I was looking into the data and the number of images we have in each folder. Later I found this is the following distribution of the number of images corresponding to each scanning method/class. this. I also read the following notebook from a previous competition and now I just can't get it as to how to use the 3-d factor of various scans. As you will see that the author of the shared notebook used 160 depth dimensions which were lower than all the images he had in the scan whereas in our case we have very few images in some scans and very large in some other scans. So how should we proceed with it? Also in some scans, there are no images at all.
", "messages": [ { "id": "1392629", "postDate": "07/18/2021 23:11:34", "content": "I was looking into the data and the number of images we have in each folder. Later I found this is the following distribution of the number of images corresponding to each scanning method/class. this. I also read the following notebook from a previous competition and now I just can't get it as to how to use the 3-d factor of various scans. As you will see that the author of the shared notebook used 160 depth dimensions which were lower than all the images he had in the scan whereas in our case we have very few images in some scans and very large in some other scans. So how should we proceed with it? Also in some scans, there are no images at all.
", "rawMarkdown": "I was looking into the data and the number of images we have in each folder. Later I found this is the following distribution of the number of images corresponding to each scanning method/class. this. I also read the following notebook from a previous competition and now I just can't get it as to how to use the 3-d factor of various scans. As you will see that the author of the shared notebook used 160 depth dimensions which were lower than all the images he had in the scan whereas in our case we have very few images in some scans and very large in some other scans. So how should we proceed with it? Also in some scans, there are no images at all.", "votes": null }, { "id": "1393729", "postDate": "07/19/2021 20:56:28", "content": "It's not proper to build 3D reconstructions from these images by stacking exported JPGs. They are already reconstructed images. Although possible, the results would be sub-optimal. Reconstructing should be done from the raw data (that we don't have).
\nIf you do want to recon these into 3D volumes though, you'll need to make use of a few DICOM tags to determine the orientation, position, thickness, gap and a few other things. However, the resulting images will lose diagnostic information.
", "rawMarkdown": "It's not proper to build 3D reconstructions from these images by stacking exported JPGs. They are already reconstructed images. Although possible, the results would be sub-optimal. Reconstructing should be done from the raw data (that we don't have). \n\nIf you do want to recon these into 3D volumes though, you'll need to make use of a few DICOM tags to determine the orientation, position, thickness, gap and a few other things. However, the resulting images will lose diagnostic information.", "votes": null }, { "id": "1394976", "postDate": "07/20/2021 19:04:53", "content": "Then how are we supposed to leverage the property of the data being in the form of 3D slices?
", "rawMarkdown": "Then how are we supposed to leverage the property of the data being in the form of 3D slices?", "votes": null }, { "id": "1394988", "postDate": "07/20/2021 19:24:31", "content": "Some of the series have enough images that reconstructing them is possible, but it's not as simple as layers slices on top of each other.
\nI'm good at radiography/radiology, but I'm not an expert in ML/AI yet .. so take my opinion with a grain of salt here. I don't think reconstructing these into 3D will offer any more information than can be extracted from the slices individually. It seems this approach would offer less diagnostic information than individual slices to a CNN.
\nIn general diagnostic radiology, doctors don't normally look at things in 3D .. there's too much hidden data that cannot be seen on 3D models.
\nHere's a semi-poor analogy ..
\nImagine if I performed an MR of a loaf of bread with a hot dog baked into the middle of it. If I sliced the images, I could see the boundaries of the hotdog inside .. If I look at a 3D view, I wouldn't even know there was a hotdog inside.
", "rawMarkdown": "Some of the series have enough images that reconstructing them is possible, but it's not as simple as layers slices on top of each other. \n\nI'm good at radiography/radiology, but I'm not an expert in ML/AI yet .. so take my opinion with a grain of salt here. I don't think reconstructing these into 3D will offer any more information than can be extracted from the slices individually. It seems this approach would offer less diagnostic information than individual slices to a CNN.\n\nIn general diagnostic radiology, doctors don't normally look at things in 3D .. there's too much hidden data that cannot be seen on 3D models.\n\nHere's a semi-poor analogy ..\n\nImagine if I performed an MR of a loaf of bread with a hot dog baked into the middle of it. If I sliced the images, I could see the boundaries of the hotdog inside .. If I look at a 3D view, I wouldn't even know there *was* a hotdog inside.", "votes": null }, { "id": "1395001", "postDate": "07/20/2021 19:45:09", "content": "Hi, can you give an example where a tumor is seen on a image? I imagine that just some of the slices are relevant.
\nHow do we have to choose the relevant slices
", "rawMarkdown": "Hi, can you give an example where a tumor is seen on a image? I imagine that just some of the slices are relevant.\n\nHow do we have to choose the relevant slices", "votes": null }, { "id": "1395011", "postDate": "07/20/2021 20:01:50", "content": "Take a look at the images in this notebook I just made on reference lines -> https://www.kaggle.com/davidbroberts/mr-reference-lines
\nYou can clearly see a tumor in the slices that capture it.
\nYou can eliminate images where the max pixel value = 0 (all black) .. or determine a mean pixel value threshold that includes enough non-black pixels to represent some anatomy.
\nOtherwise, to get just the slices with tumor tissue, I suppose you'd have to segment it. There is a brain tumor segmentation model floating around from a past competition.
", "rawMarkdown": "Take a look at the images in this notebook I just made on reference lines -> https://www.kaggle.com/davidbroberts/mr-reference-lines\n\nYou can clearly see a tumor in the slices that capture it.\n\nYou can eliminate images where the max pixel value = 0 (all black) .. or determine a mean pixel value threshold that includes enough non-black pixels to represent *some* anatomy.\n\nOtherwise, to get just the slices with tumor tissue, I suppose you'd have to segment it. There is a brain tumor segmentation model floating around from a past competition.", "votes": null }, { "id": "1395019", "postDate": "07/20/2021 20:08:17", "content": "I get a 404 error clicking on your link
", "rawMarkdown": "I get a 404 error clicking on your link", "votes": null }, { "id": "1395026", "postDate": "07/20/2021 20:19:47", "content": "Sorry, I forgot to make it public. Try now.
", "rawMarkdown": "Sorry, I forgot to make it public. Try now.", "votes": null }, { "id": "1395035", "postDate": "07/20/2021 20:27:17", "content": "Now its working, thanks
", "rawMarkdown": "Now its working, thanks", "votes": null } ], "comments": [ { "id": 1393729, "author_name": "davidbroberts", "author_url": "", "post_date": "07/19/2021 20:56:28", "content": "It's not proper to build 3D reconstructions from these images by stacking exported JPGs. They are already reconstructed images. Although possible, the results would be sub-optimal. Reconstructing should be done from the raw data (that we don't have).
\nIf you do want to recon these into 3D volumes though, you'll need to make use of a few DICOM tags to determine the orientation, position, thickness, gap and a few other things. However, the resulting images will lose diagnostic information.
", "votes": null, "replies": [ { "id": 1394976, "author_name": "abhishekprajapat", "author_url": "", "post_date": "07/20/2021 19:04:53", "content": "Then how are we supposed to leverage the property of the data being in the form of 3D slices?
", "votes": null, "replies": [] }, { "id": 1394988, "author_name": "davidbroberts", "author_url": "", "post_date": "07/20/2021 19:24:31", "content": "Some of the series have enough images that reconstructing them is possible, but it's not as simple as layers slices on top of each other.
\nI'm good at radiography/radiology, but I'm not an expert in ML/AI yet .. so take my opinion with a grain of salt here. I don't think reconstructing these into 3D will offer any more information than can be extracted from the slices individually. It seems this approach would offer less diagnostic information than individual slices to a CNN.
\nIn general diagnostic radiology, doctors don't normally look at things in 3D .. there's too much hidden data that cannot be seen on 3D models.
\nHere's a semi-poor analogy ..
\nImagine if I performed an MR of a loaf of bread with a hot dog baked into the middle of it. If I sliced the images, I could see the boundaries of the hotdog inside .. If I look at a 3D view, I wouldn't even know there was a hotdog inside.
", "votes": null, "replies": [] }, { "id": 1395001, "author_name": "lucamtb", "author_url": "", "post_date": "07/20/2021 19:45:09", "content": "Hi, can you give an example where a tumor is seen on a image? I imagine that just some of the slices are relevant.
\nHow do we have to choose the relevant slices
", "votes": null, "replies": [] }, { "id": 1395011, "author_name": "davidbroberts", "author_url": "", "post_date": "07/20/2021 20:01:50", "content": "Take a look at the images in this notebook I just made on reference lines -> https://www.kaggle.com/davidbroberts/mr-reference-lines
\nYou can clearly see a tumor in the slices that capture it.
\nYou can eliminate images where the max pixel value = 0 (all black) .. or determine a mean pixel value threshold that includes enough non-black pixels to represent some anatomy.
\nOtherwise, to get just the slices with tumor tissue, I suppose you'd have to segment it. There is a brain tumor segmentation model floating around from a past competition.
", "votes": null, "replies": [] }, { "id": 1395019, "author_name": "lucamtb", "author_url": "", "post_date": "07/20/2021 20:08:17", "content": "I get a 404 error clicking on your link
", "votes": null, "replies": [] }, { "id": 1395026, "author_name": "davidbroberts", "author_url": "", "post_date": "07/20/2021 20:19:47", "content": "Sorry, I forgot to make it public. Try now.
", "votes": null, "replies": [] }, { "id": 1395035, "author_name": "lucamtb", "author_url": "", "post_date": "07/20/2021 20:27:17", "content": "Now its working, thanks
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1392629": "I was looking into the data and the number of images we have in each folder. Later I found this is the following distribution of the number of images corresponding to each scanning method/class. this. I also read the following notebook from a previous competition and now I just can't get it as to how to use the 3-d factor of various scans. As you will see that the author of the shared notebook used 160 depth dimensions which were lower than all the images he had in the scan whereas in our case we have very few images in some scans and very large in some other scans. So how should we proceed with it? Also in some scans, there are no images at all.", "1393729": "It's not proper to build 3D reconstructions from these images by stacking exported JPGs. They are already reconstructed images. Although possible, the results would be sub-optimal. Reconstructing should be done from the raw data (that we don't have). \n\nIf you do want to recon these into 3D volumes though, you'll need to make use of a few DICOM tags to determine the orientation, position, thickness, gap and a few other things. However, the resulting images will lose diagnostic information.", "1394976": "Then how are we supposed to leverage the property of the data being in the form of 3D slices?", "1394988": "Some of the series have enough images that reconstructing them is possible, but it's not as simple as layers slices on top of each other. \n\nI'm good at radiography/radiology, but I'm not an expert in ML/AI yet .. so take my opinion with a grain of salt here. I don't think reconstructing these into 3D will offer any more information than can be extracted from the slices individually. It seems this approach would offer less diagnostic information than individual slices to a CNN.\n\nIn general diagnostic radiology, doctors don't normally look at things in 3D .. there's too much hidden data that cannot be seen on 3D models.\n\nHere's a semi-poor analogy ..\n\nImagine if I performed an MR of a loaf of bread with a hot dog baked into the middle of it. If I sliced the images, I could see the boundaries of the hotdog inside .. If I look at a 3D view, I wouldn't even know there *was* a hotdog inside.", "1395001": "Hi, can you give an example where a tumor is seen on a image? I imagine that just some of the slices are relevant.\n\nHow do we have to choose the relevant slices", "1395011": "Take a look at the images in this notebook I just made on reference lines -> https://www.kaggle.com/davidbroberts/mr-reference-lines\n\nYou can clearly see a tumor in the slices that capture it.\n\nYou can eliminate images where the max pixel value = 0 (all black) .. or determine a mean pixel value threshold that includes enough non-black pixels to represent *some* anatomy.\n\nOtherwise, to get just the slices with tumor tissue, I suppose you'd have to segment it. There is a brain tumor segmentation model floating around from a past competition.", "1395019": "I get a 404 error clicking on your link", "1395026": "Sorry, I forgot to make it public. Try now.", "1395035": "Now its working, thanks" }, "source": "meta" }![this](\"https://github.com/Abhishek-Prajapat/Data-Science-Projects-Case-Studies/blob/master/Images/new.png\"){kind=link}