![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F15806223%2Ffd9e90f351b3f00e8e653527f7b31df0%2FScreenshot%202024-07-20%20at%2017.36.32.png?generation=1721468204017089&alt=media\")
I was reading this tutorial about MR images, https://www.kaggle.com/code/gzuidhof/full-preprocessing-tutorial.
\nI found that a step that it recommend doing to resampling the dcm images to 1mm size in real world. I haven't tried this on my experient yet. But I assume it help models' performance. Or simply the change is insignificant?
\n\n\nResampling
\n
\n A scan may have a pixel spacing of [2.5, 0.5, 0.5], which means that the distance between slices is 2.5 millimeters. For a different scan this may be [1.5, 0.725, 0.725], this can be problematic for automatic analysis (e.g. using ConvNets)!
\n A common method of dealing with this is resampling the full dataset to a certain isotropic resolution. If we choose to resample everything to 1mm1mm1mm pixels we can use 3D convnets without worrying about learning zoom/slice thickness invariance.
\n Whilst this may seem like a very simple step, it has quite some edge cases due to rounding. Also, it takes quite a while.
\n Below code worked well for us (and deals with the edge cases):
():\n \n spacing = np.array([scan[].SliceThickness] + scan[].PixelSpacing, dtype=np.float32)\n\n resize_factor = spacing / new_spacing\n new_real_shape = image.shape * resize_factor\n new_shape = np.(new_real_shape)\n real_resize_factor = new_shape / image.shape\n new_spacing = spacing / real_resize_factor\n\n image = scipy.ndimage.interpolation.zoom(image, real_resize_factor, mode=)\n\n image, new_spacing\nI was reading this tutorial about MR images, https://www.kaggle.com/code/gzuidhof/full-preprocessing-tutorial.
\nI found that a step that it recommend doing to resampling the dcm images to 1mm size in real world. I haven't tried this on my experient yet. But I assume it help models' performance. Or simply the change is insignificant?
\n\n\nResampling
\n
\n A scan may have a pixel spacing of [2.5, 0.5, 0.5], which means that the distance between slices is 2.5 millimeters. For a different scan this may be [1.5, 0.725, 0.725], this can be problematic for automatic analysis (e.g. using ConvNets)!
\n A common method of dealing with this is resampling the full dataset to a certain isotropic resolution. If we choose to resample everything to 1mm1mm1mm pixels we can use 3D convnets without worrying about learning zoom/slice thickness invariance.
\n Whilst this may seem like a very simple step, it has quite some edge cases due to rounding. Also, it takes quite a while.
\n Below code worked well for us (and deals with the edge cases):
():\n \n spacing = np.array([scan[].SliceThickness] + scan[].PixelSpacing, dtype=np.float32)\n\n resize_factor = spacing / new_spacing\n new_real_shape = image.shape * resize_factor\n new_shape = np.(new_real_shape)\n real_resize_factor = new_shape / image.shape\n new_spacing = spacing / real_resize_factor\n\n image = scipy.ndimage.interpolation.zoom(image, real_resize_factor, mode=)\n\n image, new_spacing\n\". Also, it takes quite a while.\"
\nanother way to resample is to use pytorch F interpolation for batch size = N,C,D,H,W
", "rawMarkdown": "\". Also, it takes quite a while.\"\n\nanother way to resample is to use pytorch F interpolation for batch size = N,C,D,H,W", "votes": null }, { "id": "2929129", "postDate": "07/19/2024 21:03:05", "content": "This will most likely hurt and not help because the slice thickness is much greater than the pixel spacing.
\nFor example, series_id 108597120 has a pixel spacing of 0.25 x 0.25 mm and a slice thickness of 3.5 mm, meaning each voxel is 0.25 x 0.25 x 3.5 mm. If you resample this to 1 x 1 x 1 mm, you are downsampling the in-plane resolution by 4 and upsampling the out-of-plane resolution by 3.5 (which is only adding redundant information).
That is true, but what if we chose the median spacing of all the training image series? That would alleviate some of the downsampling and upsampling issues mentioned above.
\nIf I understand correctly, nnU-Net, does something similar to this.
Can you explain please, how did you figured out a pixel spacing and a slice thickness?
", "rawMarkdown": "Can you explain please, how did you figured out a pixel spacing and a slice thickness?", "votes": null }, { "id": "2929723", "postDate": "07/20/2024 09:36:45", "content": "It's in the dicom metadata
\n
thank you!
", "rawMarkdown": "thank you!", "votes": null }, { "id": "2931847", "postDate": "07/22/2024 11:49:49", "content": "I think resampling all studies to a consistent spacing (e.g., 0.3 x 0.3 x 4 mm) is reasonable. But I still don't think isotropic resampling is the best idea.
", "rawMarkdown": "I think resampling all studies to a consistent spacing (e.g., 0.3 x 0.3 x 4 mm) is reasonable. But I still don't think isotropic resampling is the best idea.", "votes": null }, { "id": "2937891", "postDate": "07/27/2024 14:45:35", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F3df64bc42e897cd0c8f25fc91778a9ad%2FSelection_198.png?generation=1722091481735297&alt=media\")
i note that some axial scans are not cuboid. It could be a parallelepiped.
\n(notice that BOTH y and z are changing in the position array)
It says 18x640x640 but is trully that volume? Because there is some axial with double shape along the slices.
", "rawMarkdown": "It says 18x640x640 but is trully that volume? Because there is some axial with double shape along the slices.", "votes": null } ], "comments": [ { "id": 2928598, "author_name": "hengck23", "author_url": "", "post_date": "07/19/2024 12:08:02", "content": "\". Also, it takes quite a while.\"
\nanother way to resample is to use pytorch F interpolation for batch size = N,C,D,H,W
", "votes": null, "replies": [] }, { "id": 2929129, "author_name": "vaillant", "author_url": "", "post_date": "07/19/2024 21:03:05", "content": "This will most likely hurt and not help because the slice thickness is much greater than the pixel spacing.
\nFor example, series_id 108597120 has a pixel spacing of 0.25 x 0.25 mm and a slice thickness of 3.5 mm, meaning each voxel is 0.25 x 0.25 x 3.5 mm. If you resample this to 1 x 1 x 1 mm, you are downsampling the in-plane resolution by 4 and upsampling the out-of-plane resolution by 3.5 (which is only adding redundant information).
That is true, but what if we chose the median spacing of all the training image series? That would alleviate some of the downsampling and upsampling issues mentioned above.
\nIf I understand correctly, nnU-Net, does something similar to this.
I think resampling all studies to a consistent spacing (e.g., 0.3 x 0.3 x 4 mm) is reasonable. But I still don't think isotropic resampling is the best idea.
", "votes": null, "replies": [] } ] }, { "id": 2929713, "author_name": "samson8", "author_url": "", "post_date": "07/20/2024 09:22:32", "content": "Can you explain please, how did you figured out a pixel spacing and a slice thickness?
", "votes": null, "replies": [ { "id": 2929723, "author_name": "llleeeoooh", "author_url": "", "post_date": "07/20/2024 09:36:45", "content": "It's in the dicom metadata
\n
thank you!
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2937891, "author_name": "hengck23", "author_url": "", "post_date": "07/27/2024 14:45:35", "content": "
i note that some axial scans are not cuboid. It could be a parallelepiped.
\n(notice that BOTH y and z are changing in the position array)
It says 18x640x640 but is trully that volume? Because there is some axial with double shape along the slices.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2928196": "I was reading this tutorial about MR images, https://www.kaggle.com/code/gzuidhof/full-preprocessing-tutorial.\n\nI found that a step that it recommend doing to resampling the dcm images to 1mm size in real world. I haven't tried this on my experient yet. But I assume it help models' performance. Or simply the change is insignificant?\n\n>Resampling\nA scan may have a pixel spacing of [2.5, 0.5, 0.5], which means that the distance between slices is 2.5 millimeters. For a different scan this may be [1.5, 0.725, 0.725], this can be problematic for automatic analysis (e.g. using ConvNets)!\nA common method of dealing with this is resampling the full dataset to a certain isotropic resolution. If we choose to resample everything to 1mm1mm1mm pixels we can use 3D convnets without worrying about learning zoom/slice thickness invariance.\nWhilst this may seem like a very simple step, it has quite some edge cases due to rounding. Also, it takes quite a while.\nBelow code worked well for us (and deals with the edge cases):\n\n```python\ndef resample(image, scan, new_spacing=[1,1,1]):\n # Determine current pixel spacing\n spacing = np.array([scan[0].SliceThickness] + scan[0].PixelSpacing, dtype=np.float32)\n\n resize_factor = spacing / new_spacing\n new_real_shape = image.shape * resize_factor\n new_shape = np.round(new_real_shape)\n real_resize_factor = new_shape / image.shape\n new_spacing = spacing / real_resize_factor\n \n image = scipy.ndimage.interpolation.zoom(image, real_resize_factor, mode='nearest')\n \n return image, new_spacing\n```", "2928598": "\". Also, it takes quite a while.\"\n\nanother way to resample is to use pytorch F interpolation for batch size = N,C,D,H,W", "2929129": "This will most likely hurt and not help because the slice thickness is much greater than the pixel spacing. \n\nFor example, `series_id` 108597120 has a pixel spacing of 0.25 x 0.25 mm and a slice thickness of 3.5 mm, meaning each voxel is 0.25 x 0.25 x 3.5 mm. If you resample this to 1 x 1 x 1 mm, you are downsampling the in-plane resolution by 4 and upsampling the out-of-plane resolution by 3.5 (which is only adding redundant information).", "2929441": "That is true, but what if we chose the median spacing of all the training image series? That would alleviate some of the downsampling and upsampling issues mentioned above.\nIf I understand correctly, [nnU-Net](https://github.com/MIC-DKFZ/nnUNet), does something similar to this.", "2929713": "Can you explain please, how did you figured out a pixel spacing and a slice thickness?", "2929723": "It's in the dicom metadata\n\n", "2929730": "thank you!", "2931847": "I think resampling all studies to a consistent spacing (e.g., 0.3 x 0.3 x 4 mm) is reasonable. But I still don't think isotropic resampling is the best idea.", "2937891": "\n\ni note that some axial scans are not cuboid. It could be a parallelepiped.\n(notice that BOTH y and z are changing in the position array)", "2938014": "It says 18x640x640 but is trully that volume? Because there is some axial with double shape along the slices." }, "source": "meta" }