{
  "id": 276628,
  "title": "2d cnn using single image is very overfitting",
  "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/276628",
  "author_name": "",
  "post_date": "2021-10-05T14:42:33.681435200Z",
  "votes": 10,
  "comment_count": 9,
  "views": 0,
  "content": "<p>I tried 2d efficientnet-b0, but it was very overfitting.</p>\n<p>e.x) after 10 epochs, 3 stratified group kfold<br>\nepoch:10/10 ite:490/490 [Train]loss:0.3410 score:0.88050 [Val]loss:1.5523 score:0.5211</p>\n<p>I stacked RNN, but result did not change.</p>\n<p>In training above Effnet, we use single image as input. I think single image always don't have tumor, but training auc is very high. This looks like leakage?</p>\n<p>Any public notebooks don't look having good convergence. </p>\n<p>This confuses me a a lot. </p>",
  "messages": [
    {
      "id": "1535182",
      "postDate": "10/05/2021 14:42:33",
      "content": "<p>I tried 2d efficientnet-b0, but it was very overfitting.</p>\n<p>e.x) after 10 epochs, 3 stratified group kfold<br>\nepoch:10/10 ite:490/490 [Train]loss:0.3410 score:0.88050 [Val]loss:1.5523 score:0.5211</p>\n<p>I stacked RNN, but result did not change.</p>\n<p>In training above Effnet, we use single image as input. I think single image always don't have tumor, but training auc is very high. This looks like leakage?</p>\n<p>Any public notebooks don't look having good convergence. </p>\n<p>This confuses me a a lot. </p>",
      "rawMarkdown": "I tried 2d efficientnet-b0, but it was very overfitting.\n\ne.x) after 10 epochs, 3 stratified group kfold\nepoch:10/10 ite:490/490 [Train]loss:0.3410 score:0.88050 [Val]loss:1.5523 score:0.5211\n\nI stacked RNN, but result did not change.\n\nIn training above Effnet, we use single image as input. I think single image always don't have tumor, but training auc is very high. This looks like leakage?\n\nAny public notebooks don't look having good convergence. \n\nThis confuses me a a lot.",
      "votes": null
    },
    {
      "id": "1535239",
      "postDate": "10/05/2021 15:43:32",
      "content": "<p>Indeed. <br>\n<img src=\"https://user-images.githubusercontent.com/17668390/136056528-5a8d1a12-1253-444f-81bb-fb93b4447f49.jpg\" alt=\"05xlvswuhfg41\"></p>",
      "rawMarkdown": "Indeed. \n![05xlvswuhfg41](https://user-images.githubusercontent.com/17668390/136056528-5a8d1a12-1253-444f-81bb-fb93b4447f49.jpg)",
      "votes": null
    },
    {
      "id": "1537566",
      "postDate": "10/07/2021 15:25:08",
      "content": "<p>lol , yeah this is just my cnn:)</p>",
      "rawMarkdown": "lol , yeah this is just my cnn:)",
      "votes": null
    },
    {
      "id": "1538164",
      "postDate": "10/08/2021 06:06:36",
      "content": "<p>What labels are you using for individual slices?<br>\nI don't think all slices of a positive scan will be positive.</p>",
      "rawMarkdown": "What labels are you using for individual slices?\nI don't think all slices of a positive scan will be positive.",
      "votes": null
    },
    {
      "id": "1538472",
      "postDate": "10/08/2021 13:11:14",
      "content": "<p>Why would you use single image for this problem?</p>",
      "rawMarkdown": "Why would you use single image for this problem?",
      "votes": null
    },
    {
      "id": "1538603",
      "postDate": "10/08/2021 15:19:18",
      "content": "<p>Above is just trial.<br>\nI am trying another way now.</p>",
      "rawMarkdown": "Above is just trial.\nI am trying another way now.",
      "votes": null
    },
    {
      "id": "1538608",
      "postDate": "10/08/2021 15:20:39",
      "content": "<p>Yeah I think so too.</p>\n<p>I said same thing above.</p>\n<blockquote>\n  <p>I think single image always don't have tumor</p>\n</blockquote>",
      "rawMarkdown": "Yeah I think so too.\n\nI said same thing above.\n> I think single image always don't have tumor",
      "votes": null
    },
    {
      "id": "1541003",
      "postDate": "10/11/2021 06:10:01",
      "content": "<p>are all dicom images in the competition dataset 3d?</p>",
      "rawMarkdown": "are all dicom images in the competition dataset 3d?",
      "votes": null
    },
    {
      "id": "1541025",
      "postDate": "10/11/2021 06:38:08",
      "content": "<p>Dicoms generally contain 2D image which are cross sectional slice of the 3D volume.<br>\nEach folder represent a sequence of the scan, there are multiple Dicoms inside each folder.</p>",
      "rawMarkdown": "Dicoms generally contain 2D image which are cross sectional slice of the 3D volume.\nEach folder represent a sequence of the scan, there are multiple Dicoms inside each folder.",
      "votes": null
    },
    {
      "id": "1541129",
      "postDate": "10/11/2021 07:47:10",
      "content": "<p>Is mr acquisition type in all images 3d then?</p>",
      "rawMarkdown": "Is mr acquisition type in all images 3d then?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1535239,
      "author_name": "ipythonx",
      "author_url": "",
      "post_date": "10/05/2021 15:43:32",
      "content": "<p>Indeed. <br>\n<img src=\"https://user-images.githubusercontent.com/17668390/136056528-5a8d1a12-1253-444f-81bb-fb93b4447f49.jpg\" alt=\"05xlvswuhfg41\"></p>",
      "votes": null,
      "replies": [
        {
          "id": 1537566,
          "author_name": "yoshito",
          "author_url": "",
          "post_date": "10/07/2021 15:25:08",
          "content": "<p>lol , yeah this is just my cnn:)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1538164,
      "author_name": "pranshu15",
      "author_url": "",
      "post_date": "10/08/2021 06:06:36",
      "content": "<p>What labels are you using for individual slices?<br>\nI don't think all slices of a positive scan will be positive.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1538608,
          "author_name": "yoshito",
          "author_url": "",
          "post_date": "10/08/2021 15:20:39",
          "content": "<p>Yeah I think so too.</p>\n<p>I said same thing above.</p>\n<blockquote>\n  <p>I think single image always don't have tumor</p>\n</blockquote>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1538472,
      "author_name": "gunesevitan",
      "author_url": "",
      "post_date": "10/08/2021 13:11:14",
      "content": "<p>Why would you use single image for this problem?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1538603,
          "author_name": "yoshito",
          "author_url": "",
          "post_date": "10/08/2021 15:19:18",
          "content": "<p>Above is just trial.<br>\nI am trying another way now.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1541003,
      "author_name": "aniarya",
      "author_url": "",
      "post_date": "10/11/2021 06:10:01",
      "content": "<p>are all dicom images in the competition dataset 3d?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1541025,
          "author_name": "pranshu15",
          "author_url": "",
          "post_date": "10/11/2021 06:38:08",
          "content": "<p>Dicoms generally contain 2D image which are cross sectional slice of the 3D volume.<br>\nEach folder represent a sequence of the scan, there are multiple Dicoms inside each folder.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1541129,
          "author_name": "aniarya",
          "author_url": "",
          "post_date": "10/11/2021 07:47:10",
          "content": "<p>Is mr acquisition type in all images 3d then?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1535182": "I tried 2d efficientnet-b0, but it was very overfitting.\n\ne.x) after 10 epochs, 3 stratified group kfold\nepoch:10/10 ite:490/490 [Train]loss:0.3410 score:0.88050 [Val]loss:1.5523 score:0.5211\n\nI stacked RNN, but result did not change.\n\nIn training above Effnet, we use single image as input. I think single image always don't have tumor, but training auc is very high. This looks like leakage?\n\nAny public notebooks don't look having good convergence. \n\nThis confuses me a a lot.",
    "1535239": "Indeed. \n![05xlvswuhfg41](https://user-images.githubusercontent.com/17668390/136056528-5a8d1a12-1253-444f-81bb-fb93b4447f49.jpg)",
    "1537566": "lol , yeah this is just my cnn:)",
    "1538164": "What labels are you using for individual slices?\nI don't think all slices of a positive scan will be positive.",
    "1538472": "Why would you use single image for this problem?",
    "1538603": "Above is just trial.\nI am trying another way now.",
    "1538608": "Yeah I think so too.\n\nI said same thing above.\n> I think single image always don't have tumor",
    "1541003": "are all dicom images in the competition dataset 3d?",
    "1541025": "Dicoms generally contain 2D image which are cross sectional slice of the 3D volume.\nEach folder represent a sequence of the scan, there are multiple Dicoms inside each folder.",
    "1541129": "Is mr acquisition type in all images 3d then?"
  },
  "source": "meta"
}