{
  "id": 249778,
  "title": "To RGB or not to RGB, that is the question ...",
  "url": "/competitions/siim-covid19-detection/discussion/249778",
  "author_name": "",
  "post_date": "2021-06-29T19:21:23.059847400Z",
  "votes": 4,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hello,</p>\n<p>This is the first time I participate in a competition with medical images. I'm not certain if I should convert images to RGB (3 channels) or keep them with gray scale (1 channel).</p>\n<p>On one hand, converting images to RGB let me use transfer learning hands on with pre-trained models. On the other hand, I'm not sure that this solution will be efficient both in terms of accuracy and performance.</p>\n<p>What are your thoughts on this question?</p>",
  "messages": [
    {
      "id": "1369979",
      "postDate": "06/29/2021 19:21:23",
      "content": "<p>Hello,</p>\n<p>This is the first time I participate in a competition with medical images. I'm not certain if I should convert images to RGB (3 channels) or keep them with gray scale (1 channel).</p>\n<p>On one hand, converting images to RGB let me use transfer learning hands on with pre-trained models. On the other hand, I'm not sure that this solution will be efficient both in terms of accuracy and performance.</p>\n<p>What are your thoughts on this question?</p>",
      "rawMarkdown": "Hello,\n\nThis is the first time I participate in a competition with medical images. I'm not certain if I should convert images to RGB (3 channels) or keep them with gray scale (1 channel).\n\nOn one hand, converting images to RGB let me use transfer learning hands on with pre-trained models. On the other hand, I'm not sure that this solution will be efficient both in terms of accuracy and performance.\n\nWhat are your thoughts on this question?",
      "votes": null
    },
    {
      "id": "1381751",
      "postDate": "07/09/2021 07:45:28",
      "content": "<p>Depends on your approach, but using pre-trained model could be a good help. With this approach, you don't have to train your model from the beginning. If you want to use them, the basic approach will be to duplicate the grayscale channel.</p>\n<p>An another approach that I'm trying is to use is image enhancement. The idea is to apply some modification as CLAHE, histogram equalization… in order to have a better visualization of our image. Then, with the created image, you can decide to use them in a specific channel. For example, you can decide that the channel 1 will be the original grayscale image and the channel 2 and 3 a specific transformation. If you want to visualize that, I create a notebook where I show the transformed image: <a href=\"https://www.kaggle.com/rerere/covid-19-image-enhancement-in-progress#Combine-the-different-approaches\" target=\"_blank\">https://www.kaggle.com/rerere/covid-19-image-enhancement-in-progress#Combine-the-different-approaches</a><br>\nBut, for the moment, I didn't made result on that, so I don't know if this approach could be interesting.</p>",
      "rawMarkdown": "Depends on your approach, but using pre-trained model could be a good help. With this approach, you don't have to train your model from the beginning. If you want to use them, the basic approach will be to duplicate the grayscale channel.\n\nAn another approach that I'm trying is to use is image enhancement. The idea is to apply some modification as CLAHE, histogram equalization... in order to have a better visualization of our image. Then, with the created image, you can decide to use them in a specific channel. For example, you can decide that the channel 1 will be the original grayscale image and the channel 2 and 3 a specific transformation. If you want to visualize that, I create a notebook where I show the transformed image: https://www.kaggle.com/rerere/covid-19-image-enhancement-in-progress#Combine-the-different-approaches\nBut, for the moment, I didn't made result on that, so I don't know if this approach could be interesting.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1381751,
      "author_name": "rerere",
      "author_url": "",
      "post_date": "07/09/2021 07:45:28",
      "content": "<p>Depends on your approach, but using pre-trained model could be a good help. With this approach, you don't have to train your model from the beginning. If you want to use them, the basic approach will be to duplicate the grayscale channel.</p>\n<p>An another approach that I'm trying is to use is image enhancement. The idea is to apply some modification as CLAHE, histogram equalization… in order to have a better visualization of our image. Then, with the created image, you can decide to use them in a specific channel. For example, you can decide that the channel 1 will be the original grayscale image and the channel 2 and 3 a specific transformation. If you want to visualize that, I create a notebook where I show the transformed image: <a href=\"https://www.kaggle.com/rerere/covid-19-image-enhancement-in-progress#Combine-the-different-approaches\" target=\"_blank\">https://www.kaggle.com/rerere/covid-19-image-enhancement-in-progress#Combine-the-different-approaches</a><br>\nBut, for the moment, I didn't made result on that, so I don't know if this approach could be interesting.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1369979": "Hello,\n\nThis is the first time I participate in a competition with medical images. I'm not certain if I should convert images to RGB (3 channels) or keep them with gray scale (1 channel).\n\nOn one hand, converting images to RGB let me use transfer learning hands on with pre-trained models. On the other hand, I'm not sure that this solution will be efficient both in terms of accuracy and performance.\n\nWhat are your thoughts on this question?",
    "1381751": "Depends on your approach, but using pre-trained model could be a good help. With this approach, you don't have to train your model from the beginning. If you want to use them, the basic approach will be to duplicate the grayscale channel.\n\nAn another approach that I'm trying is to use is image enhancement. The idea is to apply some modification as CLAHE, histogram equalization... in order to have a better visualization of our image. Then, with the created image, you can decide to use them in a specific channel. For example, you can decide that the channel 1 will be the original grayscale image and the channel 2 and 3 a specific transformation. If you want to visualize that, I create a notebook where I show the transformed image: https://www.kaggle.com/rerere/covid-19-image-enhancement-in-progress#Combine-the-different-approaches\nBut, for the moment, I didn't made result on that, so I don't know if this approach could be interesting."
  },
  "source": "meta"
}