![\"predicted](\"https://i.postimg.cc/4xZZQJYS/trachea-bifurcation.jpg\")
Hi there,
\ni'm a radiologist and I hope some of you find the dataset i made useful:
\ni manually annotated 5281 trachea bifurcation on x-rays of the dataset of the current challenge.
You can find them here:
\nhttps://www.kaggle.com/sandorkonya/5k-trachea-bifurcation-on-chest-xray
A faster RCNN (200 px bounding box around the point) trained on it performs pretty good, the average distance to GT is below 50 px, see histogram (X distance in pixel from GT) - thanks for @sainatarajan7 with the help on this:
\n
In comparison, the model trained on the images from @raddar , to be found here vs my hand annotated GT:
\n![\"predicted](\"https://i.postimg.cc/zvV35qCj/trachea-bifurcation-raddar.jpg\")
Hi there,
\ni'm a radiologist and I hope some of you find the dataset i made useful:
\ni manually annotated 5281 trachea bifurcation on x-rays of the dataset of the current challenge.
You can find them here:
\nhttps://www.kaggle.com/sandorkonya/5k-trachea-bifurcation-on-chest-xray
A faster RCNN (200 px bounding box around the point) trained on it performs pretty good, the average distance to GT is below 50 px, see histogram (X distance in pixel from GT) - thanks for @sainatarajan7 with the help on this:
\n
In comparison, the model trained on the images from @raddar , to be found here vs my hand annotated GT:
\n
This is tremendous work… did you use DICOM viewer for that? - im interested in your tools:) some images just have bad windowing to properly see something
", "rawMarkdown": "This is tremendous work... did you use DICOM viewer for that? - im interested in your tools:) some images just have bad windowing to properly see something", "votes": null }, { "id": "1212182", "postDate": "02/21/2021 01:15:30", "content": "i am using just GIMP to mark a circle for the endpoint
", "rawMarkdown": "i am using just GIMP to mark a circle for the endpoint", "votes": null }, { "id": "1212650", "postDate": "02/21/2021 12:49:27", "content": "Hi there,
\ntook like 3-4 nights…
\ni use VGG for such easy tasks.
\nYes, i think from this 5k there were like 20-30 where i needed windowing.
\nI programmed my own browser based tool, (with the help of @sainatarajan7 ) a Cornerstone & CornerstoneTools based image annotation tool running on Flask. It has all basic functions of a dicom viewer, it enables to primary annotate images, but also for the correction of the segment proposals from real time (in background) inferenced images.
", "rawMarkdown": "Hi there,\n\ntook like 3-4 nights... \n\ni use [VGG](https://www.robots.ox.ac.uk/~vgg/software/via/) for such easy tasks.\n\nYes, i think from this 5k there were like 20-30 where i needed windowing.\n\nI programmed my own browser based tool, (with the help of @sainatarajan7 ) a Cornerstone & [CornerstoneTools](https://tools.cornerstonejs.org/examples/tools/wwwc.html) based image annotation tool running on Flask. It has all basic functions of a dicom viewer, it enables to primary annotate images, but also for the correction of the segment proposals from real time (in background) inferenced images.", "votes": null }, { "id": "1212652", "postDate": "02/21/2021 12:50:45", "content": "Try VGG, it is standalone one page html and produces json or csv.
\nI use it extensively for various things!
I had been using https://github.com/heartexlabs/label-studio which seem to be similar to VGG. label-studio has very nice backend and good multi-user management. Thats why I like it more.
", "rawMarkdown": "I had been using https://github.com/heartexlabs/label-studio which seem to be similar to VGG. label-studio has very nice backend and good multi-user management. Thats why I like it more.", "votes": null }, { "id": "1212733", "postDate": "02/21/2021 14:26:25", "content": "Nice, did not know this!
\nWith the cornerstone integration it would be bombastic =)
Hey! Sorry for the stupid question. How will this data help in competition?
", "rawMarkdown": "Hey! Sorry for the stupid question. How will this data help in competition?", "votes": null }, { "id": "1213028", "postDate": "02/21/2021 19:27:45", "content": "\ntake a look at the picture below:
\n![\"trachea](\"https://i.postimg.cc/gjLbyy1V/trachea-bifurcation.jpg\")
If you can determine the trachea bifurcation on the image and the tip of the ETT, you can easily conclude whether the ETT is abnormal or not.
\nIf the ETT-Tip is near (but above) the bifurcation, it can be borderline. If the ETT is well above the bifurcation, it is normal - how many pixels?
\n You can simply make a mean of distances of ETT-Tip -> Trachea bifurcation in each of the categories (ETT Normal, Borderline, Abnormal) and take the mean for each class as a starter…
BTW: i would love to see a notebook that shows these above - anyone? @hengck23 , @raddar , @antonyo314 ? =)
\nCAVE: there are other signs to that define the abnormal position of the ETT (possible contralateral opacity, mediastinal shift for example) but i think with the very low incidence of abnormally placed ETT this may be one good way to select those cases… the other way could might be synthetic data like this:
\n![\"synthetic](\"https://i.postimg.cc/pXpzM9Mg/synthetic-ETT.jpg\")
\n(open in new tab to see the details)
I wrote a script to render random syntetic ETT and place along a curved path (annotated few hundred input images manually) and then add these images on the input image… within the coming days i am going to upload this dataset to.
", "rawMarkdown": "antonyo314 ,\n\ntake a look at the picture below:\n\n\nIf you can determine the trachea bifurcation on the image and the tip of the ETT, you can easily conclude whether the ETT is abnormal or not. \n\nIf the ETT-Tip is near (but above) the bifurcation, it can be borderline. If the ETT is well above the bifurcation, it is normal - how many pixels?\n You can simply make a mean of distances of ETT-Tip -> Trachea bifurcation in each of the categories (ETT Normal, Borderline, Abnormal) and take the mean for each class as a starter...\n\nBTW: i would love to see a notebook that shows these above - anyone? @hengck23 , @raddar , @antonyo314 ? =)\n\nCAVE: there are other signs to that define the abnormal position of the ETT (possible contralateral opacity, mediastinal shift for example) but i think with the very low incidence of abnormally placed ETT this may be one good way to select those cases... the other way could might be synthetic data like this:\n\n \n(open in new tab to see the details)\n\nI wrote a script to render random syntetic ETT and place along a curved path (annotated few hundred input images manually) and then add these images on the input image... within the coming days i am going to upload this dataset to.", "votes": null }, { "id": "1214915", "postDate": "02/23/2021 07:49:37", "content": "hi,
\nHow can I have this synthetically created supplement dataset.
\nThanks
I have added a notebook for this dataset -
\nhttps://www.kaggle.com/something4kag/ranzcr-tracheal-bifurcations-datasets-viz-5k-csv
It outputs a csv from the json and includes a column for raddar's tracheal bifurcations as well.
\nThe notebook has some visualisations for ETT showing the train annotation and the tracheal bifurcations for both 5k and raddar's prediction for comparison.
Is there a reason why these 5k were selected?
", "rawMarkdown": "I have added a notebook for this dataset -\nhttps://www.kaggle.com/something4kag/ranzcr-tracheal-bifurcations-datasets-viz-5k-csv\n\nIt outputs a csv from the json and includes a column for raddar's tracheal bifurcations as well.\nThe notebook has some visualisations for ETT showing the train annotation and the tracheal bifurcations for both 5k and raddar's prediction for comparison. \n\nIs there a reason why these 5k were selected?", "votes": null }, { "id": "1219891", "postDate": "02/27/2021 10:21:54", "content": "@something4kag,
\nNice notebook, already commented there!
\nI think these were the first 5k images in the train folder…
This is tremendous work… did you use DICOM viewer for that? - im interested in your tools:) some images just have bad windowing to properly see something
", "votes": null, "replies": [ { "id": 1212650, "author_name": "sandorkonya", "author_url": "", "post_date": "02/21/2021 12:49:27", "content": "Hi there,
\ntook like 3-4 nights…
\ni use VGG for such easy tasks.
\nYes, i think from this 5k there were like 20-30 where i needed windowing.
\nI programmed my own browser based tool, (with the help of @sainatarajan7 ) a Cornerstone & CornerstoneTools based image annotation tool running on Flask. It has all basic functions of a dicom viewer, it enables to primary annotate images, but also for the correction of the segment proposals from real time (in background) inferenced images.
", "votes": null, "replies": [] }, { "id": 1212712, "author_name": "raddar", "author_url": "", "post_date": "02/21/2021 14:01:42", "content": "I had been using https://github.com/heartexlabs/label-studio which seem to be similar to VGG. label-studio has very nice backend and good multi-user management. Thats why I like it more.
", "votes": null, "replies": [] }, { "id": 1212733, "author_name": "sandorkonya", "author_url": "", "post_date": "02/21/2021 14:26:25", "content": "Nice, did not know this!
\nWith the cornerstone integration it would be bombastic =)
i am using just GIMP to mark a circle for the endpoint
", "votes": null, "replies": [ { "id": 1212652, "author_name": "sandorkonya", "author_url": "", "post_date": "02/21/2021 12:50:45", "content": "Try VGG, it is standalone one page html and produces json or csv.
\nI use it extensively for various things!
Hey! Sorry for the stupid question. How will this data help in competition?
", "votes": null, "replies": [ { "id": 1213028, "author_name": "sandorkonya", "author_url": "", "post_date": "02/21/2021 19:27:45", "content": "\ntake a look at the picture below:
\n
If you can determine the trachea bifurcation on the image and the tip of the ETT, you can easily conclude whether the ETT is abnormal or not.
\nIf the ETT-Tip is near (but above) the bifurcation, it can be borderline. If the ETT is well above the bifurcation, it is normal - how many pixels?
\n You can simply make a mean of distances of ETT-Tip -> Trachea bifurcation in each of the categories (ETT Normal, Borderline, Abnormal) and take the mean for each class as a starter…
BTW: i would love to see a notebook that shows these above - anyone? @hengck23 , @raddar , @antonyo314 ? =)
\nCAVE: there are other signs to that define the abnormal position of the ETT (possible contralateral opacity, mediastinal shift for example) but i think with the very low incidence of abnormally placed ETT this may be one good way to select those cases… the other way could might be synthetic data like this:
\n
\n(open in new tab to see the details)
I wrote a script to render random syntetic ETT and place along a curved path (annotated few hundred input images manually) and then add these images on the input image… within the coming days i am going to upload this dataset to.
", "votes": null, "replies": [] } ] }, { "id": 1214915, "author_name": "sapthrishi007", "author_url": "", "post_date": "02/23/2021 07:49:37", "content": "hi,
\nHow can I have this synthetically created supplement dataset.
\nThanks
I have added a notebook for this dataset -
\nhttps://www.kaggle.com/something4kag/ranzcr-tracheal-bifurcations-datasets-viz-5k-csv
It outputs a csv from the json and includes a column for raddar's tracheal bifurcations as well.
\nThe notebook has some visualisations for ETT showing the train annotation and the tracheal bifurcations for both 5k and raddar's prediction for comparison.
Is there a reason why these 5k were selected?
", "votes": null, "replies": [ { "id": 1219891, "author_name": "sandorkonya", "author_url": "", "post_date": "02/27/2021 10:21:54", "content": "@something4kag,
\nNice notebook, already commented there!
\nI think these were the first 5k images in the train folder…