Hi all, I am new here and thank everyone in the community for providing such broad and detail knowledge.
\nWith the assistance of https://www.kaggle.com/code/enriquezaf/totalsegmentator-offline, we have generated segmentator images for DCM files.
\nHowever, I find myself at an impasse when it comes to integrating these generated images as channels during the training of PNG files in a CNN architecture.
\nI would be interested to know how members of this community have leveraged total_segmentator in their workflows. Additionally, any guidance on employing total_segmentator during inference would be greatly appreciated.
", "messages": [ { "id": 2441901, "postDate": "2023-09-16T14:28:23.257Z", "content": "I use Totalsegmentator to generate masks to extract ROI. In most cases, the masks are relatively good, but there are also some samples where the masks are very poor. For submission, I thought TotalSegmentator's inference speed was too slow, so I use Totalsegmentator inferred pseudo-labels and 206 real labels to train a basic unet for submission.
", "rawMarkdown": "I use Totalsegmentator to generate masks to extract ROI. In most cases, the masks are relatively good, but there are also some samples where the masks are very poor. For submission, I thought TotalSegmentator's inference speed was too slow, so I use Totalsegmentator inferred pseudo-labels and 206 real labels to train a basic unet for submission.", "votes": 4, "replies": [ { "id": 2445976, "postDate": "2023-09-19T07:10:27.853Z", "content": "I try to use TotalSegmentator to generate masks. It takes large amount of time to generate masks for all the series. What is your appraoch to generate pseudo-labels? Do you reduce the size of dcm?
", "rawMarkdown": "I try to use TotalSegmentator to generate masks. It takes large amount of time to generate masks for all the series. What is your appraoch to generate pseudo-labels? Do you reduce the size of dcm? ", "replies": [ { "id": 2446701, "postDate": "2023-09-19T14:53:35.197Z", "content": "I dont reduce the size of dcm, I just use the default setting of TotalSegmentator, It does require a lot of time to generate pseudo-labels. If you dont have too much time, you can check this.
", "rawMarkdown": "I dont reduce the size of dcm, I just use the default setting of TotalSegmentator, It does require a lot of time to generate pseudo-labels. If you dont have too much time, you can check [this](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/436096)." } ] } ] }, { "id": 2443174, "postDate": "2023-09-17T15:06:26.820Z", "content": "I will integrate the masks as colors, since the images are greyscale anyway.
\nAs for inference, I am thinking about not using all the images for inference.
I am trying this approach.
\nHow do you make sure the nii and dcm overlay on each other correctly? Any resource that can help me make sure they are in the same orientation?
In the Dicom metadata you can find the orientation of the Dicom image
", "rawMarkdown": "In the Dicom metadata you can find the orientation of the Dicom image" } ] } ] }, { "id": 2440788, "postDate": "2023-09-15T18:27:53.890Z", "content": "Hi all, I am new here and thank everyone in the community for providing such broad and detail knowledge.
\nWith the assistance of https://www.kaggle.com/code/enriquezaf/totalsegmentator-offline, we have generated segmentator images for DCM files.
\nHowever, I find myself at an impasse when it comes to integrating these generated images as channels during the training of PNG files in a CNN architecture.
\nI would be interested to know how members of this community have leveraged total_segmentator in their workflows. Additionally, any guidance on employing total_segmentator during inference would be greatly appreciated.
", "rawMarkdown": "Hi all, I am new here and thank everyone in the community for providing such broad and detail knowledge. \n\nWith the assistance of [https://www.kaggle.com/code/enriquezaf/totalsegmentator-offline](url), we have generated segmentator images for DCM files.\n\nHowever, I find myself at an impasse when it comes to integrating these generated images as channels during the training of PNG files in a CNN architecture.\n\nI would be interested to know how members of this community have leveraged total_segmentator in their workflows. Additionally, any guidance on employing total_segmentator during inference would be greatly appreciated.\n", "votes": 1 } ], "comments": [ { "id": 2441901, "author_name": "m1dsolo", "author_url": "", "post_date": "2023-09-16T14:28:23.257000", "content": "I use Totalsegmentator to generate masks to extract ROI. In most cases, the masks are relatively good, but there are also some samples where the masks are very poor. For submission, I thought TotalSegmentator's inference speed was too slow, so I use Totalsegmentator inferred pseudo-labels and 206 real labels to train a basic unet for submission.
", "votes": 4, "replies": [ { "id": 2445976, "author_name": "Franklin Shih0617", "author_url": "", "post_date": "2023-09-19T07:10:27.853000", "content": "I try to use TotalSegmentator to generate masks. It takes large amount of time to generate masks for all the series. What is your appraoch to generate pseudo-labels? Do you reduce the size of dcm?
", "votes": 0, "replies": [ { "id": 2446701, "author_name": "m1dsolo", "author_url": "", "post_date": "2023-09-19T14:53:35.197000", "content": "I dont reduce the size of dcm, I just use the default setting of TotalSegmentator, It does require a lot of time to generate pseudo-labels. If you dont have too much time, you can check this.
", "votes": 0, "replies": [] } ] } ] }, { "id": 2443174, "author_name": "Gyula Maloveczky4", "author_url": "", "post_date": "2023-09-17T15:06:26.820000", "content": "I will integrate the masks as colors, since the images are greyscale anyway.
\nAs for inference, I am thinking about not using all the images for inference.
I am trying this approach.
\nHow do you make sure the nii and dcm overlay on each other correctly? Any resource that can help me make sure they are in the same orientation?
In the Dicom metadata you can find the orientation of the Dicom image
", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2441901": "I use Totalsegmentator to generate masks to extract ROI. In most cases, the masks are relatively good, but there are also some samples where the masks are very poor. For submission, I thought TotalSegmentator's inference speed was too slow, so I use Totalsegmentator inferred pseudo-labels and 206 real labels to train a basic unet for submission.", "2443174": "I will integrate the masks as colors, since the images are greyscale anyway.\nAs for inference, I am thinking about not using all the images for inference.", "2440788": "Hi all, I am new here and thank everyone in the community for providing such broad and detail knowledge. \n\nWith the assistance of [https://www.kaggle.com/code/enriquezaf/totalsegmentator-offline](url), we have generated segmentator images for DCM files.\n\nHowever, I find myself at an impasse when it comes to integrating these generated images as channels during the training of PNG files in a CNN architecture.\n\nI would be interested to know how members of this community have leveraged total_segmentator in their workflows. Additionally, any guidance on employing total_segmentator during inference would be greatly appreciated.\n" } }