![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9961690%2Fef98870b9168423f53dbc57749fbbd7c%2FScreenshot_20240724_001724.png?generation=1721776657392507&alt=media\")
Hi all! I was looking through which slices/instance numbers from the MRI scans are used to classify each condition according to train_series_descriptions.csv, and I noticed the following regarding foraminal narrowing -- let's look at left for simplicity. Below is a histogram of which instance number is used to diagnose the condition according to train_series_descriptions.csv:
\n
(x axis represents percentage through the instance numbers: so 1.dcm would be 0.0, and say 30.dcm or whatever would be 1.0)
\nWhy is this bimodal? If the MRI scans were taken in a consistent direction: say, always right to left, you would expect all the instances to be in the right hand mode; but about a third of them are on the left. If it is indeed true that we don't know which direction the sagittal scan was taken in (left to right or right to left), how can a model hope to distinguish between conditions in the left vs right foramina on the test data?
\nI may well be misinterpreting something so would be glad if anyone has any helpful insights :)
", "messages": [ { "id": "2933768", "postDate": "07/23/2024 23:27:01", "content": "Hi all! I was looking through which slices/instance numbers from the MRI scans are used to classify each condition according to train_series_descriptions.csv, and I noticed the following regarding foraminal narrowing -- let's look at left for simplicity. Below is a histogram of which instance number is used to diagnose the condition according to train_series_descriptions.csv:
\n
(x axis represents percentage through the instance numbers: so 1.dcm would be 0.0, and say 30.dcm or whatever would be 1.0)
\nWhy is this bimodal? If the MRI scans were taken in a consistent direction: say, always right to left, you would expect all the instances to be in the right hand mode; but about a third of them are on the left. If it is indeed true that we don't know which direction the sagittal scan was taken in (left to right or right to left), how can a model hope to distinguish between conditions in the left vs right foramina on the test data?
\nI may well be misinterpreting something so would be glad if anyone has any helpful insights :)
", "rawMarkdown": "Hi all! I was looking through which slices/instance numbers from the MRI scans are used to classify each condition according to train_series_descriptions.csv, and I noticed the following regarding foraminal narrowing -- let's look at left for simplicity. Below is a histogram of which instance number is used to diagnose the condition according to train_series_descriptions.csv:\n\n\n\n(x axis represents percentage through the instance numbers: so 1.dcm would be 0.0, and say 30.dcm or whatever would be 1.0)\n\nWhy is this bimodal? If the MRI scans were taken in a consistent direction: say, always right to left, you would expect all the instances to be in the right hand mode; but about a third of them are on the left. If it is indeed true that we don't know which direction the sagittal scan was taken in (left to right or right to left), how can a model hope to distinguish between conditions in the left vs right foramina on the test data?\n\nI may well be misinterpreting something so would be glad if anyone has any helpful insights :)", "votes": null }, { "id": "2933880", "postDate": "07/24/2024 03:36:30", "content": "Histogram is similar to value counts of patient positions. Check this discussion post. https://www.kaggle.com/competitions/rsna-2024-lumbar-spine-degenerative-classification/discussion/521389
", "rawMarkdown": "Histogram is similar to value counts of patient positions. Check this discussion post. https://www.kaggle.com/competitions/rsna-2024-lumbar-spine-degenerative-classification/discussion/521389", "votes": null }, { "id": "2934310", "postDate": "07/24/2024 11:18:44", "content": "Thanks for point it out. Fortunately they look easy separable. We should also check L1 to S1 in axial just in case?
\nFor L1/L2 diagnosis on subarticular:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8722753%2F684c042f64e6473d42ff667669cb8f83%2Fsubarticular.png?generation=1721832997474015&alt=media\")
\"…how can a model hope to distinguish between conditions in the left vs right foramina on the test data?\" I haven't though into that. We can correct train with train_label_coordinates.csv but obviously no possible corrections to test.
", "rawMarkdown": "\"...how can a model hope to distinguish between conditions in the left vs right foramina on the test data?\" I haven't though into that. We can correct train with train_label_coordinates.csv but obviously no possible corrections to test.", "votes": null }, { "id": "2934619", "postDate": "07/24/2024 15:42:33", "content": "\n\nWhy is this bimodal
\n
Just like you suspected, this is due to the orientation of the series volume. I calculated the difference also:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1303569%2Fd93a5a3c7db7e048f09145103cbc23e6%2FScreenshot%202024-07-24%20210737.png?generation=1721835479773343&alt=media\")
I tried to do what @gunesevitan suggested and correlate it with PatientPosition attribute, but the results are not satisfactory:
| PatientPosition | \nSign (of difference between right and left) | \nValue Counts | \n
|---|---|---|
| HFS | \n1.0 | \n1100 | \n
| HFS | \n-1.0 | \n477 | \n
| FFS | \n1.0 | \n223 | \n
| FFS | \n-1.0 | \n161 | \n
yes but patient position groupwise plots are not unimodal they are also bimodal showing 2-humps.
", "rawMarkdown": "yes but patient position groupwise plots are not unimodal they are also bimodal showing 2-humps.", "votes": null }, { "id": "2937362", "postDate": "07/27/2024 02:36:00", "content": "@coderrkj thanks for provide this informative message
", "rawMarkdown": "coderrkj thanks for provide this informative message", "votes": null }, { "id": "2940125", "postDate": "07/29/2024 19:17:38", "content": "From here I've googleled a bit about IPP and IOP and I've found that.
", "rawMarkdown": "From [here](https://www.kaggle.com/competitions/rsna-2024-lumbar-spine-degenerative-classification/discussion/522956) I've googleled a bit about IPP and IOP and I've found [that](https://stackoverflow.com/questions/68494649/sorting-dicom-images-without-slicelocation-attribute).", "votes": null } ], "comments": [ { "id": 2933880, "author_name": "gunesevitan", "author_url": "", "post_date": "07/24/2024 03:36:30", "content": "Histogram is similar to value counts of patient positions. Check this discussion post. https://www.kaggle.com/competitions/rsna-2024-lumbar-spine-degenerative-classification/discussion/521389
", "votes": null, "replies": [ { "id": 2935728, "author_name": "rohitchaudhari25", "author_url": "", "post_date": "07/25/2024 14:28:47", "content": "yes but patient position groupwise plots are not unimodal they are also bimodal showing 2-humps.
", "votes": null, "replies": [] } ] }, { "id": 2934310, "author_name": "sacuscreed", "author_url": "", "post_date": "07/24/2024 11:18:44", "content": "Thanks for point it out. Fortunately they look easy separable. We should also check L1 to S1 in axial just in case?
\nFor L1/L2 diagnosis on subarticular:
\n
\"…how can a model hope to distinguish between conditions in the left vs right foramina on the test data?\" I haven't though into that. We can correct train with train_label_coordinates.csv but obviously no possible corrections to test.
", "votes": null, "replies": [] }, { "id": 2934619, "author_name": "coderrkj", "author_url": "", "post_date": "07/24/2024 15:42:33", "content": "\n\nWhy is this bimodal
\n
Just like you suspected, this is due to the orientation of the series volume. I calculated the difference also:
\n
I tried to do what @gunesevitan suggested and correlate it with PatientPosition attribute, but the results are not satisfactory:
| PatientPosition | \nSign (of difference between right and left) | \nValue Counts | \n
|---|---|---|
| HFS | \n1.0 | \n1100 | \n
| HFS | \n-1.0 | \n477 | \n
| FFS | \n1.0 | \n223 | \n
| FFS | \n-1.0 | \n161 | \n
@coderrkj thanks for provide this informative message
", "votes": null, "replies": [] } ] }, { "id": 2940125, "author_name": "sacuscreed", "author_url": "", "post_date": "07/29/2024 19:17:38", "content": "From here I've googleled a bit about IPP and IOP and I've found that.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2933768": "Hi all! I was looking through which slices/instance numbers from the MRI scans are used to classify each condition according to train_series_descriptions.csv, and I noticed the following regarding foraminal narrowing -- let's look at left for simplicity. Below is a histogram of which instance number is used to diagnose the condition according to train_series_descriptions.csv:\n\n\n\n(x axis represents percentage through the instance numbers: so 1.dcm would be 0.0, and say 30.dcm or whatever would be 1.0)\n\nWhy is this bimodal? If the MRI scans were taken in a consistent direction: say, always right to left, you would expect all the instances to be in the right hand mode; but about a third of them are on the left. If it is indeed true that we don't know which direction the sagittal scan was taken in (left to right or right to left), how can a model hope to distinguish between conditions in the left vs right foramina on the test data?\n\nI may well be misinterpreting something so would be glad if anyone has any helpful insights :)", "2933880": "Histogram is similar to value counts of patient positions. Check this discussion post. https://www.kaggle.com/competitions/rsna-2024-lumbar-spine-degenerative-classification/discussion/521389", "2934310": "Thanks for point it out. Fortunately they look easy separable. We should also check L1 to S1 in axial just in case?\n\nFor L1/L2 diagnosis on subarticular:\n", "2934601": "\"...how can a model hope to distinguish between conditions in the left vs right foramina on the test data?\" I haven't though into that. We can correct train with train_label_coordinates.csv but obviously no possible corrections to test.", "2934619": ">Why is this bimodal\n\nJust like you suspected, this is due to the orientation of the series volume. I calculated the difference also:\n\n\nI tried to do what @gunesevitan suggested and correlate it with `PatientPosition` attribute, but the results are not satisfactory:\n| PatientPosition | Sign (of difference between right and left) | Value Counts |\n| --- | --- | --- |\n| HFS | 1.0 | 1100 |\n| HFS | -1.0 | 477 |\n| FFS | 1.0 | 223 |\n| FFS | -1.0 | 161 |", "2935728": "yes but patient position groupwise plots are not unimodal they are also bimodal showing 2-humps.", "2937362": "coderrkj thanks for provide this informative message", "2940125": "From [here](https://www.kaggle.com/competitions/rsna-2024-lumbar-spine-degenerative-classification/discussion/522956) I've googleled a bit about IPP and IOP and I've found [that](https://stackoverflow.com/questions/68494649/sorting-dicom-images-without-slicelocation-attribute)." }, "source": "meta" }