{ "id": 344819, "title": "1/4 fractured cases have bounding box annotations", "url": "/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/344819", "author_name": "Jirka", "post_date": "2022-08-16T17:49:30.769000", "votes": 1, "comment_count": 3, "views": 0, "content": "

It seems that only a fraction of the positive cases also has bounding box annotation. Also, I have observed that some boxes focus on the fractured part of the spine compared to others that have annotated the whole spine.

\n

\"results-15-0.png\"

\n
\n

https://www.kaggle.com/code/jirkaborovec/spine-fracture-eda-loading-dicom-3d-browse

\n
\n

So what would be a good way to use annotation as guidance but not the only source of truth?

\n

An alternative would be using just classification on the whole 3D volume:

\n
\n

https://www.kaggle.com/code/jirkaborovec/spinefrac-classif-e2e-lightning-monai-3d

\n
", "messages": [ { "id": 1901522, "postDate": "2022-08-16T17:49:30.770Z", "content": "

It seems that only a fraction of the positive cases also has bounding box annotation. Also, I have observed that some boxes focus on the fractured part of the spine compared to others that have annotated the whole spine.

\n

\"results-15-0.png\"

\n
\n

https://www.kaggle.com/code/jirkaborovec/spine-fracture-eda-loading-dicom-3d-browse

\n
\n

So what would be a good way to use annotation as guidance but not the only source of truth?

\n

An alternative would be using just classification on the whole 3D volume:

\n
\n

https://www.kaggle.com/code/jirkaborovec/spinefrac-classif-e2e-lightning-monai-3d

\n
", "rawMarkdown": "It seems that only a fraction of the positive cases also has bounding box annotation. Also, I have observed that some boxes focus on the fractured part of the spine compared to others that have annotated the whole spine.\n\n[![results-15-0.png](https://i.postimg.cc/3xYyZn1r/results-15-0.png)](https://postimg.cc/JGY4r3SS)\n\n>https://www.kaggle.com/code/jirkaborovec/spine-fracture-eda-loading-dicom-3d-browse\n\nSo what would be a good way to use annotation as guidance but not the only source of truth?\n\nAn alternative would be using just classification on the whole 3D volume:\n\n>https://www.kaggle.com/code/jirkaborovec/spinefrac-classif-e2e-lightning-monai-3d", "votes": 1 }, { "id": 1903291, "postDate": "2022-08-17T09:22:49.813Z", "content": "

This notebook gave me some clues, which i checkd on 3d viewer, atleast 87 segmented images have intensity values(1-7) which segments c1-c7,
\nhttps://www.kaggle.com/code/samuelcortinhas/rsna-fracture-detection-in-depth-eda#6.-Extract-metadata

", "rawMarkdown": "This notebook gave me some clues, which i checkd on 3d viewer, atleast 87 segmented images have intensity values(1-7) which segments c1-c7, \nhttps://www.kaggle.com/code/samuelcortinhas/rsna-fracture-detection-in-depth-eda#6.-Extract-metadata" }, { "id": 1903164, "postDate": "2022-08-17T06:52:09.530Z", "content": "

IMHO, looks like we need to work on the data a bit more, bruteforce whole 3d volume based classification seems too crude, most probably won't give good results.

", "rawMarkdown": "IMHO, looks like we need to work on the data a bit more, bruteforce whole 3d volume based classification seems too crude, most probably won't give good results.", "replies": [ { "id": 1903186, "postDate": "2022-08-17T07:10:12.150Z", "content": "

Yes, I was wondering if it is possible to inject into the training something like teacher's attention?

", "rawMarkdown": "Yes, I was wondering if it is possible to inject into the training something like teacher's attention?", "votes": 1 } ] } ], "comments": [ { "id": 1903291, "author_name": "Krish78", "author_url": "", "post_date": "2022-08-17T09:22:49.813000", "content": "

This notebook gave me some clues, which i checkd on 3d viewer, atleast 87 segmented images have intensity values(1-7) which segments c1-c7,
\nhttps://www.kaggle.com/code/samuelcortinhas/rsna-fracture-detection-in-depth-eda#6.-Extract-metadata

", "votes": 0, "replies": [] }, { "id": 1903164, "author_name": "Krish78", "author_url": "", "post_date": "2022-08-17T06:52:09.530000", "content": "

IMHO, looks like we need to work on the data a bit more, bruteforce whole 3d volume based classification seems too crude, most probably won't give good results.

", "votes": 0, "replies": [ { "id": 1903186, "author_name": "Jirka", "author_url": "", "post_date": "2022-08-17T07:10:12.150000", "content": "

Yes, I was wondering if it is possible to inject into the training something like teacher's attention?

", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "1901522": "It seems that only a fraction of the positive cases also has bounding box annotation. Also, I have observed that some boxes focus on the fractured part of the spine compared to others that have annotated the whole spine.\n\n[![results-15-0.png](https://i.postimg.cc/3xYyZn1r/results-15-0.png)](https://postimg.cc/JGY4r3SS)\n\n>https://www.kaggle.com/code/jirkaborovec/spine-fracture-eda-loading-dicom-3d-browse\n\nSo what would be a good way to use annotation as guidance but not the only source of truth?\n\nAn alternative would be using just classification on the whole 3D volume:\n\n>https://www.kaggle.com/code/jirkaborovec/spinefrac-classif-e2e-lightning-monai-3d", "1903291": "This notebook gave me some clues, which i checkd on 3d viewer, atleast 87 segmented images have intensity values(1-7) which segments c1-c7, \nhttps://www.kaggle.com/code/samuelcortinhas/rsna-fracture-detection-in-depth-eda#6.-Extract-metadata", "1903164": "IMHO, looks like we need to work on the data a bit more, bruteforce whole 3d volume based classification seems too crude, most probably won't give good results." } }