{
  "id": 281911,
  "title": "5th place solution",
  "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/writeups/random-5th-place-solution",
  "author_name": "",
  "post_date": "2021-10-25T22:20:24.219720600Z",
  "votes": 9,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I would like to thank Kaggle and the competition hosts for this learning opportunity. </p>\n<p>Link to the <a href=\"https://github.com/bhaveshtangirala786/RSNA\" target=\"_blank\">Github</a> code.</p>\n<h1>Dataset</h1>\n<p>This kaggle <a href=\"https://www.kaggle.com/jonathanbesomi/rsna-miccai-png\" target=\"_blank\">dataset</a> was used for training the models. I thank <a href=\"https://www.kaggle.com/jonathanbesomi\" target=\"_blank\">@jonathanbesomi</a> for providing us this compact dataset</p>\n<h1>Approach</h1>\n<ul>\n<li>I sampled 10 images for each type of mpMRI scan (namely FLAIR, T1w, T1Gd, T2) and calculated the mean for each type to get 4 2D images.</li>\n<li>I then concatenated these 4 images to get a 4 channel (4 x h x w) image which I passed through a 1x1 convolution bottleneck to get a 3 channel (3 x h x w) feature map.</li>\n<li>This feature map is passed through a CNN (Efficientnet) to predict the MGMT value</li>\n<li>I used Taylor Cross Entropy loss for training as I thought this dataset might be a bit noisy due to less number of samples.</li>\n</ul>\n<h1>Notebook</h1>\n<p>Training notebook : <a href=\"https://www.kaggle.com/abhimanyukarshni/rsna-training/notebook\" target=\"_blank\">https://www.kaggle.com/abhimanyukarshni/rsna-training/notebook</a><br>\nInference notebook : <a href=\"https://www.kaggle.com/abhimanyukarshni/rsna-inference/notebook\" target=\"_blank\">https://www.kaggle.com/abhimanyukarshni/rsna-inference/notebook</a></p>",
  "messages": [
    {
      "id": "1557869",
      "postDate": "10/25/2021 22:20:24",
      "content": "<p>I would like to thank Kaggle and the competition hosts for this learning opportunity. </p>\n<p>Link to the <a href=\"https://github.com/bhaveshtangirala786/RSNA\" target=\"_blank\">Github</a> code.</p>\n<h1>Dataset</h1>\n<p>This kaggle <a href=\"https://www.kaggle.com/jonathanbesomi/rsna-miccai-png\" target=\"_blank\">dataset</a> was used for training the models. I thank <a href=\"https://www.kaggle.com/jonathanbesomi\" target=\"_blank\">@jonathanbesomi</a> for providing us this compact dataset</p>\n<h1>Approach</h1>\n<ul>\n<li>I sampled 10 images for each type of mpMRI scan (namely FLAIR, T1w, T1Gd, T2) and calculated the mean for each type to get 4 2D images.</li>\n<li>I then concatenated these 4 images to get a 4 channel (4 x h x w) image which I passed through a 1x1 convolution bottleneck to get a 3 channel (3 x h x w) feature map.</li>\n<li>This feature map is passed through a CNN (Efficientnet) to predict the MGMT value</li>\n<li>I used Taylor Cross Entropy loss for training as I thought this dataset might be a bit noisy due to less number of samples.</li>\n</ul>\n<h1>Notebook</h1>\n<p>Training notebook : <a href=\"https://www.kaggle.com/abhimanyukarshni/rsna-training/notebook\" target=\"_blank\">https://www.kaggle.com/abhimanyukarshni/rsna-training/notebook</a><br>\nInference notebook : <a href=\"https://www.kaggle.com/abhimanyukarshni/rsna-inference/notebook\" target=\"_blank\">https://www.kaggle.com/abhimanyukarshni/rsna-inference/notebook</a></p>",
      "rawMarkdown": "I would like to thank Kaggle and the competition hosts for this learning opportunity. \n\nLink to the [Github](https://github.com/bhaveshtangirala786/RSNA) code.\n\n# Dataset\nThis kaggle [dataset](https://www.kaggle.com/jonathanbesomi/rsna-miccai-png) was used for training the models. I thank @jonathanbesomi for providing us this compact dataset\n\n# Approach\n* I sampled 10 images for each type of mpMRI scan (namely FLAIR, T1w, T1Gd, T2) and calculated the mean for each type to get 4 2D images.\n* I then concatenated these 4 images to get a 4 channel (4 x h x w) image which I passed through a 1x1 convolution bottleneck to get a 3 channel (3 x h x w) feature map.\n* This feature map is passed through a CNN (Efficientnet) to predict the MGMT value\n* I used Taylor Cross Entropy loss for training as I thought this dataset might be a bit noisy due to less number of samples.\n\n# Notebook\nTraining notebook : https://www.kaggle.com/abhimanyukarshni/rsna-training/notebook\nInference notebook : https://www.kaggle.com/abhimanyukarshni/rsna-inference/notebook",
      "votes": null
    },
    {
      "id": "1558914",
      "postDate": "10/26/2021 14:53:30",
      "content": "<p>Congratulations!</p>",
      "rawMarkdown": "Congratulations!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1558914,
      "author_name": "atsunorifujita",
      "author_url": "",
      "post_date": "10/26/2021 14:53:30",
      "content": "<p>Congratulations!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1557869": "I would like to thank Kaggle and the competition hosts for this learning opportunity. \n\nLink to the [Github](https://github.com/bhaveshtangirala786/RSNA) code.\n\n# Dataset\nThis kaggle [dataset](https://www.kaggle.com/jonathanbesomi/rsna-miccai-png) was used for training the models. I thank @jonathanbesomi for providing us this compact dataset\n\n# Approach\n* I sampled 10 images for each type of mpMRI scan (namely FLAIR, T1w, T1Gd, T2) and calculated the mean for each type to get 4 2D images.\n* I then concatenated these 4 images to get a 4 channel (4 x h x w) image which I passed through a 1x1 convolution bottleneck to get a 3 channel (3 x h x w) feature map.\n* This feature map is passed through a CNN (Efficientnet) to predict the MGMT value\n* I used Taylor Cross Entropy loss for training as I thought this dataset might be a bit noisy due to less number of samples.\n\n# Notebook\nTraining notebook : https://www.kaggle.com/abhimanyukarshni/rsna-training/notebook\nInference notebook : https://www.kaggle.com/abhimanyukarshni/rsna-inference/notebook",
    "1558914": "Congratulations!"
  },
  "source": "meta"
}