Is anyone using rib suppression techniques?
\nI made this notebook showing how to do it with a couple masks.
\nThis is a quick project and the results are nowhere near perfect, but I thought someone here might find it useful.
\nhttps://www.kaggle.com/davidbroberts/rib-suppression-poc
", "messages": [ { "id": "1341820", "postDate": "06/09/2021 03:10:01", "content": "Is anyone using rib suppression techniques?
\nI made this notebook showing how to do it with a couple masks.
\nThis is a quick project and the results are nowhere near perfect, but I thought someone here might find it useful.
\nhttps://www.kaggle.com/davidbroberts/rib-suppression-poc
", "rawMarkdown": "Is anyone using rib suppression techniques? \n\nI made this notebook showing how to do it with a couple masks.\n\nThis is a quick project and the results are nowhere near perfect, but I thought someone here might find it useful.\n\nhttps://www.kaggle.com/davidbroberts/rib-suppression-poc", "votes": null }, { "id": "1342940", "postDate": "06/09/2021 21:08:58", "content": "I have done some work in this area in the past - all done with \"synthetic\" X-rays:
\nhttps://www.kaggle.com/raddar/bone-drr-unet
", "rawMarkdown": "I have done some work in this area in the past - all done with \"synthetic\" X-rays:\n\nhttps://www.kaggle.com/raddar/bone-drr-unet", "votes": null }, { "id": "1342967", "postDate": "06/09/2021 22:22:57", "content": "Interesting work @raddar
\nI'm not sure I understand exactly what's going on though.
\nI have done some multiplanar reconstruction from CT's and fiddled with convolution filters and segmentation. Big modality vendors are starting to include it directly on the devices. I don't see much in the way of x-ray segmentation yet.
", "rawMarkdown": "Interesting work @raddar \n\nI'm not sure I understand exactly what's going on though.\n\nI have done some multiplanar reconstruction from CT's and fiddled with convolution filters and segmentation. Big modality vendors are starting to include it directly on the devices. I don't see much in the way of x-ray segmentation yet.", "votes": null }, { "id": "1343221", "postDate": "06/10/2021 05:45:35", "content": "Tanks @raddar, this is useful for pre-processing images before inserting into ML. It greatly improved my training
", "rawMarkdown": "Tanks @raddar, this is useful for pre-processing images before inserting into ML. It greatly improved my training", "votes": null }, { "id": "1343261", "postDate": "06/10/2021 06:17:20", "content": "@David. bones have different density and thus different pixel values in CT. So the idea is to filter these bone tissue related pixels in 3D space, and mimic by some model how X-ray device would produce X-ray film based on that 3D space (imagine a skeleton standing while X-ray is performed). Thus you get bone X-ray. Similarly you can repeat the same idea with bone pixels removed from 3D space - in the end you get 2 different images - bone layer X-ray and bone subtracted X-ray.
", "rawMarkdown": "David. bones have different density and thus different pixel values in CT. So the idea is to filter these bone tissue related pixels in 3D space, and mimic by some model how X-ray device would produce X-ray film based on that 3D space (imagine a skeleton standing while X-ray is performed). Thus you get bone X-ray. Similarly you can repeat the same idea with bone pixels removed from 3D space - in the end you get 2 different images - bone layer X-ray and bone subtracted X-ray.", "votes": null }, { "id": "1343271", "postDate": "06/10/2021 06:21:29", "content": "idea how to extract bone pixels in general:
\nhttps://www.kaggle.com/gzuidhof/full-preprocessing-tutorial
Interesting @raddar. Thanks for sharing. This looks like MIPing to me (maximum intensity projection).
", "rawMarkdown": "Interesting @raddar. Thanks for sharing. This looks like MIPing to me (maximum intensity projection).", "votes": null }, { "id": "1344288", "postDate": "06/10/2021 19:16:14", "content": "Compared to CT, a challenge from general X-ray image is the great variation of the absolution intensity due to different device & post processing. This makes it hard to normalize to the standard tissue HU window. You may find intercept (background air) varies in the training set from -1000 to -2000, and it is harder when image was taken with little air in the background. Foreign object, e.g., metal and digital label, on the other hand, could be very bright. Heterogenicity is the biggest hurdle.
\nThis is a pre/post processing & augmentation competition. Radiologists knows what they are looking for. Attention is all they need.
", "rawMarkdown": "Compared to CT, a challenge from general X-ray image is the great variation of the absolution intensity due to different device & post processing. This makes it hard to normalize to the standard tissue HU window. You may find intercept (background air) varies in the training set from -1000 to -2000, and it is harder when image was taken with little air in the background. Foreign object, e.g., metal and digital label, on the other hand, could be very bright. Heterogenicity is the biggest hurdle.\n\nThis is a pre/post processing & augmentation competition. Radiologists knows what they are looking for. *Attention is all they need*.", "votes": null } ], "comments": [ { "id": 1342940, "author_name": "raddar", "author_url": "", "post_date": "06/09/2021 21:08:58", "content": "I have done some work in this area in the past - all done with \"synthetic\" X-rays:
\nhttps://www.kaggle.com/raddar/bone-drr-unet
", "votes": null, "replies": [ { "id": 1342967, "author_name": "davidbroberts", "author_url": "", "post_date": "06/09/2021 22:22:57", "content": "Interesting work @raddar
\nI'm not sure I understand exactly what's going on though.
\nI have done some multiplanar reconstruction from CT's and fiddled with convolution filters and segmentation. Big modality vendors are starting to include it directly on the devices. I don't see much in the way of x-ray segmentation yet.
", "votes": null, "replies": [] }, { "id": 1343221, "author_name": "lucasvianna360", "author_url": "", "post_date": "06/10/2021 05:45:35", "content": "Tanks @raddar, this is useful for pre-processing images before inserting into ML. It greatly improved my training
", "votes": null, "replies": [] }, { "id": 1343261, "author_name": "raddar", "author_url": "", "post_date": "06/10/2021 06:17:20", "content": "@David. bones have different density and thus different pixel values in CT. So the idea is to filter these bone tissue related pixels in 3D space, and mimic by some model how X-ray device would produce X-ray film based on that 3D space (imagine a skeleton standing while X-ray is performed). Thus you get bone X-ray. Similarly you can repeat the same idea with bone pixels removed from 3D space - in the end you get 2 different images - bone layer X-ray and bone subtracted X-ray.
", "votes": null, "replies": [] }, { "id": 1343271, "author_name": "raddar", "author_url": "", "post_date": "06/10/2021 06:21:29", "content": "idea how to extract bone pixels in general:
\nhttps://www.kaggle.com/gzuidhof/full-preprocessing-tutorial
Interesting @raddar. Thanks for sharing. This looks like MIPing to me (maximum intensity projection).
", "votes": null, "replies": [] }, { "id": 1344288, "author_name": "houndcl", "author_url": "", "post_date": "06/10/2021 19:16:14", "content": "Compared to CT, a challenge from general X-ray image is the great variation of the absolution intensity due to different device & post processing. This makes it hard to normalize to the standard tissue HU window. You may find intercept (background air) varies in the training set from -1000 to -2000, and it is harder when image was taken with little air in the background. Foreign object, e.g., metal and digital label, on the other hand, could be very bright. Heterogenicity is the biggest hurdle.
\nThis is a pre/post processing & augmentation competition. Radiologists knows what they are looking for. Attention is all they need.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1341820": "Is anyone using rib suppression techniques? \n\nI made this notebook showing how to do it with a couple masks.\n\nThis is a quick project and the results are nowhere near perfect, but I thought someone here might find it useful.\n\nhttps://www.kaggle.com/davidbroberts/rib-suppression-poc", "1342940": "I have done some work in this area in the past - all done with \"synthetic\" X-rays:\n\nhttps://www.kaggle.com/raddar/bone-drr-unet", "1342967": "Interesting work @raddar \n\nI'm not sure I understand exactly what's going on though.\n\nI have done some multiplanar reconstruction from CT's and fiddled with convolution filters and segmentation. Big modality vendors are starting to include it directly on the devices. I don't see much in the way of x-ray segmentation yet.", "1343221": "Tanks @raddar, this is useful for pre-processing images before inserting into ML. It greatly improved my training", "1343261": "David. bones have different density and thus different pixel values in CT. So the idea is to filter these bone tissue related pixels in 3D space, and mimic by some model how X-ray device would produce X-ray film based on that 3D space (imagine a skeleton standing while X-ray is performed). Thus you get bone X-ray. Similarly you can repeat the same idea with bone pixels removed from 3D space - in the end you get 2 different images - bone layer X-ray and bone subtracted X-ray.", "1343271": "idea how to extract bone pixels in general: \nhttps://www.kaggle.com/gzuidhof/full-preprocessing-tutorial", "1343779": "Interesting @raddar. Thanks for sharing. This looks like MIPing to me (maximum intensity projection).", "1344288": "Compared to CT, a challenge from general X-ray image is the great variation of the absolution intensity due to different device & post processing. This makes it hard to normalize to the standard tissue HU window. You may find intercept (background air) varies in the training set from -1000 to -2000, and it is harder when image was taken with little air in the background. Foreign object, e.g., metal and digital label, on the other hand, could be very bright. Heterogenicity is the biggest hurdle.\n\nThis is a pre/post processing & augmentation competition. Radiologists knows what they are looking for. *Attention is all they need*." }, "source": "meta" }