{ "id": 222630, "title": "Making sense from annotations + Starter notebook", "url": "/competitions/ranzcr-clip-catheter-line-classification/discussion/222630", "author_name": "", "post_date": "2021-02-28T10:34:31.968239100Z", "votes": 14, "comment_count": 1, "views": 0, "content": "

Hello!

\n

There's a lot of topics and comments on annotations. Some users claim them to be useful for the teacher-student approach (topic by @hengck23, another topic with a detailed strategy and notebook links by @yasufuminakama) or for segmentation approach (topic by @ryches) while some others claim that there are no relevant improvements from these approaches for relatively big models. Some of the annotations highlighted here are probably mislabeled.

\n

I've done a quick EDA showing that out of 9095 samples with annotations available:

\n\n

Regarding the last one, it's been a recent topic by @cdeotte wondering why e.g. CVC - Normal and CVC - Abnormal can be positive for the same sample (the answer is - there can be multiple CVC inserted for different purposes). It's now not so wondering to spot samples where multiple catheters have the same class, e.g.:
\n

\n

Label distribution across annotated samples seems to be fine, e.g. we have only 79 ETT - Abnormal cases and 40 of them do have annotations.
\n

\n

Lastly, I've made this starter notebook for preparing dataset with images annotated via OpenCV methods in .jpeg and .tfrec format.

\n

Hope this helps someone. Happy coding!

", "messages": [ { "id": "1220761", "postDate": "02/28/2021 10:34:31", "content": "

Hello!

\n

There's a lot of topics and comments on annotations. Some users claim them to be useful for the teacher-student approach (topic by @hengck23, another topic with a detailed strategy and notebook links by @yasufuminakama) or for segmentation approach (topic by @ryches) while some others claim that there are no relevant improvements from these approaches for relatively big models. Some of the annotations highlighted here are probably mislabeled.

\n

I've done a quick EDA showing that out of 9095 samples with annotations available:

\n\n

Regarding the last one, it's been a recent topic by @cdeotte wondering why e.g. CVC - Normal and CVC - Abnormal can be positive for the same sample (the answer is - there can be multiple CVC inserted for different purposes). It's now not so wondering to spot samples where multiple catheters have the same class, e.g.:
\n

\n

Label distribution across annotated samples seems to be fine, e.g. we have only 79 ETT - Abnormal cases and 40 of them do have annotations.
\n

\n

Lastly, I've made this starter notebook for preparing dataset with images annotated via OpenCV methods in .jpeg and .tfrec format.

\n

Hope this helps someone. Happy coding!

", "rawMarkdown": "### Hello!\n\nThere's a lot of topics and comments on annotations. Some users claim them to be useful for the teacher-student approach (**[topic](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/205243)** by @hengck23, **[another topic](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/207577)** with a detailed strategy and notebook links by @yasufuminakama) or for segmentation approach (**[topic](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/212698)** by @ryches) while some others claim that there are no relevant improvements from these approaches for relatively big models. Some of the annotations highlighted **[here](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/210064)** are probably mislabeled. \n\nI've done a quick EDA showing that out of 9095 samples with annotations available:\n* 7723 samples are completely annotated i.e. have annotations for all the catheters inserted\n* only 24 samples have incomplete annotations\n* 1349 samples have more than one catheter of the same **class** (not just the same **type**)\n\nRegarding the last one, it's been a **[recent topic](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/221513)** by @cdeotte wondering why e.g. `CVC - Normal` and `CVC - Abnormal` can be positive for the same sample (the answer is - there can be multiple CVC inserted for different purposes). It's now not so wondering to spot samples where multiple catheters have the same class, e.g.:\n\n\nLabel distribution across annotated samples seems to be fine, e.g. we have only 79 `ETT - Abnormal` cases and 40 of them do have annotations.\n\n\nLastly, I've made this **[starter notebook](https://www.kaggle.com/nickuzmenkov/ranzcr-clip-making-sense-from-annotations)** for preparing dataset with images annotated via **OpenCV** methods in `.jpeg` and `.tfrec` format. \n\nHope this helps someone. Happy coding!", "votes": null }, { "id": "1221008", "postDate": "02/28/2021 15:42:02", "content": "

One interesting study is looking at the relationship between the number of catheters of a given type and loss. These multi catheter scenarios appear to be hard for the model

", "rawMarkdown": "One interesting study is looking at the relationship between the number of catheters of a given type and loss. These multi catheter scenarios appear to be hard for the model", "votes": null } ], "comments": [ { "id": 1221008, "author_name": "ryches", "author_url": "", "post_date": "02/28/2021 15:42:02", "content": "

One interesting study is looking at the relationship between the number of catheters of a given type and loss. These multi catheter scenarios appear to be hard for the model

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1220761": "### Hello!\n\nThere's a lot of topics and comments on annotations. Some users claim them to be useful for the teacher-student approach (**[topic](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/205243)** by @hengck23, **[another topic](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/207577)** with a detailed strategy and notebook links by @yasufuminakama) or for segmentation approach (**[topic](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/212698)** by @ryches) while some others claim that there are no relevant improvements from these approaches for relatively big models. Some of the annotations highlighted **[here](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/210064)** are probably mislabeled. \n\nI've done a quick EDA showing that out of 9095 samples with annotations available:\n* 7723 samples are completely annotated i.e. have annotations for all the catheters inserted\n* only 24 samples have incomplete annotations\n* 1349 samples have more than one catheter of the same **class** (not just the same **type**)\n\nRegarding the last one, it's been a **[recent topic](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/221513)** by @cdeotte wondering why e.g. `CVC - Normal` and `CVC - Abnormal` can be positive for the same sample (the answer is - there can be multiple CVC inserted for different purposes). It's now not so wondering to spot samples where multiple catheters have the same class, e.g.:\n\n\nLabel distribution across annotated samples seems to be fine, e.g. we have only 79 `ETT - Abnormal` cases and 40 of them do have annotations.\n\n\nLastly, I've made this **[starter notebook](https://www.kaggle.com/nickuzmenkov/ranzcr-clip-making-sense-from-annotations)** for preparing dataset with images annotated via **OpenCV** methods in `.jpeg` and `.tfrec` format. \n\nHope this helps someone. Happy coding!", "1221008": "One interesting study is looking at the relationship between the number of catheters of a given type and loss. These multi catheter scenarios appear to be hard for the model" }, "source": "meta" }