things that initially look like mistakes are actually real. Hope this helps, if you see anything that looks really off feel free to post it to the sticky thread about annotation errors.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "285329": "Browsing through the training masks, I can see the mistakes that are being made are typical of a segmentation routine.  Specifically, I see many examples of false detection of small nuclei (likely from a local threshold algorithm).  If submissions are to be judged by how well they compare to the routine that created these masks then the winning entry will be no better.  What am I missing here?\n\nFor those of you applying deep learning, be aware that the false positive rate in the training set is quite high while the false negative rate is much lower.",
    "285358": "A thresholding method wasn’t used. Each of the annotations was annotated by hand.",
    "285647": "Yeah. \nMichael, if you see some small objects they are likely real (often called micronuclei, which we decided to include in the competition). Sometimes the brightness of your screen can reveal that things that initially look like mistakes are actually real. Hope this helps, if you see anything that looks really off feel free to post it to the sticky thread about annotation errors."
  },
  "source": "meta"
}