Which one's better? or there would be any difference in the performance? Normalizing each slice and then stacking the images to form 3d image/volume or stacking the images, form a 3d image/volume and then normalize?
", "messages": [ { "id": "1501332", "postDate": "09/03/2021 05:59:44", "content": "Which one's better? or there would be any difference in the performance? Normalizing each slice and then stacking the images to form 3d image/volume or stacking the images, form a 3d image/volume and then normalize?
", "rawMarkdown": "Which one's better? or there would be any difference in the performance? Normalizing each slice and then stacking the images to form 3d image/volume or stacking the images, form a 3d image/volume and then normalize?", "votes": null }, { "id": "1501826", "postDate": "09/03/2021 15:14:03", "content": "I think normalization will be tricky .. especially across studies.
\nDICOM images are larger than 8 bit and frequently have pixel ranges from 0-1000+ . Standard linear normalization to 8 bit (0-255) can eliminate contrast that is necessary to separate tissue types.
\n'Windowing' with a LUT is how DICOM image viewers allow radiologists to bin the pixels to demonstrate necessary brightness and contrast for specific tissue types.
\nSome of the issue with normalization:
\nI made a notebook that demonstrates how to apply a LUT to maintain proper contrast across a series -> https://www.kaggle.com/davidbroberts/standardizing-mr-images
", "rawMarkdown": "I think normalization will be tricky .. especially across studies. \n\nDICOM images are larger than 8 bit and frequently have pixel ranges from 0-1000+ . Standard linear normalization to 8 bit (0-255) can eliminate contrast that is necessary to separate tissue types.\n\n'Windowing' with a LUT is how DICOM image viewers allow radiologists to bin the pixels to demonstrate necessary brightness and contrast for specific tissue types.\n\nSome of the issue with normalization:\n- Scans performed on various modality units with different protocols, varying magnet strength, post-processing etc.\n- These images have had the bone removed, yet some bone pixels remain, causing outliers in the pixel distribution that skew distribution when normalized.\n- Various acquisition matrices (image sizes)\n\nI made a notebook that demonstrates how to apply a LUT to maintain proper contrast across a series -> https://www.kaggle.com/davidbroberts/standardizing-mr-images", "votes": null }, { "id": "1501963", "postDate": "09/03/2021 17:12:21", "content": "I agree with you, also not much enhancing could be done on a stack of images like, histogram eq which you have done in your notebook which enhances the tumor region very well. But since the problem statement states that the promoter MGMT is a genetic sequence in the tumor, I was wondering if classifier would have any advantage of normalizing the stack rather than slices. BTW your notebook was great, well documented and neatly presented. Thanks for sharing.
I am thinking about similar thing as well about how do we normalize for the 3D volume, and also how about across different MRI type. I haven't got good way to move forward. Any sharing will be appreciated.
", "rawMarkdown": "I am thinking about similar thing as well about how do we normalize for the 3D volume, and also how about across different MRI type. I haven't got good way to move forward. Any sharing will be appreciated.", "votes": null } ], "comments": [ { "id": 1501826, "author_name": "davidbroberts", "author_url": "", "post_date": "09/03/2021 15:14:03", "content": "I think normalization will be tricky .. especially across studies.
\nDICOM images are larger than 8 bit and frequently have pixel ranges from 0-1000+ . Standard linear normalization to 8 bit (0-255) can eliminate contrast that is necessary to separate tissue types.
\n'Windowing' with a LUT is how DICOM image viewers allow radiologists to bin the pixels to demonstrate necessary brightness and contrast for specific tissue types.
\nSome of the issue with normalization:
\nI made a notebook that demonstrates how to apply a LUT to maintain proper contrast across a series -> https://www.kaggle.com/davidbroberts/standardizing-mr-images
", "votes": null, "replies": [ { "id": 1501963, "author_name": "tanmaymane18", "author_url": "", "post_date": "09/03/2021 17:12:21", "content": "I agree with you, also not much enhancing could be done on a stack of images like, histogram eq which you have done in your notebook which enhances the tumor region very well. But since the problem statement states that the promoter MGMT is a genetic sequence in the tumor, I was wondering if classifier would have any advantage of normalizing the stack rather than slices. BTW your notebook was great, well documented and neatly presented. Thanks for sharing.
I am thinking about similar thing as well about how do we normalize for the 3D volume, and also how about across different MRI type. I haven't got good way to move forward. Any sharing will be appreciated.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1501332": "Which one's better? or there would be any difference in the performance? Normalizing each slice and then stacking the images to form 3d image/volume or stacking the images, form a 3d image/volume and then normalize?", "1501826": "I think normalization will be tricky .. especially across studies. \n\nDICOM images are larger than 8 bit and frequently have pixel ranges from 0-1000+ . Standard linear normalization to 8 bit (0-255) can eliminate contrast that is necessary to separate tissue types.\n\n'Windowing' with a LUT is how DICOM image viewers allow radiologists to bin the pixels to demonstrate necessary brightness and contrast for specific tissue types.\n\nSome of the issue with normalization:\n- Scans performed on various modality units with different protocols, varying magnet strength, post-processing etc.\n- These images have had the bone removed, yet some bone pixels remain, causing outliers in the pixel distribution that skew distribution when normalized.\n- Various acquisition matrices (image sizes)\n\nI made a notebook that demonstrates how to apply a LUT to maintain proper contrast across a series -> https://www.kaggle.com/davidbroberts/standardizing-mr-images", "1501963": "I agree with you, also not much enhancing could be done on a stack of images like, histogram eq which you have done in your notebook which enhances the tumor region very well. But since the problem statement states that the promoter MGMT is a `genetic` sequence in the tumor, I was wondering if classifier would have any advantage of normalizing the stack rather than slices. BTW your notebook was great, well documented and neatly presented. Thanks for sharing.", "1503064": "I am thinking about similar thing as well about how do we normalize for the 3D volume, and also how about across different MRI type. I haven't got good way to move forward. Any sharing will be appreciated." }, "source": "meta" }