Hi everyone! I am new to this area and I notice some tutorials only use tabular data. I am confused that how should I use dicom files?
\nCan anyone give me some hint?
Hi everyone! I am new to this area and I notice some tutorials only use tabular data. I am confused that how should I use dicom files?
\nCan anyone give me some hint?
There are two aspects, some meta data that is available in the dicom files that you could use and of course you can create 3D images out of them and feed them into some type of conv net. Challenge is the limited amount of training data for large conv nets, so challenge is how to be able to use them without overfitting. One approach could be to first perform some auto-encoding.
", "rawMarkdown": "There are two aspects, some meta data that is available in the dicom files that you could use and of course you can create 3D images out of them and feed them into some type of conv net. Challenge is the limited amount of training data for large conv nets, so challenge is how to be able to use them without overfitting. One approach could be to first perform some auto-encoding.", "votes": null }, { "id": "980776", "postDate": "08/21/2020 21:02:54", "content": "Hi, @peterdekkers101. I am trying to understand how to use 3d images with other features provided in the csv data. Should I derive a feature from the images (like the volume of the lung) and then use it in a classifier?
", "rawMarkdown": "Hi, @peterdekkers101. I am trying to understand how to use 3d images with other features provided in the csv data. Should I derive a feature from the images (like the volume of the lung) and then use it in a classifier?", "votes": null }, { "id": "980794", "postDate": "08/21/2020 21:39:51", "content": "Although I suspect that getting useful information from the images might be very difficult with only so few samples, in theory you can:
\n1) feed the images into a few conv3d layers with some pooling. The kernels of the conv layers should take into consideration that the z dimension is very low compared to x and y.
\n2) Then flatten them.
\n3) The flatten layer you can merge now with other input features.
\n4) Finally output two values (one for the FVC value and one for the confidence level).
You can apply the competition evaluation as a loss function. So it is a single neural network from image to loss. That all being said, there is so few data that a linear regression almost scores as good as more complex networks.
", "rawMarkdown": "Although I suspect that getting useful information from the images might be very difficult with only so few samples, in theory you can: \n\n1) feed the images into a few conv3d layers with some pooling. The kernels of the conv layers should take into consideration that the z dimension is very low compared to x and y. \n2) Then flatten them. \n3) The flatten layer you can merge now with other input features.\n4) Finally output two values (one for the FVC value and one for the confidence level). \n\nYou can apply the competition evaluation as a loss function. So it is a single neural network from image to loss. That all being said, there is so few data that a linear regression almost scores as good as more complex networks.", "votes": null } ], "comments": [ { "id": 977898, "author_name": "peterdekkers101", "author_url": "", "post_date": "08/19/2020 19:34:29", "content": "There are two aspects, some meta data that is available in the dicom files that you could use and of course you can create 3D images out of them and feed them into some type of conv net. Challenge is the limited amount of training data for large conv nets, so challenge is how to be able to use them without overfitting. One approach could be to first perform some auto-encoding.
", "votes": null, "replies": [ { "id": 980776, "author_name": "krishnakumar0", "author_url": "", "post_date": "08/21/2020 21:02:54", "content": "Hi, @peterdekkers101. I am trying to understand how to use 3d images with other features provided in the csv data. Should I derive a feature from the images (like the volume of the lung) and then use it in a classifier?
", "votes": null, "replies": [] }, { "id": 980794, "author_name": "peterdekkers101", "author_url": "", "post_date": "08/21/2020 21:39:51", "content": "Although I suspect that getting useful information from the images might be very difficult with only so few samples, in theory you can:
\n1) feed the images into a few conv3d layers with some pooling. The kernels of the conv layers should take into consideration that the z dimension is very low compared to x and y.
\n2) Then flatten them.
\n3) The flatten layer you can merge now with other input features.
\n4) Finally output two values (one for the FVC value and one for the confidence level).
You can apply the competition evaluation as a loss function. So it is a single neural network from image to loss. That all being said, there is so few data that a linear regression almost scores as good as more complex networks.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "977742": "Hi everyone! I am new to this area and I notice some tutorials only use tabular data. I am confused that how should I use dicom files?\nCan anyone give me some hint?", "977898": "There are two aspects, some meta data that is available in the dicom files that you could use and of course you can create 3D images out of them and feed them into some type of conv net. Challenge is the limited amount of training data for large conv nets, so challenge is how to be able to use them without overfitting. One approach could be to first perform some auto-encoding.", "980776": "Hi, @peterdekkers101. I am trying to understand how to use 3d images with other features provided in the csv data. Should I derive a feature from the images (like the volume of the lung) and then use it in a classifier?", "980794": "Although I suspect that getting useful information from the images might be very difficult with only so few samples, in theory you can: \n\n1) feed the images into a few conv3d layers with some pooling. The kernels of the conv layers should take into consideration that the z dimension is very low compared to x and y. \n2) Then flatten them. \n3) The flatten layer you can merge now with other input features.\n4) Finally output two values (one for the FVC value and one for the confidence level). \n\nYou can apply the competition evaluation as a loss function. So it is a single neural network from image to loss. That all being said, there is so few data that a linear regression almost scores as good as more complex networks." }, "source": "meta" }