\n\nIt’s a double loss for the host(they didn’t get what they want) and for those kagglers who made tons of work but didn’t survive in the big shakeup.
\n
I don't know about this. The hosts certainly got something. They got the best (advertised) minds in the DS field to direct their attention at this problem for three months. It's just that the determination is clear that at least for now, the problem of determining the MGMT genetic marker from the modalities presented at the resolutions provided isn't tractable. Science ends like that sometimes. In the future, this problem might be possible with higher resolution scans, or with additional clinical biomarkers added into the mix. As for those that got shaken down, it would be nice if all the submissions had their private scores revealed because that'd help people learn a little about what worked and what didn't. But I understand withholding that information. The labeled test set probably was very expensive to produce and they want to use it again for other purposes, which wouldn't be possible if the labels became known widespread.
", "rawMarkdown": "> It’s a double loss for the host(they didn’t get what they want) and for those kagglers who made tons of work but didn’t survive in the big shakeup.\n\nI don't know about this. The hosts certainly got something. They got the best (advertised) minds in the DS field to direct their attention at this problem for three months. It's just that the determination is clear that at least for now, the problem of determining the MGMT genetic marker from the modalities presented at the resolutions provided isn't tractable. Science ends like that sometimes. In the future, this problem might be possible with higher resolution scans, or with additional clinical biomarkers added into the mix. As for those that got shaken down, it would be nice if all the submissions had their private scores revealed because that'd help people learn a little about what worked and what didn't. But I understand withholding that information. The labeled test set probably was very expensive to produce and they want to use it again for other purposes, which wouldn't be possible if the labels became known widespread.", "votes": null }, { "id": "1549453", "postDate": "10/19/2021 00:19:36", "content": "\n\nScience ends like that sometimes
\n
exactly! and is part of the scientific process
", "rawMarkdown": "> Science ends like that sometimes\n\nexactly! and is part of the scientific process", "votes": null }, { "id": "1549466", "postDate": "10/19/2021 00:54:16", "content": "You are right that 60% is clinically useless. I wonder how some papers have achieved 90% AUC with just T2 weighted images… Some papers should be re-examined.
", "rawMarkdown": "You are right that 60% is clinically useless. I wonder how some papers have achieved 90% AUC with just T2 weighted images... Some papers should be re-examined.", "votes": null }, { "id": "1549659", "postDate": "10/19/2021 05:56:25", "content": "I will say this is definitely waste of my time on this comp.
", "rawMarkdown": "I will say this is definitely waste of my time on this comp.", "votes": null }, { "id": "1549663", "postDate": "10/19/2021 05:58:18", "content": "I wonder how our score will change if we have 100x training data.
", "rawMarkdown": "I wonder how our score will change if we have 100x training data.", "votes": null }, { "id": "1549690", "postDate": "10/19/2021 06:18:09", "content": "I don’t think 100x training data will significantly improve the score. This kind of task is just similar to those that ask you to predict their genetic mutation based on their face images. The two things are hardly related, and you wouldn’t expect the model to predict well.
", "rawMarkdown": "I don’t think 100x training data will significantly improve the score. This kind of task is just similar to those that ask you to predict their genetic mutation based on their face images. The two things are hardly related, and you wouldn’t expect the model to predict well.", "votes": null }, { "id": "1549691", "postDate": "10/19/2021 06:19:01", "content": "Totally agree. Be careful when you choose to join a comp.
", "rawMarkdown": "Totally agree. Be careful when you choose to join a comp.", "votes": null }, { "id": "1549873", "postDate": "10/19/2021 08:54:26", "content": "@authman you're absolutely right on that!
", "rawMarkdown": "authman you're absolutely right on that!", "votes": null } ], "comments": [ { "id": 1549446, "author_name": "authman", "author_url": "", "post_date": "10/19/2021 00:08:03", "content": "\n\nIt’s a double loss for the host(they didn’t get what they want) and for those kagglers who made tons of work but didn’t survive in the big shakeup.
\n
I don't know about this. The hosts certainly got something. They got the best (advertised) minds in the DS field to direct their attention at this problem for three months. It's just that the determination is clear that at least for now, the problem of determining the MGMT genetic marker from the modalities presented at the resolutions provided isn't tractable. Science ends like that sometimes. In the future, this problem might be possible with higher resolution scans, or with additional clinical biomarkers added into the mix. As for those that got shaken down, it would be nice if all the submissions had their private scores revealed because that'd help people learn a little about what worked and what didn't. But I understand withholding that information. The labeled test set probably was very expensive to produce and they want to use it again for other purposes, which wouldn't be possible if the labels became known widespread.
", "votes": null, "replies": [ { "id": 1549453, "author_name": "mutantspore", "author_url": "", "post_date": "10/19/2021 00:19:36", "content": "\n\nScience ends like that sometimes
\n
exactly! and is part of the scientific process
", "votes": null, "replies": [] }, { "id": 1549873, "author_name": "sandorkonya", "author_url": "", "post_date": "10/19/2021 08:54:26", "content": "@authman you're absolutely right on that!
", "votes": null, "replies": [] } ] }, { "id": 1549466, "author_name": "lars123", "author_url": "", "post_date": "10/19/2021 00:54:16", "content": "You are right that 60% is clinically useless. I wonder how some papers have achieved 90% AUC with just T2 weighted images… Some papers should be re-examined.
", "votes": null, "replies": [ { "id": 1549663, "author_name": "drtausamaru", "author_url": "", "post_date": "10/19/2021 05:58:18", "content": "I wonder how our score will change if we have 100x training data.
", "votes": null, "replies": [] }, { "id": 1549690, "author_name": "lililycai", "author_url": "", "post_date": "10/19/2021 06:18:09", "content": "I don’t think 100x training data will significantly improve the score. This kind of task is just similar to those that ask you to predict their genetic mutation based on their face images. The two things are hardly related, and you wouldn’t expect the model to predict well.
", "votes": null, "replies": [] } ] }, { "id": 1549659, "author_name": "steaphan", "author_url": "", "post_date": "10/19/2021 05:56:25", "content": "I will say this is definitely waste of my time on this comp.
", "votes": null, "replies": [ { "id": 1549691, "author_name": "lililycai", "author_url": "", "post_date": "10/19/2021 06:19:01", "content": "Totally agree. Be careful when you choose to join a comp.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1549418": "1. ~0.6 AUC score shows that it is really a hard task to predict genetic marker based on MRI image. Future work of genetic marker prediction that based on pure MRI images might be given up.\n2. The poor prediction performance makes the model useless in clinical settings, as you still need to sequence patients’ biopsies to get the genetic result.\n3. It’s a double loss for the host(they didn’t get what they want) and for those kagglers who made tons of work but didn’t survive in the big shakeup.", "1549446": "> It’s a double loss for the host(they didn’t get what they want) and for those kagglers who made tons of work but didn’t survive in the big shakeup.\n\nI don't know about this. The hosts certainly got something. They got the best (advertised) minds in the DS field to direct their attention at this problem for three months. It's just that the determination is clear that at least for now, the problem of determining the MGMT genetic marker from the modalities presented at the resolutions provided isn't tractable. Science ends like that sometimes. In the future, this problem might be possible with higher resolution scans, or with additional clinical biomarkers added into the mix. As for those that got shaken down, it would be nice if all the submissions had their private scores revealed because that'd help people learn a little about what worked and what didn't. But I understand withholding that information. The labeled test set probably was very expensive to produce and they want to use it again for other purposes, which wouldn't be possible if the labels became known widespread.", "1549453": "> Science ends like that sometimes\n\nexactly! and is part of the scientific process", "1549466": "You are right that 60% is clinically useless. I wonder how some papers have achieved 90% AUC with just T2 weighted images... Some papers should be re-examined.", "1549659": "I will say this is definitely waste of my time on this comp.", "1549663": "I wonder how our score will change if we have 100x training data.", "1549690": "I don’t think 100x training data will significantly improve the score. This kind of task is just similar to those that ask you to predict their genetic mutation based on their face images. The two things are hardly related, and you wouldn’t expect the model to predict well.", "1549691": "Totally agree. Be careful when you choose to join a comp.", "1549873": "authman you're absolutely right on that!" }, "source": "meta" }