{
  "id": 171098,
  "title": "Histogram features",
  "url": "/competitions/osic-pulmonary-fibrosis-progression/discussion/171098",
  "author_name": "",
  "post_date": "2020-07-30T11:59:35.729006600Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I've trained quantile regression model with <a href=\"https://www.kaggle.com/jameschapman19/histogram-features\">histogram features</a> (big thanks for your work BTW) but my CV has not improved in comparison with tabular features only. I have not implemented the inference yet.\nDid anyone experience a boost ?</p>",
  "messages": [
    {
      "id": "951776",
      "postDate": "07/30/2020 11:59:35",
      "content": "<p>I've trained quantile regression model with <a href=\"https://www.kaggle.com/jameschapman19/histogram-features\">histogram features</a> (big thanks for your work BTW) but my CV has not improved in comparison with tabular features only. I have not implemented the inference yet.\nDid anyone experience a boost ?</p>",
      "rawMarkdown": "I've trained quantile regression model with [histogram features](https://www.kaggle.com/jameschapman19/histogram-features) (big thanks for your work BTW) but my CV has not improved in comparison with tabular features only. I have not implemented the inference yet.\nDid anyone experience a boost ?",
      "votes": null
    },
    {
      "id": "951994",
      "postDate": "07/30/2020 14:43:57",
      "content": "<p>I've also not noticed a reproducible improvement. I suspect I might need to do a bit more work e.g. one idea I have is to follow <a href=\"https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/165727\">https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/165727</a> and maybe do the histograms by slices/patches rather than the whole 3D segmented lung for each patient which is what i've done in the notebook you've linked</p>",
      "rawMarkdown": "I've also not noticed a reproducible improvement. I suspect I might need to do a bit more work e.g. one idea I have is to follow https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/165727 and maybe do the histograms by slices/patches rather than the whole 3D segmented lung for each patient which is what i've done in the notebook you've linked",
      "votes": null
    },
    {
      "id": "952015",
      "postDate": "07/30/2020 15:00:40",
      "content": "<p>I'm also suspecting that the number of patients is to small to learn something valuable when using histogram features. These extracted features could be too noisy.\nOne could also compute the chest area based on slices of the segmented lungs but I did not try yet. </p>",
      "rawMarkdown": "I'm also suspecting that the number of patients is to small to learn something valuable when using histogram features. These extracted features could be too noisy.\nOne could also compute the chest area based on slices of the segmented lungs but I did not try yet.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 951994,
      "author_name": "jameschapman19",
      "author_url": "",
      "post_date": "07/30/2020 14:43:57",
      "content": "<p>I've also not noticed a reproducible improvement. I suspect I might need to do a bit more work e.g. one idea I have is to follow <a href=\"https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/165727\">https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/165727</a> and maybe do the histograms by slices/patches rather than the whole 3D segmented lung for each patient which is what i've done in the notebook you've linked</p>",
      "votes": null,
      "replies": [
        {
          "id": 952015,
          "author_name": "alexj21",
          "author_url": "",
          "post_date": "07/30/2020 15:00:40",
          "content": "<p>I'm also suspecting that the number of patients is to small to learn something valuable when using histogram features. These extracted features could be too noisy.\nOne could also compute the chest area based on slices of the segmented lungs but I did not try yet. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "951776": "I've trained quantile regression model with [histogram features](https://www.kaggle.com/jameschapman19/histogram-features) (big thanks for your work BTW) but my CV has not improved in comparison with tabular features only. I have not implemented the inference yet.\nDid anyone experience a boost ?",
    "951994": "I've also not noticed a reproducible improvement. I suspect I might need to do a bit more work e.g. one idea I have is to follow https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/165727 and maybe do the histograms by slices/patches rather than the whole 3D segmented lung for each patient which is what i've done in the notebook you've linked",
    "952015": "I'm also suspecting that the number of patients is to small to learn something valuable when using histogram features. These extracted features could be too noisy.\nOne could also compute the chest area based on slices of the segmented lungs but I did not try yet."
  },
  "source": "meta"
}