{ "id": 221117, "title": "How to use external dataset?", "url": "/competitions/ranzcr-clip-catheter-line-classification/discussion/221117", "author_name": "Zekun", "post_date": "2021-02-21T09:08:56.115000", "votes": 5, "comment_count": 16, "views": 0, "content": "

If we use external dataset for pretraining,we should train on their original task or this games task ?

", "messages": [ { "id": 1212489, "postDate": "2021-02-21T09:08:56.117Z", "content": "

If we use external dataset for pretraining,we should train on their original task or this games task ?

", "rawMarkdown": "If we use external dataset for pretraining,we should train on their original task or this games task ?", "votes": 5 }, { "id": 1212623, "postDate": "2021-02-21T12:08:22.610Z", "content": "

Well, if you can find external data with the right labels for this task, that would be ideal. However, the x-ray datasets I'm aware off mostly have labels for diagnoses rather than this particular task (I'd love to find a dataset directly for this task, let me know if you know one…).

\n

I guess if enough of the x-rays in other datasets show catheters/lines, you could extract those and try labelling them. However, that's a big project.

\n

More likely, external data can teach neural networks what structures in x-rays are important (hopefully those other tasks lead to learning features that also help in this competition, which is not guaranteed).

", "rawMarkdown": "Well, if you can find external data with the right labels for this task, that would be ideal. However, the x-ray datasets I'm aware off mostly have labels for diagnoses rather than this particular task (I'd love to find a dataset directly for this task, let me know if you know one...). \n\nI guess if enough of the x-rays in other datasets show catheters/lines, you could extract those and try labelling them. However, that's a big project.\n\nMore likely, external data can teach neural networks what structures in x-rays are important (hopefully those other tasks lead to learning features that also help in this competition, which is not guaranteed).", "votes": 2, "replies": [ { "id": 1212668, "postDate": "2021-02-21T13:17:56.397Z", "rawMarkdown": "", "votes": 2, "isDeleted": true }, { "id": 1212675, "postDate": "2021-02-21T13:27:29.337Z", "content": "

They may overfit….

", "rawMarkdown": "They may overfit....", "votes": 1 }, { "id": 1212679, "postDate": "2021-02-21T13:30:54.457Z", "content": "

the tracheostomy tube dataset is very useful if you are doing segmentation, just saying… :)

", "rawMarkdown": "the tracheostomy tube dataset is very useful if you are doing segmentation, just saying... :)", "votes": 2 }, { "id": 1212691, "postDate": "2021-02-21T13:41:18.523Z", "content": "

Is that not a somewhat different task (if I read those comments)?

", "rawMarkdown": "Is that not a somewhat different task (if I read those comments)?", "votes": 1 }, { "id": 1212692, "postDate": "2021-02-21T13:41:35.440Z", "content": "

Can i ask which labels of our dataset corresponds in it?

", "rawMarkdown": "Can i ask which labels of our dataset corresponds in it?" }, { "id": 1212696, "postDate": "2021-02-21T13:45:46.487Z", "content": "

Tracheostomy tubes is a different type of ETT.

\n

So ETT - Normal can be seen as ETT - Normal - Tracheostomy if file is in that dataset. and ETT - Normal - Endotracheal if not in that dataset

", "rawMarkdown": "Tracheostomy tubes is a different type of ETT. \n\nSo `ETT - Normal` can be seen as `ETT - Normal - Tracheostomy` if file is in that dataset. and `ETT - Normal - Endotracheal` if not in that dataset", "votes": 2 }, { "id": 1212702, "postDate": "2021-02-21T13:50:00.030Z", "content": "

@raddar can you please say what labels are corresponding in NIH chest 14 dataset?

\n
\n

So ETT - Normal can be seen as ETT - Normal - Tracheostomy if file is in that dataset. and ETT - Normal - Endotracheal if not in that dataset

\n
\n

that means we are splitting ETT-Normal into 2

", "rawMarkdown": "@raddar can you please say what labels are corresponding in NIH chest 14 dataset?\n\n>So ETT - Normal can be seen as ETT - Normal - Tracheostomy if file is in that dataset. and ETT - Normal - Endotracheal if not in that dataset\n\nthat means we are splitting ETT-Normal into 2", "votes": 1 }, { "id": 1212707, "postDate": "2021-02-21T13:56:46.730Z", "content": "

We need to relabel , right ?

", "rawMarkdown": "We need to relabel , right ?", "votes": 1 }, { "id": 1212709, "postDate": "2021-02-21T14:00:00.040Z", "content": "

You need to look at this creatively :) there is no good or bad way of doing things. Just experiment

", "rawMarkdown": "You need to look at this creatively :) there is no good or bad way of doing things. Just experiment", "votes": 3 }, { "id": 1212713, "postDate": "2021-02-21T14:02:03.957Z", "content": "

Thanks for your reply !

", "rawMarkdown": "Thanks for your reply !", "votes": 1 }, { "id": 1212718, "postDate": "2021-02-21T14:04:52.487Z", "content": "

can you @raddar please refer/say what labels are corresponding in NIH chest 14 dataset? please. And thanks for last reply

", "rawMarkdown": "can you @raddar please refer/say what labels are corresponding in NIH chest 14 dataset? please. And thanks for last reply", "votes": 1 }, { "id": 1212720, "postDate": "2021-02-21T14:07:18.200Z", "content": "

NIH chest 14 does not have these labels.

", "rawMarkdown": "NIH chest 14 does not have these labels.", "votes": 2 }, { "id": 1212726, "postDate": "2021-02-21T14:21:25.517Z", "content": "

thanks for your reply @raddar . but i saw you saying in discussion chest14 contain lots of abnormal patients and i think also CVC
\nis that true?

", "rawMarkdown": "thanks for your reply @raddar . but i saw you saying in discussion chest14 contain lots of abnormal patients and i think also CVC\nis that true?" }, { "id": 1212729, "postDate": "2021-02-21T14:22:36.263Z", "content": "

I think he wants to know how to use these dataset…

", "rawMarkdown": "I think he wants to know how to use these dataset..." }, { "id": 1212752, "postDate": "2021-02-21T14:49:30.047Z", "content": "

The labels you get in chest 14 seem to more or less match the labels (but without giving bounding boxes for most) you have in the VinBigData Chest X-ray Abnormalities Detection, not this competition.

", "rawMarkdown": "The labels you get in chest 14 seem to more or less match the labels (but without giving bounding boxes for most) you have in the [VinBigData Chest X-ray Abnormalities Detection](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/overview), not this competition.", "votes": 1 } ] } ], "comments": [ { "id": 1212623, "author_name": "Björn", "author_url": "", "post_date": "2021-02-21T12:08:22.610000", "content": "

Well, if you can find external data with the right labels for this task, that would be ideal. However, the x-ray datasets I'm aware off mostly have labels for diagnoses rather than this particular task (I'd love to find a dataset directly for this task, let me know if you know one…).

\n

I guess if enough of the x-rays in other datasets show catheters/lines, you could extract those and try labelling them. However, that's a big project.

\n

More likely, external data can teach neural networks what structures in x-rays are important (hopefully those other tasks lead to learning features that also help in this competition, which is not guaranteed).

", "votes": 2, "replies": [ { "id": 1212668, "author_name": "", "author_url": "", "post_date": "2021-02-21T13:17:56.397000", "content": "", "votes": 2, "replies": [] }, { "id": 1212675, "author_name": "Zekun", "author_url": "", "post_date": "2021-02-21T13:27:29.337000", "content": "

They may overfit….

", "votes": 1, "replies": [] }, { "id": 1212679, "author_name": "raddar", "author_url": "", "post_date": "2021-02-21T13:30:54.457000", "content": "

the tracheostomy tube dataset is very useful if you are doing segmentation, just saying… :)

", "votes": 2, "replies": [] }, { "id": 1212691, "author_name": "Björn", "author_url": "", "post_date": "2021-02-21T13:41:18.523000", "content": "

Is that not a somewhat different task (if I read those comments)?

", "votes": 1, "replies": [] }, { "id": 1212692, "author_name": "Mohammed Rizin V K", "author_url": "", "post_date": "2021-02-21T13:41:35.440000", "content": "

Can i ask which labels of our dataset corresponds in it?

", "votes": 0, "replies": [] }, { "id": 1212696, "author_name": "raddar", "author_url": "", "post_date": "2021-02-21T13:45:46.487000", "content": "

Tracheostomy tubes is a different type of ETT.

\n

So ETT - Normal can be seen as ETT - Normal - Tracheostomy if file is in that dataset. and ETT - Normal - Endotracheal if not in that dataset

", "votes": 2, "replies": [] }, { "id": 1212702, "author_name": "Mohammed Rizin V K", "author_url": "", "post_date": "2021-02-21T13:50:00.030000", "content": "

@raddar can you please say what labels are corresponding in NIH chest 14 dataset?

\n
\n

So ETT - Normal can be seen as ETT - Normal - Tracheostomy if file is in that dataset. and ETT - Normal - Endotracheal if not in that dataset

\n
\n

that means we are splitting ETT-Normal into 2

", "votes": 1, "replies": [] }, { "id": 1212707, "author_name": "Zekun", "author_url": "", "post_date": "2021-02-21T13:56:46.730000", "content": "

We need to relabel , right ?

", "votes": 1, "replies": [] }, { "id": 1212709, "author_name": "raddar", "author_url": "", "post_date": "2021-02-21T14:00:00.040000", "content": "

You need to look at this creatively :) there is no good or bad way of doing things. Just experiment

", "votes": 3, "replies": [] }, { "id": 1212713, "author_name": "Zekun", "author_url": "", "post_date": "2021-02-21T14:02:03.957000", "content": "

Thanks for your reply !

", "votes": 1, "replies": [] }, { "id": 1212718, "author_name": "Mohammed Rizin V K", "author_url": "", "post_date": "2021-02-21T14:04:52.487000", "content": "

can you @raddar please refer/say what labels are corresponding in NIH chest 14 dataset? please. And thanks for last reply

", "votes": 1, "replies": [] }, { "id": 1212720, "author_name": "raddar", "author_url": "", "post_date": "2021-02-21T14:07:18.200000", "content": "

NIH chest 14 does not have these labels.

", "votes": 2, "replies": [] }, { "id": 1212726, "author_name": "Mohammed Rizin V K", "author_url": "", "post_date": "2021-02-21T14:21:25.517000", "content": "

thanks for your reply @raddar . but i saw you saying in discussion chest14 contain lots of abnormal patients and i think also CVC
\nis that true?

", "votes": 0, "replies": [] }, { "id": 1212729, "author_name": "Zekun", "author_url": "", "post_date": "2021-02-21T14:22:36.263000", "content": "

I think he wants to know how to use these dataset…

", "votes": 0, "replies": [] }, { "id": 1212752, "author_name": "Björn", "author_url": "", "post_date": "2021-02-21T14:49:30.047000", "content": "

The labels you get in chest 14 seem to more or less match the labels (but without giving bounding boxes for most) you have in the VinBigData Chest X-ray Abnormalities Detection, not this competition.

", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "1212489": "If we use external dataset for pretraining,we should train on their original task or this games task ?", "1212623": "Well, if you can find external data with the right labels for this task, that would be ideal. However, the x-ray datasets I'm aware off mostly have labels for diagnoses rather than this particular task (I'd love to find a dataset directly for this task, let me know if you know one...). \n\nI guess if enough of the x-rays in other datasets show catheters/lines, you could extract those and try labelling them. However, that's a big project.\n\nMore likely, external data can teach neural networks what structures in x-rays are important (hopefully those other tasks lead to learning features that also help in this competition, which is not guaranteed)." } }