![](\"https://i.postimg.cc/sD4ZyMGN/sample-df.png\")
I haven't seen much progress on finding which vertebrae is in each image yet so I have just released my NOTEBOOK. The idea is that we can identify the targets from the unique values in the segmentation masks. I have collected this info and stored it in this DATASET.
\n
The next step would be to build a supervised model (using metadata/images) to predict which vertebrae is in each image for all the other patients in the train (& test) set which don't have segmentation masks. The challenging part is preserving the monotonicity (C1->C2->etc). I haven't figured this out yet so I'm sharing this resource to hopefully kickstart some progress.
", "messages": [ { "id": 1898870, "postDate": "2022-08-14T22:53:17.417Z", "content": "I haven't seen much progress on finding which vertebrae is in each image yet so I have just released my NOTEBOOK. The idea is that we can identify the targets from the unique values in the segmentation masks. I have collected this info and stored it in this DATASET.
\n
The next step would be to build a supervised model (using metadata/images) to predict which vertebrae is in each image for all the other patients in the train (& test) set which don't have segmentation masks. The challenging part is preserving the monotonicity (C1->C2->etc). I haven't figured this out yet so I'm sharing this resource to hopefully kickstart some progress.
", "rawMarkdown": "I haven't seen much progress on **finding which vertebrae is in each image** yet so I have just released my [NOTEBOOK](https://www.kaggle.com/code/samuelcortinhas/extracting-vertebrae-c1-c7). The idea is that we can identify the targets from the **unique values in the segmentation masks**. I have collected this info and stored it in this [DATASET](https://www.kaggle.com/datasets/samuelcortinhas/rsna-2022-spine-fracture-detection-metadata/code). \n\n
\n\nThe **next step** would be to **build a supervised model** (using metadata/images) to predict which vertebrae is in each image for **all the other patients** in the train (& test) set which **don't have segmentation masks**. The **challenging** part is **preserving the monotonicity** (C1->C2->etc). I haven't figured this out yet so I'm sharing this resource to hopefully kickstart some progress. ",
"votes": 7
},
{
"id": 1900969,
"postDate": "2022-08-16T11:42:32.963Z",
"content": "Excellent work and great initiative in sharing your progress!
\nDid you try to train a model end2end with this and submit it? Is it improving it's performance?
\nAgain, nice work!
The Devastator.
", "rawMarkdown": "Excellent work and great initiative in sharing your progress!\nDid you try to train a model end2end with this and submit it? Is it improving it's performance? \nAgain, nice work! \n\nThe Devastator.\n", "votes": 1, "replies": [ { "id": 1901595, "postDate": "2022-08-16T19:16:39.467Z", "content": "I am in the process of doing this. Will let you how it goes.
", "rawMarkdown": "I am in the process of doing this. Will let you how it goes." } ] }, { "id": 1898913, "postDate": "2022-08-14T23:57:09.450Z", "content": "A simple enough supervised model gets like 85% accuracy (in my tests), won't making it 2.5d solve your problem with monotonicity most of the time, if that does not work, you can always just make a 3d model.
", "rawMarkdown": "A simple enough supervised model gets like 85% accuracy (in my tests), won't making it 2.5d solve your problem with monotonicity most of the time, if that does not work, you can always just make a 3d model.", "votes": 1, "replies": [ { "id": 1899530, "postDate": "2022-08-15T10:27:22.507Z", "content": "@harshitsheoran thanks for the suggestion. I trained a random forest classifier and got 95% accuracy. The code has been added it to my notebook linked above.
", "rawMarkdown": "@harshitsheoran thanks for the suggestion. I trained a random forest classifier and got 95% accuracy. The code has been added it to my notebook linked above. " } ] }, { "id": 1904524, "postDate": "2022-08-18T08:52:19.253Z", "content": "Big Data Leak! Not good result if the leak is solved, leak is in train_test_split, you are not splitting it by StudyInstanceUID, instead taking really similar values in train and validation
", "rawMarkdown": "Big Data Leak! Not good result if the leak is solved, leak is in train_test_split, you are not splitting it by StudyInstanceUID, instead taking really similar values in train and validation", "votes": 2, "replies": [ { "id": 1913262, "postDate": "2022-08-25T09:03:43.850Z", "content": "Thank you for spotting this! I've fixed it now.
", "rawMarkdown": "Thank you for spotting this! I've fixed it now." } ] } ], "comments": [ { "id": 1900969, "author_name": "The Devastator", "author_url": "", "post_date": "2022-08-16T11:42:32.963000", "content": "Excellent work and great initiative in sharing your progress!
\nDid you try to train a model end2end with this and submit it? Is it improving it's performance?
\nAgain, nice work!
The Devastator.
", "votes": 1, "replies": [ { "id": 1901595, "author_name": "Samuel Cortinhas", "author_url": "", "post_date": "2022-08-16T19:16:39.467000", "content": "I am in the process of doing this. Will let you how it goes.
", "votes": 0, "replies": [] } ] }, { "id": 1898913, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2022-08-14T23:57:09.450000", "content": "A simple enough supervised model gets like 85% accuracy (in my tests), won't making it 2.5d solve your problem with monotonicity most of the time, if that does not work, you can always just make a 3d model.
", "votes": 1, "replies": [ { "id": 1899530, "author_name": "Samuel Cortinhas", "author_url": "", "post_date": "2022-08-15T10:27:22.507000", "content": "@harshitsheoran thanks for the suggestion. I trained a random forest classifier and got 95% accuracy. The code has been added it to my notebook linked above.
", "votes": 0, "replies": [] } ] }, { "id": 1904524, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2022-08-18T08:52:19.253000", "content": "Big Data Leak! Not good result if the leak is solved, leak is in train_test_split, you are not splitting it by StudyInstanceUID, instead taking really similar values in train and validation
", "votes": 2, "replies": [ { "id": 1913262, "author_name": "Samuel Cortinhas", "author_url": "", "post_date": "2022-08-25T09:03:43.850000", "content": "Thank you for spotting this! I've fixed it now.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1898870": "I haven't seen much progress on **finding which vertebrae is in each image** yet so I have just released my [NOTEBOOK](https://www.kaggle.com/code/samuelcortinhas/extracting-vertebrae-c1-c7). The idea is that we can identify the targets from the **unique values in the segmentation masks**. I have collected this info and stored it in this [DATASET](https://www.kaggle.com/datasets/samuelcortinhas/rsna-2022-spine-fracture-detection-metadata/code). \n\n\n\nThe **next step** would be to **build a supervised model** (using metadata/images) to predict which vertebrae is in each image for **all the other patients** in the train (& test) set which **don't have segmentation masks**. The **challenging** part is **preserving the monotonicity** (C1->C2->etc). I haven't figured this out yet so I'm sharing this resource to hopefully kickstart some progress. ",
"1900969": "Excellent work and great initiative in sharing your progress!\nDid you try to train a model end2end with this and submit it? Is it improving it's performance? \nAgain, nice work! \n\nThe Devastator.\n",
"1898913": "A simple enough supervised model gets like 85% accuracy (in my tests), won't making it 2.5d solve your problem with monotonicity most of the time, if that does not work, you can always just make a 3d model.",
"1904524": "Big Data Leak! Not good result if the leak is solved, leak is in train_test_split, you are not splitting it by StudyInstanceUID, instead taking really similar values in train and validation"
}
}