i read a lot disscution and notbook for a while but there are some things that are realy not clerify to me and I would like to some help here .
\n1)2.5D Methods:
\nMany people tend to use 2.5D methods. Could someone clarify this for me?
\n2) Segmentation Data:
\nI have not been able to figure out how to use segmentation data. I understand that it probably helps to minimize irrelevant images. It might also be used in some attention mechanisms. However, I'm really not sure how to attach one segmentation NIfTI file to another DICOM file. I would appreciate it if anyone could clarify how they can be used.
3)Patient-Level Labeling:
\nBecause labeling is done at the patient level, it seems weak to assign the same label to every image in a session. Furthermore, for patients with more than one session, sometimes the sessions do not agree. How do you deal with this?
4)Converting Images to PNG:
\nI saw some threads that convert the images to PNG format. However, this seems to result in the loss of the ability to build a volume model. How do people use this data?
5)Dealing with Imbalance:
\nI would appreciate some help on how to deal with the imbalance between healthy and injured individuals.
I have been trying to grasp this topic for a long time,reading a lot of discussion and notebooks and would greatly appreciate any help.
\nThank you all
1) 2.5D methods actually use 2D convolution operations for feature extraction, but with different depths of volumes. 2D convolutions only move in two directions, so in theory, they cannot extract information about the Z direction. However, when you use a 2D convolution over a 3D volume (assuming you choose 20 slices in the middle of the CT scan), the 2D conv layer treats the 3D volume as a single photo with 20 channels and tries to extract information. These models can give good results if you determine how many and which slices you will use for input.
", "rawMarkdown": "1) 2.5D methods actually use 2D convolution operations for feature extraction, but with different depths of volumes. 2D convolutions only move in two directions, so in theory, they cannot extract information about the Z direction. However, when you use a 2D convolution over a 3D volume (assuming you choose 20 slices in the middle of the CT scan), the 2D conv layer treats the 3D volume as a single photo with 20 channels and tries to extract information. These models can give good results if you determine how many and which slices you will use for input.", "votes": 1 }, { "id": 2438792, "postDate": "2023-09-14T13:53:40.077Z", "content": "i read a lot disscution and notbook for a while but there are some things that are realy not clerify to me and I would like to some help here .
\n1)2.5D Methods:
\nMany people tend to use 2.5D methods. Could someone clarify this for me?
\n2) Segmentation Data:
\nI have not been able to figure out how to use segmentation data. I understand that it probably helps to minimize irrelevant images. It might also be used in some attention mechanisms. However, I'm really not sure how to attach one segmentation NIfTI file to another DICOM file. I would appreciate it if anyone could clarify how they can be used.
3)Patient-Level Labeling:
\nBecause labeling is done at the patient level, it seems weak to assign the same label to every image in a session. Furthermore, for patients with more than one session, sometimes the sessions do not agree. How do you deal with this?
4)Converting Images to PNG:
\nI saw some threads that convert the images to PNG format. However, this seems to result in the loss of the ability to build a volume model. How do people use this data?
5)Dealing with Imbalance:
\nI would appreciate some help on how to deal with the imbalance between healthy and injured individuals.
I have been trying to grasp this topic for a long time,reading a lot of discussion and notebooks and would greatly appreciate any help.
\nThank you all
2) For example, I used this data to create segmentation model to predict region of interest for example liver, and than used cropped images for classification only for liver.
\n3) In my case I used subsample of images to fit it to volume tensor, for example shape will be 96256256.
2) but than in the infernce the images will not be croped and segmanted, or you semangt also through infernce?
\n3) you subsample randomly?
what you mean by label smoothing? how you deal with diffrent label to same patient over diffrent session ?
", "rawMarkdown": "what you mean by label smoothing? how you deal with diffrent label to same patient over diffrent session ? ", "replies": [ { "id": 2440702, "postDate": "2023-09-15T17:09:22.173Z", "rawMarkdown": "", "isDeleted": true } ] } ] } ], "comments": [ { "id": 2440402, "author_name": "turkenm", "author_url": "", "post_date": "2023-09-15T13:29:53.350000", "content": "1) 2.5D methods actually use 2D convolution operations for feature extraction, but with different depths of volumes. 2D convolutions only move in two directions, so in theory, they cannot extract information about the Z direction. However, when you use a 2D convolution over a 3D volume (assuming you choose 20 slices in the middle of the CT scan), the 2D conv layer treats the 3D volume as a single photo with 20 channels and tries to extract information. These models can give good results if you determine how many and which slices you will use for input.
", "votes": 1, "replies": [] }, { "id": 2440317, "author_name": "Pavel Bakhvalov", "author_url": "", "post_date": "2023-09-15T12:37:55.043000", "content": "2) For example, I used this data to create segmentation model to predict region of interest for example liver, and than used cropped images for classification only for liver.
\n3) In my case I used subsample of images to fit it to volume tensor, for example shape will be 96256256.
2) but than in the infernce the images will not be croped and segmanted, or you semangt also through infernce?
\n3) you subsample randomly?
what you mean by label smoothing? how you deal with diffrent label to same patient over diffrent session ?
", "votes": 0, "replies": [ { "id": 2440702, "author_name": "", "author_url": "", "post_date": "2023-09-15T17:09:22.173000", "content": "", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2440402": "1) 2.5D methods actually use 2D convolution operations for feature extraction, but with different depths of volumes. 2D convolutions only move in two directions, so in theory, they cannot extract information about the Z direction. However, when you use a 2D convolution over a 3D volume (assuming you choose 20 slices in the middle of the CT scan), the 2D conv layer treats the 3D volume as a single photo with 20 channels and tries to extract information. These models can give good results if you determine how many and which slices you will use for input.", "2438792": "i read a lot disscution and notbook for a while but there are some things that are realy not clerify to me and I would like to some help here .\n1)2.5D Methods:\nMany people tend to use 2.5D methods. Could someone clarify this for me?\n2) Segmentation Data:\nI have not been able to figure out how to use segmentation data. I understand that it probably helps to minimize irrelevant images. It might also be used in some attention mechanisms. However, I'm really not sure how to attach one segmentation NIfTI file to another DICOM file. I would appreciate it if anyone could clarify how they can be used.\n\n3)Patient-Level Labeling:\nBecause labeling is done at the patient level, it seems weak to assign the same label to every image in a session. Furthermore, for patients with more than one session, sometimes the sessions do not agree. How do you deal with this?\n\n4)Converting Images to PNG:\nI saw some threads that convert the images to PNG format. However, this seems to result in the loss of the ability to build a volume model. How do people use this data?\n\n5)Dealing with Imbalance:\nI would appreciate some help on how to deal with the imbalance between healthy and injured individuals.\n\nI have been trying to grasp this topic for a long time,reading a lot of discussion and notebooks and would greatly appreciate any help.\nThank you all", "2440317": "2) For example, I used this data to create segmentation model to predict region of interest for example liver, and than used cropped images for classification only for liver.\n3) In my case I used subsample of images to fit it to volume tensor, for example shape will be 96*256*256.", "2439329": "" } }