![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4791839%2Fb3192382344b4935a5743d7b339780f3%2FScreenshot%20from%202020-08-26%2017-27-27.png?generation=1598459892612918&alt=media\")
I'm struggling to convert the 2D dicoms to 3D arrays. I followed this notebook but I'm finding issues:
\nhttps://www.kaggle.com/allunia/pulmonary-fibrosis-dicom-preprocessing
For example, see the image below for subject ID00355637202295106567614. The scan seems to be squashed in the z axis. There are 37 slices, so the shape before resampling is(37, 512, 512).
\nApparently the slice thickness is 1, and the pixel spacing is 0.679688, i.e. [1.0, 0.679688, 0.679688]. Therefore the shape after resampling becomes (37, 348, 348), but this is still squashed in the z axis.
Is the slice thickness wrong??
\n
I'm struggling to convert the 2D dicoms to 3D arrays. I followed this notebook but I'm finding issues:
\nhttps://www.kaggle.com/allunia/pulmonary-fibrosis-dicom-preprocessing
For example, see the image below for subject ID00355637202295106567614. The scan seems to be squashed in the z axis. There are 37 slices, so the shape before resampling is(37, 512, 512).
\nApparently the slice thickness is 1, and the pixel spacing is 0.679688, i.e. [1.0, 0.679688, 0.679688]. Therefore the shape after resampling becomes (37, 348, 348), but this is still squashed in the z axis.
Is the slice thickness wrong??
\n
Slice thickness is different than slice spacing. I'm guessing your example is 1 mm slices taken every 10 mm. This is like takes a loaf of sliced bread and only taking every tenth slice. Many of the studies take every slice, but it is not unusual to have slices with skips in between.
", "rawMarkdown": "Slice thickness is different than slice spacing. I'm guessing your example is 1 mm slices taken every 10 mm. This is like takes a loaf of sliced bread and only taking every tenth slice. Many of the studies take every slice, but it is not unusual to have slices with skips in between.", "votes": null }, { "id": "987407", "postDate": "08/27/2020 08:24:26", "content": "Okay, so I can use the difference between the SliceLocation of adjacent scans, or the difference in ImagePositionPatient[2] to get the slice spacing. However, some patients don't have this information, for example patient ID00128637202219474716089 doesn't have this information. So what to do? I wonder if it is better to just resample everything to a uniform cube, 128x128x128 for instance.
", "rawMarkdown": "Okay, so I can use the difference between the SliceLocation of adjacent scans, or the difference in ImagePositionPatient[2] to get the slice spacing. However, some patients don't have this information, for example patient ID00128637202219474716089 doesn't have this information. So what to do? I wonder if it is better to just resample everything to a uniform cube, 128x128x128 for instance.", "votes": null }, { "id": "987740", "postDate": "08/27/2020 13:36:59", "content": "I tried doing something similar and this is what I did.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1887174%2F434cb4a45e2eff9ed40cb13d06871ec9%2Ffsafsed.png?generation=1598535294840985&alt=media\")
\nI think it's hacky but I read all files for a patient into a 8 * 256 * 256 array and used a 3D convolution on it. Could not really get good validation score for the first try.
@richardoshaw
\nwhat is pixel spacing..
\n\"Apparently the slice thickness is 1, and the pixel spacing is 0.679688, i.e. [1.0, 0.679688, 0.679688]. Therefore the shape after resampling becomes (37, 348, 348), but this is still squashed in the z axis.\"
what is purpose of resampling here.. how is pixel spacing used ?
", "rawMarkdown": "richardoshaw \nwhat is pixel spacing..\n\"Apparently the slice thickness is 1, and the pixel spacing is 0.679688, i.e. [1.0, 0.679688, 0.679688]. Therefore the shape after resampling becomes (37, 348, 348), but this is still squashed in the z axis.\"\n\nwhat is purpose of resampling here.. how is pixel spacing used ?", "votes": null } ], "comments": [ { "id": 986673, "author_name": "richardepstein", "author_url": "", "post_date": "08/26/2020 17:15:11", "content": "Slice thickness is different than slice spacing. I'm guessing your example is 1 mm slices taken every 10 mm. This is like takes a loaf of sliced bread and only taking every tenth slice. Many of the studies take every slice, but it is not unusual to have slices with skips in between.
", "votes": null, "replies": [ { "id": 987407, "author_name": "richardoshaw", "author_url": "", "post_date": "08/27/2020 08:24:26", "content": "Okay, so I can use the difference between the SliceLocation of adjacent scans, or the difference in ImagePositionPatient[2] to get the slice spacing. However, some patients don't have this information, for example patient ID00128637202219474716089 doesn't have this information. So what to do? I wonder if it is better to just resample everything to a uniform cube, 128x128x128 for instance.
", "votes": null, "replies": [] } ] }, { "id": 987740, "author_name": "jonykarki", "author_url": "", "post_date": "08/27/2020 13:36:59", "content": "I tried doing something similar and this is what I did.
\n
\nI think it's hacky but I read all files for a patient into a 8 * 256 * 256 array and used a 3D convolution on it. Could not really get good validation score for the first try.
@richardoshaw
\nwhat is pixel spacing..
\n\"Apparently the slice thickness is 1, and the pixel spacing is 0.679688, i.e. [1.0, 0.679688, 0.679688]. Therefore the shape after resampling becomes (37, 348, 348), but this is still squashed in the z axis.\"
what is purpose of resampling here.. how is pixel spacing used ?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "986632": "I'm struggling to convert the 2D dicoms to 3D arrays. I followed this notebook but I'm finding issues:\nhttps://www.kaggle.com/allunia/pulmonary-fibrosis-dicom-preprocessing\n\nFor example, see the image below for subject ID00355637202295106567614. The scan seems to be squashed in the z axis. There are 37 slices, so the shape before resampling is(37, 512, 512).\nApparently the slice thickness is 1, and the pixel spacing is 0.679688, i.e. [1.0, 0.679688, 0.679688]. Therefore the shape after resampling becomes (37, 348, 348), but this is still squashed in the z axis.\n\nIs the slice thickness wrong??\n\n", "986673": "Slice thickness is different than slice spacing. I'm guessing your example is 1 mm slices taken every 10 mm. This is like takes a loaf of sliced bread and only taking every tenth slice. Many of the studies take every slice, but it is not unusual to have slices with skips in between.", "987407": "Okay, so I can use the difference between the SliceLocation of adjacent scans, or the difference in ImagePositionPatient[2] to get the slice spacing. However, some patients don't have this information, for example patient ID00128637202219474716089 doesn't have this information. So what to do? I wonder if it is better to just resample everything to a uniform cube, 128x128x128 for instance.", "987740": "I tried doing something similar and this is what I did.\n\nI think it's hacky but I read all files for a patient into a 8 * 256 * 256 array and used a 3D convolution on it. Could not really get good validation score for the first try.", "988801": "richardoshaw \nwhat is pixel spacing..\n\"Apparently the slice thickness is 1, and the pixel spacing is 0.679688, i.e. [1.0, 0.679688, 0.679688]. Therefore the shape after resampling becomes (37, 348, 348), but this is still squashed in the z axis.\"\n\nwhat is purpose of resampling here.. how is pixel spacing used ?" }, "source": "meta" }