{
  "id": 65300,
  "title": "Feature Engineering Ideas",
  "url": "/competitions/rsna-pneumonia-detection-challenge/discussion/65300",
  "author_name": "",
  "post_date": "2018-09-09T00:27:31.089429700Z",
  "votes": 7,
  "comment_count": 7,
  "views": 0,
  "content": "<p>There have been some good ideas expressed in other discussions on how to incorporate the DICOM metadata. Here are my ideas on the visually-based features:</p>\n\n<ul>\n<li>Localize and Crop Lungs,</li>\n<li>Analyze Gray-Scale Histograms,</li>\n<li>Clustering,</li>\n<li>and a plethora of other ideas on creating synthetic features.</li>\n</ul>\n\n<p>What are your thoughts?</p>\n\n<p><img src=\"https://raw.githubusercontent.com/PureMath86/imgs/master/workflow.png\" alt=\"enter image description here\"></p>\n\n<p><strong>Edit:</strong> More examples --as you can see, sometimes it doesn't get the lungs quite right. </p>\n\n<p><img src=\"https://raw.githubusercontent.com/PureMath86/imgs/master/workflow2.png\" alt=\"enter image description here\"></p>",
  "messages": [
    {
      "id": "383547",
      "postDate": "09/09/2018 00:27:31",
      "content": "<p>There have been some good ideas expressed in other discussions on how to incorporate the DICOM metadata. Here are my ideas on the visually-based features:</p>\n\n<ul>\n<li>Localize and Crop Lungs,</li>\n<li>Analyze Gray-Scale Histograms,</li>\n<li>Clustering,</li>\n<li>and a plethora of other ideas on creating synthetic features.</li>\n</ul>\n\n<p>What are your thoughts?</p>\n\n<p><img src=\"https://raw.githubusercontent.com/PureMath86/imgs/master/workflow.png\" alt=\"enter image description here\"></p>\n\n<p><strong>Edit:</strong> More examples --as you can see, sometimes it doesn't get the lungs quite right. </p>\n\n<p><img src=\"https://raw.githubusercontent.com/PureMath86/imgs/master/workflow2.png\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "There have been some good ideas expressed in other discussions on how to incorporate the DICOM metadata. Here are my ideas on the visually-based features:\n\n * Localize and Crop Lungs,\n * Analyze Gray-Scale Histograms,\n * Clustering,\n * and a plethora of other ideas on creating synthetic features.\n\nWhat are your thoughts?\n\n![enter image description here][1]\n\n\n  [1]: https://raw.githubusercontent.com/PureMath86/imgs/master/workflow.png\n\n**Edit:** More examples --as you can see, sometimes it doesn't get the lungs quite right. \n\n![enter image description here][2]\n\n[2]: https://raw.githubusercontent.com/PureMath86/imgs/master/workflow2.png",
      "votes": null
    },
    {
      "id": "383562",
      "postDate": "09/09/2018 01:08:57",
      "content": "<p>Did you check that all gt boxes are inside your segmentation? </p>",
      "rawMarkdown": "Did you check that all gt boxes are inside your segmentation?",
      "votes": null
    },
    {
      "id": "383569",
      "postDate": "09/09/2018 01:40:04",
      "content": "<p>I have. In the above example they both are. However, in general, it is not always true. But I'm not too concerned with this at the moment. In my current strategy, I'm trying to find the central point of \"cloudiness\" (after pre-filtering images by classifying them as pneumonia, normal, and abnormal).</p>",
      "rawMarkdown": "I have. In the above example they both are. However, in general, it is not always true. But I'm not too concerned with this at the moment. In my current strategy, I'm trying to find the central point of \"cloudiness\" (after pre-filtering images by classifying them as pneumonia, normal, and abnormal).",
      "votes": null
    },
    {
      "id": "383572",
      "postDate": "09/09/2018 01:47:45",
      "content": "<p>Some of the xrays are really difficult. Anyways, my strategy would be to clip the gt to the segmented bboxes and teain an assistive model... We shall see how this goes... </p>",
      "rawMarkdown": "Some of the xrays are really difficult. Anyways, my strategy would be to clip the gt to the segmented bboxes and teain an assistive model... We shall see how this goes...",
      "votes": null
    },
    {
      "id": "383580",
      "postDate": "09/09/2018 02:27:08",
      "content": "<p>I like it. Good luck!</p>",
      "rawMarkdown": "I like it. Good luck!",
      "votes": null
    },
    {
      "id": "388005",
      "postDate": "09/16/2018 04:55:24",
      "content": "<p>I'm becoming increasingly convinced that dealing with the annotation misalignment will be the key to winning solution.</p>",
      "rawMarkdown": "I'm becoming increasingly convinced that dealing with the annotation misalignment will be the key to winning solution.",
      "votes": null
    },
    {
      "id": "388098",
      "postDate": "09/16/2018 10:18:40",
      "content": "<p>Hi Bryan, can you expand on this at all and maybe give an example of misalignment? Would be very grateful, thanks.</p>",
      "rawMarkdown": "Hi Bryan, can you expand on this at all and maybe give an example of misalignment? Would be very grateful, thanks.",
      "votes": null
    },
    {
      "id": "388829",
      "postDate": "09/17/2018 16:54:42",
      "content": "<p>Sure. I'll cook up some examples after work. Though the very first image above is a good example. The box \"Pneumonia Target 1\" is sub-optimal, for example. It covers almost as much \"lung area\" as \"outside-of-the-lung area\". And the cloudy, opaque pneumonia regions drift outside of the box. Hence, misalignment.</p>\n\n<p>You could train a model, which becomes really good at detecting and localizing pneumonia --but that may not win. We score points for drawing boxes close to what the annotator would draw. That's a different problem entirely.</p>",
      "rawMarkdown": "Sure. I'll cook up some examples after work. Though the very first image above is a good example. The box \"Pneumonia Target 1\" is sub-optimal, for example. It covers almost as much \"lung area\" as \"outside-of-the-lung area\". And the cloudy, opaque pneumonia regions drift outside of the box. Hence, misalignment.\n\nYou could train a model, which becomes really good at detecting and localizing pneumonia --but that may not win. We score points for drawing boxes close to what the annotator would draw. That's a different problem entirely.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 383562,
      "author_name": "moshel",
      "author_url": "",
      "post_date": "09/09/2018 01:08:57",
      "content": "<p>Did you check that all gt boxes are inside your segmentation? </p>",
      "votes": null,
      "replies": [
        {
          "id": 383569,
          "author_name": "puremath86",
          "author_url": "",
          "post_date": "09/09/2018 01:40:04",
          "content": "<p>I have. In the above example they both are. However, in general, it is not always true. But I'm not too concerned with this at the moment. In my current strategy, I'm trying to find the central point of \"cloudiness\" (after pre-filtering images by classifying them as pneumonia, normal, and abnormal).</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 383572,
          "author_name": "moshel",
          "author_url": "",
          "post_date": "09/09/2018 01:47:45",
          "content": "<p>Some of the xrays are really difficult. Anyways, my strategy would be to clip the gt to the segmented bboxes and teain an assistive model... We shall see how this goes... </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 383580,
          "author_name": "puremath86",
          "author_url": "",
          "post_date": "09/09/2018 02:27:08",
          "content": "<p>I like it. Good luck!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 388005,
      "author_name": "puremath86",
      "author_url": "",
      "post_date": "09/16/2018 04:55:24",
      "content": "<p>I'm becoming increasingly convinced that dealing with the annotation misalignment will be the key to winning solution.</p>",
      "votes": null,
      "replies": [
        {
          "id": 388098,
          "author_name": "taindow",
          "author_url": "",
          "post_date": "09/16/2018 10:18:40",
          "content": "<p>Hi Bryan, can you expand on this at all and maybe give an example of misalignment? Would be very grateful, thanks.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 388829,
          "author_name": "puremath86",
          "author_url": "",
          "post_date": "09/17/2018 16:54:42",
          "content": "<p>Sure. I'll cook up some examples after work. Though the very first image above is a good example. The box \"Pneumonia Target 1\" is sub-optimal, for example. It covers almost as much \"lung area\" as \"outside-of-the-lung area\". And the cloudy, opaque pneumonia regions drift outside of the box. Hence, misalignment.</p>\n\n<p>You could train a model, which becomes really good at detecting and localizing pneumonia --but that may not win. We score points for drawing boxes close to what the annotator would draw. That's a different problem entirely.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "383547": "There have been some good ideas expressed in other discussions on how to incorporate the DICOM metadata. Here are my ideas on the visually-based features:\n\n * Localize and Crop Lungs,\n * Analyze Gray-Scale Histograms,\n * Clustering,\n * and a plethora of other ideas on creating synthetic features.\n\nWhat are your thoughts?\n\n![enter image description here][1]\n\n\n  [1]: https://raw.githubusercontent.com/PureMath86/imgs/master/workflow.png\n\n**Edit:** More examples --as you can see, sometimes it doesn't get the lungs quite right. \n\n![enter image description here][2]\n\n[2]: https://raw.githubusercontent.com/PureMath86/imgs/master/workflow2.png",
    "383562": "Did you check that all gt boxes are inside your segmentation?",
    "383569": "I have. In the above example they both are. However, in general, it is not always true. But I'm not too concerned with this at the moment. In my current strategy, I'm trying to find the central point of \"cloudiness\" (after pre-filtering images by classifying them as pneumonia, normal, and abnormal).",
    "383572": "Some of the xrays are really difficult. Anyways, my strategy would be to clip the gt to the segmented bboxes and teain an assistive model... We shall see how this goes...",
    "383580": "I like it. Good luck!",
    "388005": "I'm becoming increasingly convinced that dealing with the annotation misalignment will be the key to winning solution.",
    "388098": "Hi Bryan, can you expand on this at all and maybe give an example of misalignment? Would be very grateful, thanks.",
    "388829": "Sure. I'll cook up some examples after work. Though the very first image above is a good example. The box \"Pneumonia Target 1\" is sub-optimal, for example. It covers almost as much \"lung area\" as \"outside-of-the-lung area\". And the cloudy, opaque pneumonia regions drift outside of the box. Hence, misalignment.\n\nYou could train a model, which becomes really good at detecting and localizing pneumonia --but that may not win. We score points for drawing boxes close to what the annotator would draw. That's a different problem entirely."
  },
  "source": "meta"
}