{
  "id": 66130,
  "title": "Will Data Augmentation improve robustness of model?",
  "url": "/competitions/rsna-pneumonia-detection-challenge/discussion/66130",
  "author_name": "",
  "post_date": "2018-09-18T13:09:35.280791Z",
  "votes": 2,
  "comment_count": 6,
  "views": 0,
  "content": "<p>If data augmentation works, which way is best for this task?</p>",
  "messages": [
    {
      "id": "389275",
      "postDate": "09/18/2018 13:09:35",
      "content": "<p>If data augmentation works, which way is best for this task?</p>",
      "rawMarkdown": "If data augmentation works, which way is best for this task?",
      "votes": null
    },
    {
      "id": "389277",
      "postDate": "09/18/2018 13:13:00",
      "content": "<p>Yes in my experience it should help improve the result. The imgaug library is one of the best ones for image augmentation. </p>",
      "rawMarkdown": "Yes in my experience it should help improve the result. The imgaug library is one of the best ones for image augmentation.",
      "votes": null
    },
    {
      "id": "390234",
      "postDate": "09/19/2018 22:29:16",
      "content": "<p>Make sure you are careful your augmentation does not modify the labeling. For example, a brightness augmentation on this data could change what the image represents. You can look at previous work done on x-ray images to see what augmentations are viable.</p>",
      "rawMarkdown": "Make sure you are careful your augmentation does not modify the labeling. For example, a brightness augmentation on this data could change what the image represents. You can look at previous work done on x-ray images to see what augmentations are viable.",
      "votes": null
    },
    {
      "id": "394959",
      "postDate": "09/27/2018 18:12:53",
      "content": "<p>Do you think horizontal and vertical flipping will help ?? Because In reality its very unlikely that the chest X-ray will be vertical and horizontally flipped. </p>",
      "rawMarkdown": "Do you think horizontal and vertical flipping will help ?? Because In reality its very unlikely that the chest X-ray will be vertical and horizontally flipped.",
      "votes": null
    },
    {
      "id": "395543",
      "postDate": "09/28/2018 19:26:33",
      "content": "<p>So having looked at this quite a bit (and continuing to look at it quite a bit), the quick answer is that horizontal flipping seems to perform better as an augmentation operation than vertical flipping (abstract presented at 2018 SIIM, paper in works).  The rationale I give it is that the chest X-ray is a fairly structured examination, taken the same way every time, and the body is outwardly symmetric across the midline (bilateral symmetry).  The lungs are reasonably symmetric, and the outer soft tissues and bones of the chest very symmetric.  The only area of relative assymetry is the heart/mediastinum, which is not the direct focus of this exam.</p>\n\n<p>Compare that with a vertical flip.  Now, you are putting the bottom of the lung bases up at the top of the image, with the heart at the top of the image.  The neck is down at the bottom.  This doesn't make sense, to a radiologist or a classifier.  So, your accuracy goes down with that augmentation.</p>\n\n<p>Most of the time.  The thing is, that there is an element of regularization here.  By providing the horizontal and vertical flips, you help the network to \"forget\" some of the overtrained detail it has learned making it more robust.  Depending on how much data you have, and how good the data is, these augments might help or hurt.  Bottom line is you usually have to try it out.</p>",
      "rawMarkdown": "So having looked at this quite a bit (and continuing to look at it quite a bit), the quick answer is that horizontal flipping seems to perform better as an augmentation operation than vertical flipping (abstract presented at 2018 SIIM, paper in works).  The rationale I give it is that the chest X-ray is a fairly structured examination, taken the same way every time, and the body is outwardly symmetric across the midline (bilateral symmetry).  The lungs are reasonably symmetric, and the outer soft tissues and bones of the chest very symmetric.  The only area of relative assymetry is the heart/mediastinum, which is not the direct focus of this exam.\n\nCompare that with a vertical flip.  Now, you are putting the bottom of the lung bases up at the top of the image, with the heart at the top of the image.  The neck is down at the bottom.  This doesn't make sense, to a radiologist or a classifier.  So, your accuracy goes down with that augmentation.\n\nMost of the time.  The thing is, that there is an element of regularization here.  By providing the horizontal and vertical flips, you help the network to \"forget\" some of the overtrained detail it has learned making it more robust.  Depending on how much data you have, and how good the data is, these augments might help or hurt.  Bottom line is you usually have to try it out.",
      "votes": null
    },
    {
      "id": "395545",
      "postDate": "09/28/2018 19:33:37",
      "content": "<p>@drsxr Thanks a lot for your reply.  I totally agree with you</p>",
      "rawMarkdown": "drsxr Thanks a lot for your reply.  I totally agree with you",
      "votes": null
    },
    {
      "id": "396470",
      "postDate": "09/30/2018 20:02:32",
      "content": "<p>Be careful when you make an augmentation on your image, sometime you also need to replace your bounding boxes</p>",
      "rawMarkdown": "Be careful when you make an augmentation on your image, sometime you also need to replace your bounding boxes",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 389277,
      "author_name": "mpsampat",
      "author_url": "",
      "post_date": "09/18/2018 13:13:00",
      "content": "<p>Yes in my experience it should help improve the result. The imgaug library is one of the best ones for image augmentation. </p>",
      "votes": null,
      "replies": [
        {
          "id": 394959,
          "author_name": "pradeeprathore04",
          "author_url": "",
          "post_date": "09/27/2018 18:12:53",
          "content": "<p>Do you think horizontal and vertical flipping will help ?? Because In reality its very unlikely that the chest X-ray will be vertical and horizontally flipped. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 395543,
          "author_name": "drsxr1",
          "author_url": "",
          "post_date": "09/28/2018 19:26:33",
          "content": "<p>So having looked at this quite a bit (and continuing to look at it quite a bit), the quick answer is that horizontal flipping seems to perform better as an augmentation operation than vertical flipping (abstract presented at 2018 SIIM, paper in works).  The rationale I give it is that the chest X-ray is a fairly structured examination, taken the same way every time, and the body is outwardly symmetric across the midline (bilateral symmetry).  The lungs are reasonably symmetric, and the outer soft tissues and bones of the chest very symmetric.  The only area of relative assymetry is the heart/mediastinum, which is not the direct focus of this exam.</p>\n\n<p>Compare that with a vertical flip.  Now, you are putting the bottom of the lung bases up at the top of the image, with the heart at the top of the image.  The neck is down at the bottom.  This doesn't make sense, to a radiologist or a classifier.  So, your accuracy goes down with that augmentation.</p>\n\n<p>Most of the time.  The thing is, that there is an element of regularization here.  By providing the horizontal and vertical flips, you help the network to \"forget\" some of the overtrained detail it has learned making it more robust.  Depending on how much data you have, and how good the data is, these augments might help or hurt.  Bottom line is you usually have to try it out.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 395545,
          "author_name": "pradeeprathore04",
          "author_url": "",
          "post_date": "09/28/2018 19:33:37",
          "content": "<p>@drsxr Thanks a lot for your reply.  I totally agree with you</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 390234,
      "author_name": "arpandhatt",
      "author_url": "",
      "post_date": "09/19/2018 22:29:16",
      "content": "<p>Make sure you are careful your augmentation does not modify the labeling. For example, a brightness augmentation on this data could change what the image represents. You can look at previous work done on x-ray images to see what augmentations are viable.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 396470,
      "author_name": "perlimpinpin",
      "author_url": "",
      "post_date": "09/30/2018 20:02:32",
      "content": "<p>Be careful when you make an augmentation on your image, sometime you also need to replace your bounding boxes</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "389275": "If data augmentation works, which way is best for this task?",
    "389277": "Yes in my experience it should help improve the result. The imgaug library is one of the best ones for image augmentation.",
    "390234": "Make sure you are careful your augmentation does not modify the labeling. For example, a brightness augmentation on this data could change what the image represents. You can look at previous work done on x-ray images to see what augmentations are viable.",
    "394959": "Do you think horizontal and vertical flipping will help ?? Because In reality its very unlikely that the chest X-ray will be vertical and horizontally flipped.",
    "395543": "So having looked at this quite a bit (and continuing to look at it quite a bit), the quick answer is that horizontal flipping seems to perform better as an augmentation operation than vertical flipping (abstract presented at 2018 SIIM, paper in works).  The rationale I give it is that the chest X-ray is a fairly structured examination, taken the same way every time, and the body is outwardly symmetric across the midline (bilateral symmetry).  The lungs are reasonably symmetric, and the outer soft tissues and bones of the chest very symmetric.  The only area of relative assymetry is the heart/mediastinum, which is not the direct focus of this exam.\n\nCompare that with a vertical flip.  Now, you are putting the bottom of the lung bases up at the top of the image, with the heart at the top of the image.  The neck is down at the bottom.  This doesn't make sense, to a radiologist or a classifier.  So, your accuracy goes down with that augmentation.\n\nMost of the time.  The thing is, that there is an element of regularization here.  By providing the horizontal and vertical flips, you help the network to \"forget\" some of the overtrained detail it has learned making it more robust.  Depending on how much data you have, and how good the data is, these augments might help or hurt.  Bottom line is you usually have to try it out.",
    "395545": "drsxr Thanks a lot for your reply.  I totally agree with you",
    "396470": "Be careful when you make an augmentation on your image, sometime you also need to replace your bounding boxes"
  },
  "source": "meta"
}