{
  "id": 280043,
  "title": "Were our efforts really useless?",
  "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/280043",
  "author_name": "",
  "post_date": "2021-10-20T06:39:00.836605Z",
  "votes": 10,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I am seeing many posts, discussions saying this competition has failed and the models didn't learn anything. While I completely agree with the fact that randomness helped some teams and destroyed some but saying the competition is a complete failure is not correct at all. Before this competition, I didn't know anything about dealing with 3D data or MRI scans, due to this competition I think if there is a similar competition I can perform better from all the wisdom/knowledge I learned, shared in the posts and notebooks. I think there are many people like me here.</p>\n<p>And saying this is not clinical usable, maybe the host's plan was never to use it clinically? maybe their aim was to see if ML can actually find new relations that we never observed previously.</p>\n<p>So let us not come to conclusions ourselves. This competition is interesting because we actually couldn't make our models learn good features for detecting MGMT, I don't think even with more data we can solve this problem.  Maybe there is some other way, maybe this competition would make the researchers working on this problem make them choose a different path and solve it in some other way.<br>\nBecause that's how new discoveries are made, and that's how we learn things. And I think like every time the kaggle community hasn't disappointed and maybe our efforts will bring some change.<br>\nThanks</p>",
  "messages": [
    {
      "id": "1550913",
      "postDate": "10/20/2021 06:39:00",
      "content": "<p>I am seeing many posts, discussions saying this competition has failed and the models didn't learn anything. While I completely agree with the fact that randomness helped some teams and destroyed some but saying the competition is a complete failure is not correct at all. Before this competition, I didn't know anything about dealing with 3D data or MRI scans, due to this competition I think if there is a similar competition I can perform better from all the wisdom/knowledge I learned, shared in the posts and notebooks. I think there are many people like me here.</p>\n<p>And saying this is not clinical usable, maybe the host's plan was never to use it clinically? maybe their aim was to see if ML can actually find new relations that we never observed previously.</p>\n<p>So let us not come to conclusions ourselves. This competition is interesting because we actually couldn't make our models learn good features for detecting MGMT, I don't think even with more data we can solve this problem.  Maybe there is some other way, maybe this competition would make the researchers working on this problem make them choose a different path and solve it in some other way.<br>\nBecause that's how new discoveries are made, and that's how we learn things. And I think like every time the kaggle community hasn't disappointed and maybe our efforts will bring some change.<br>\nThanks</p>",
      "rawMarkdown": "I am seeing many posts, discussions saying this competition has failed and the models didn't learn anything. While I completely agree with the fact that randomness helped some teams and destroyed some but saying the competition is a complete failure is not correct at all. Before this competition, I didn't know anything about dealing with 3D data or MRI scans, due to this competition I think if there is a similar competition I can perform better from all the wisdom/knowledge I learned, shared in the posts and notebooks. I think there are many people like me here.\n\nAnd saying this is not clinical usable, maybe the host's plan was never to use it clinically? maybe their aim was to see if ML can actually find new relations that we never observed previously.\n\nSo let us not come to conclusions ourselves. This competition is interesting because we actually couldn't make our models learn good features for detecting MGMT, I don't think even with more data we can solve this problem.  Maybe there is some other way, maybe this competition would make the researchers working on this problem make them choose a different path and solve it in some other way.\nBecause that's how new discoveries are made, and that's how we learn things. And I think like every time the kaggle community hasn't disappointed and maybe our efforts will bring some change.\nThanks",
      "votes": null
    },
    {
      "id": "1550945",
      "postDate": "10/20/2021 07:05:15",
      "content": "<p>The only doubt I have are regarding some papers where they said they achieved &gt;0.9 AUC, don't think it is possible.</p>",
      "rawMarkdown": "The only doubt I have are regarding some papers where they said they achieved >0.9 AUC, don't think it is possible.",
      "votes": null
    },
    {
      "id": "1550960",
      "postDate": "10/20/2021 07:21:34",
      "content": "<p>I think, they weren't. People from gold have started sharing their solutions and I am pleased to see that so far none of them have used random predictions as their submits. I doubt that the organizers didn't even try to solve this problem on their own. They know what scores to expect. </p>\n<p>I see two ways here. Firstly, classic approach, - use an ensemble from top performing solutions. The 12th place solution shows that it improves the score very well, at least on validation. But most likely, here will be used <code>Precision approach</code>, the purpose of which is rather to put a tumor where it is not, than not to notice it where it is.</p>",
      "rawMarkdown": "I think, they weren't. People from gold have started sharing their solutions and I am pleased to see that so far none of them have used random predictions as their submits. I doubt that the organizers didn't even try to solve this problem on their own. They know what scores to expect. \n\nI see two ways here. Firstly, classic approach, - use an ensemble from top performing solutions. The 12th place solution shows that it improves the score very well, at least on validation. But most likely, here will be used `Precision approach`, the purpose of which is rather to put a tumor where it is not, than not to notice it where it is.",
      "votes": null
    },
    {
      "id": "1550996",
      "postDate": "10/20/2021 07:51:46",
      "content": "<p>rightly said</p>",
      "rawMarkdown": "rightly said",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1550945,
      "author_name": "mrinath",
      "author_url": "",
      "post_date": "10/20/2021 07:05:15",
      "content": "<p>The only doubt I have are regarding some papers where they said they achieved &gt;0.9 AUC, don't think it is possible.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1550960,
      "author_name": "vadimtimakin",
      "author_url": "",
      "post_date": "10/20/2021 07:21:34",
      "content": "<p>I think, they weren't. People from gold have started sharing their solutions and I am pleased to see that so far none of them have used random predictions as their submits. I doubt that the organizers didn't even try to solve this problem on their own. They know what scores to expect. </p>\n<p>I see two ways here. Firstly, classic approach, - use an ensemble from top performing solutions. The 12th place solution shows that it improves the score very well, at least on validation. But most likely, here will be used <code>Precision approach</code>, the purpose of which is rather to put a tumor where it is not, than not to notice it where it is.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1550996,
          "author_name": "mrinath",
          "author_url": "",
          "post_date": "10/20/2021 07:51:46",
          "content": "<p>rightly said</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1550913": "I am seeing many posts, discussions saying this competition has failed and the models didn't learn anything. While I completely agree with the fact that randomness helped some teams and destroyed some but saying the competition is a complete failure is not correct at all. Before this competition, I didn't know anything about dealing with 3D data or MRI scans, due to this competition I think if there is a similar competition I can perform better from all the wisdom/knowledge I learned, shared in the posts and notebooks. I think there are many people like me here.\n\nAnd saying this is not clinical usable, maybe the host's plan was never to use it clinically? maybe their aim was to see if ML can actually find new relations that we never observed previously.\n\nSo let us not come to conclusions ourselves. This competition is interesting because we actually couldn't make our models learn good features for detecting MGMT, I don't think even with more data we can solve this problem.  Maybe there is some other way, maybe this competition would make the researchers working on this problem make them choose a different path and solve it in some other way.\nBecause that's how new discoveries are made, and that's how we learn things. And I think like every time the kaggle community hasn't disappointed and maybe our efforts will bring some change.\nThanks",
    "1550945": "The only doubt I have are regarding some papers where they said they achieved >0.9 AUC, don't think it is possible.",
    "1550960": "I think, they weren't. People from gold have started sharing their solutions and I am pleased to see that so far none of them have used random predictions as their submits. I doubt that the organizers didn't even try to solve this problem on their own. They know what scores to expect. \n\nI see two ways here. Firstly, classic approach, - use an ensemble from top performing solutions. The 12th place solution shows that it improves the score very well, at least on validation. But most likely, here will be used `Precision approach`, the purpose of which is rather to put a tumor where it is not, than not to notice it where it is.",
    "1550996": "rightly said"
  },
  "source": "meta"
}