![\"scatter\"](\"https://lh3.googleusercontent.com/proxy/LF03vqY3-Vg5DGlJ7TRZW5SVKAecVO7kLfdBGzei2KvnnNNtEBy55nWt2aev5_xwAd4aw5p0s2S-nNojEjaCinPgWpRkiqel3aU\")
Hi everyone,
\n\n\nTL;DR
\n
\n I describe here an idea to reduce blur on the x-ray images in order to get better segmentation & classification results.
After reading the discussion about the difficulty of the segmentation, it occured to me, that there is maybe a preprocessing step that could help to detect more lines at the end of the day.
\nThere is a phenomenon called radiation scattering.
\nThis is one of the factors that results in blurry areas on an x-ray instead of nice sharp edges!
To keep it short, the best is - for the sake of imaging - when the photons that are emitted from the x-ray source get through the body and reach the detector (or casette) - because these contribute to the image.
\nProblem is, that many of these photons changes it's path after interacting with the tissue and do not continue it's straight path to the casette, but takes a turn and ends up somewhere else on the casette, making pseudorandom noise on the image. One can reduce this effect with a so called anti-scatter grid, that is nothing else just thin lamellae that let the straight from source-to-casette flying photons through but not the ones that have a larger incline angle ( = scattered from the tissue - see image).
\nThis grid is missing by the images of this contest (for the sake of dosis minimalising).
What if we could make such an anti-scatter preprocessing filter? It had been done by many manufacturer already. This links to a white paper of a well known manufacturer and it's proprietary solution to the aforementioned problem - no advertisements here, i am in no way affiliated with any of them.
\nFigure below is from the whitepaper:
\n![\"figure\"](\"https://i.postimg.cc/8CmPZhYT/scatter.jpg\")
What if we would try to recreate a synthetic version of this? The key is to estimate the areas that are \"blurred\" due to scattering and this correlates highly with the mass the rays travel through, so with the non-attenuated areas of the image!
\n![\"scatter](\"https://i.postimg.cc/5tsMr1wR/scatter-reduction.jpg\")
The steps above (gaussian blurring of original image, multi-stage thresholdig and progressively blurring of the resulted areas) can be done fairly simple to generate the synthetic scatter image.
\nThe resultig image-pairs should be fed into one of the existing image deblurring models and at the end we get a blurry input --> non blurry output model.
\nIn other datasets and also the less- or non-scattered images from the present dataset are good candidates for this approach.
\nLet me know what you think!
", "messages": [ { "id": "1232050", "postDate": "03/09/2021 13:03:28", "content": "Hi everyone,
\n\n\nTL;DR
\n
\n I describe here an idea to reduce blur on the x-ray images in order to get better segmentation & classification results.
After reading the discussion about the difficulty of the segmentation, it occured to me, that there is maybe a preprocessing step that could help to detect more lines at the end of the day.
\nThere is a phenomenon called radiation scattering.
\nThis is one of the factors that results in blurry areas on an x-ray instead of nice sharp edges!
To keep it short, the best is - for the sake of imaging - when the photons that are emitted from the x-ray source get through the body and reach the detector (or casette) - because these contribute to the image.
\nProblem is, that many of these photons changes it's path after interacting with the tissue and do not continue it's straight path to the casette, but takes a turn and ends up somewhere else on the casette, making pseudorandom noise on the image. One can reduce this effect with a so called anti-scatter grid, that is nothing else just thin lamellae that let the straight from source-to-casette flying photons through but not the ones that have a larger incline angle ( = scattered from the tissue - see image).
\nThis grid is missing by the images of this contest (for the sake of dosis minimalising).
What if we could make such an anti-scatter preprocessing filter? It had been done by many manufacturer already. This links to a white paper of a well known manufacturer and it's proprietary solution to the aforementioned problem - no advertisements here, i am in no way affiliated with any of them.
\nFigure below is from the whitepaper:
\n
What if we would try to recreate a synthetic version of this? The key is to estimate the areas that are \"blurred\" due to scattering and this correlates highly with the mass the rays travel through, so with the non-attenuated areas of the image!
\n
The steps above (gaussian blurring of original image, multi-stage thresholdig and progressively blurring of the resulted areas) can be done fairly simple to generate the synthetic scatter image.
\nThe resultig image-pairs should be fed into one of the existing image deblurring models and at the end we get a blurry input --> non blurry output model.
\nIn other datasets and also the less- or non-scattered images from the present dataset are good candidates for this approach.
\nLet me know what you think!
", "rawMarkdown": "Hi everyone,\n\n> TL;DR\nI describe here an idea to reduce blur on the x-ray images in order to get better segmentation & classification results.\n\nAfter reading the [discussion ](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/223984) about the difficulty of the segmentation, it occured to me, that there is maybe a preprocessing step that could help to detect more lines at the end of the day.\n\nThere is a phenomenon called **radiation scattering**. \nThis is one of the factors that **results in blurry areas on an x-ray** instead of nice sharp edges!\n\nTo keep it short, the best is - for the sake of imaging - when the photons that are emitted from the x-ray source get through the body and reach the detector (or casette) - because these contribute to the image.\n\nProblem is, that many of these photons changes it's path after interacting with the tissue and do not continue it's straight path to the casette, but takes a turn and ends up somewhere else on the casette, making **pseudorandom noise on the image**. One can reduce this effect with a so called anti-scatter grid, that is nothing else just thin lamellae that let the straight from source-to-casette flying photons through but not the ones that have a larger incline angle ( = scattered from the tissue - see image). \n**This grid is missing by the images of this contest** (for the sake of dosis minimalising).\n\n\n\nWhat if we could make such an anti-scatter preprocessing filter? It had been done by many manufacturer already. [This links ](https://philipsproductcontent.blob.core.windows.net/assets/20170523/16bb7fdd54a24d369a89a77c015ccedd.pdf) to a white paper of a well known manufacturer and it's proprietary solution to the aforementioned problem - no advertisements here, i am in no way affiliated with any of them.\n\nFigure below is from the whitepaper:\n\n\nWhat if we would try to recreate a synthetic version of this? The key is to estimate the areas that are \"blurred\" due to scattering and this correlates highly with the mass the rays travel through, so with the non-attenuated areas of the image!\n\n\n\nThe steps above (gaussian blurring of original image, multi-stage thresholdig and progressively blurring of the resulted areas) can be done fairly simple to **generate the synthetic scatter image**.\n\nThe resultig image-pairs should be fed into one of the [existing image deblurring models](https://github.com/subeeshvasu/Awesome-Deblurring) and at the end we get a blurry input --> non blurry output model.\n\nIn other datasets and also the less- or non-scattered images from the present dataset are good candidates for this approach.\n\nLet me know what you think!", "votes": null }, { "id": "1232484", "postDate": "03/09/2021 19:42:42", "content": "Very interesting idea! Do you know what process the whitepaper you mentioned used to perform the ‘scatter estimation’ step?
", "rawMarkdown": "Very interesting idea! Do you know what process the whitepaper you mentioned used to perform the ‘scatter estimation’ step?", "votes": null }, { "id": "1233337", "postDate": "03/10/2021 10:11:25", "content": "@reubenschmidt,
\nthey used real life measurement data.
Very interesting idea! Do you know what process the whitepaper you mentioned used to perform the ‘scatter estimation’ step?
", "votes": null, "replies": [ { "id": 1233337, "author_name": "sandorkonya", "author_url": "", "post_date": "03/10/2021 10:11:25", "content": "@reubenschmidt,
\nthey used real life measurement data.