Hi everyone,
\n\none of the creators of the HAM10000 dataset (= ISIC 2018 Task 3 Data) here. You might be interested to hear that I created lesion segmentations for all 10015 images of before-mentioned data. They are free to use for non-commercial purposes: \n- Outside kaggle: https://dataverse.harvard.edu/file.xhtml?persistentId=doi:10.7910/DVN/DBW86T/EGDUDF&version=2.0\n- Inside kaggle: https://www.kaggle.com/tschandl/ham10000-lesion-segmentations
\n\nIt was quite the work to produce them, so I wouldn't be unhappy if you acknowledge the publication I made them for :) --> https://www.nature.com/articles/s41591-020-0942-0
\n\nLooking forward to see whether this can help making techniques like CutMix work?
\n\nphil
", "messages": [ { "id": "913004", "postDate": "07/02/2020 22:04:33", "content": "Hi everyone,
\n\none of the creators of the HAM10000 dataset (= ISIC 2018 Task 3 Data) here. You might be interested to hear that I created lesion segmentations for all 10015 images of before-mentioned data. They are free to use for non-commercial purposes: \n- Outside kaggle: https://dataverse.harvard.edu/file.xhtml?persistentId=doi:10.7910/DVN/DBW86T/EGDUDF&version=2.0\n- Inside kaggle: https://www.kaggle.com/tschandl/ham10000-lesion-segmentations
\n\nIt was quite the work to produce them, so I wouldn't be unhappy if you acknowledge the publication I made them for :) --> https://www.nature.com/articles/s41591-020-0942-0
\n\nLooking forward to see whether this can help making techniques like CutMix work?
\n\nphil
", "rawMarkdown": "Hi everyone, \n\none of the creators of the [HAM10000 dataset](https://www.nature.com/articles/sdata2018161) (= ISIC 2018 Task 3 Data) here. You might be interested to hear that I created lesion segmentations for all 10015 images of before-mentioned data. They are free to use for non-commercial purposes: \n- Outside kaggle: https://dataverse.harvard.edu/file.xhtml?persistentId=doi:10.7910/DVN/DBW86T/EGDUDF&version=2.0\n- Inside kaggle: https://www.kaggle.com/tschandl/ham10000-lesion-segmentations\n\nIt was quite the work to produce them, so I wouldn't be unhappy if you acknowledge the publication I made them for :) --> https://www.nature.com/articles/s41591-020-0942-0\n\nLooking forward to see whether this can help making techniques like CutMix work?\n\nphil", "votes": null }, { "id": "915331", "postDate": "07/04/2020 16:43:13", "content": "Hi @tschandl , thanks a lot for sharing this!\nIt looks like a loooot of work!
\n\nI have a few questions : \n- How much time would you say it took you to go through those 10K images?\n- In the end, do you know what was the Intersection Over Union (or any other metric) between what the FCN model gave and your manually revised version?\n- Did you look for different softwares than FIJI to perform your manual annotation? Any tips on how to do this as quickly as possible?
\n\nThank you very much!
", "rawMarkdown": "Hi @tschandl , thanks a lot for sharing this!\nIt looks like a loooot of work!\n\nI have a few questions : \n- How much time would you say it took you to go through those 10K images?\n- In the end, do you know what was the Intersection Over Union (or any other metric) between what the FCN model gave and your manually revised version?\n- Did you look for different softwares than FIJI to perform your manual annotation? Any tips on how to do this as quickly as possible?\n\nThank you very much!", "votes": null }, { "id": "915983", "postDate": "07/05/2020 09:04:41", "content": "\n\n\n\n
\n- How much time would you say it took you to go through those 10K images?
\n
Net-time I can't estimate easily as I did it in numerous sessions whenever I had time to spare. I can tell you the timediff between first and last image-segmentation: 496 days.
\n\n\n\n\n\n
\n- In the end, do you know what was the Intersection Over Union (or any other metric) between what the FCN model gave and your manually revised version?
\n
The mean IoU I think was around ~.86, one would have to mention a few things though.\n- This is a biased metric here, as for a large part I corrected the preexisting automated masks rather than drawing from scratch. \n- There is no reliable \"true\" border but only an approximation from dermatoscopic images. Following, any human, and automated, segmentation will be biased on what one believes to be the true lesion border. Especially delimiting lesion area for groups BKL and AKIEC surrounded by sun-damaged skin are sometimes problematic where I often made a \"best guess\". \n- As a side note, even with less training data I guess most of today's segmentation networks show an IoU comparable to human raters among themselves already - for pigmented skin lesions. In fact we have shown it for one FCN here quite a while ago.
\n\n\n\n\n\n
\n- Did you look for different softwares than FIJI to perform your manual annotation? Any tips on how to do this as quickly as possible?
\n
I did, but I can hardly imagine anything being quicker for offline single-instance binary segmentation than what I ended up with: I made a FIJI-macro (triggered by one keystroke) to automatically do the following sequence, after I drawing the lesion outline: 1) convert selection outline to a binary segmentation mask - 2) store segmentation mask as \"verified\" in new folder - 3) close mask and image - 4) load the next image with an \"unverified\" segmentation mask - 5) load automated segmentation and convert to selection outline.
", "rawMarkdown": "> - How much time would you say it took you to go through those 10K images?\n\nNet-time I can't estimate easily as I did it in numerous sessions whenever I had time to spare. I can tell you the timediff between first and last image-segmentation: 496 days.\n\n\n> - In the end, do you know what was the Intersection Over Union (or any other metric) between what the FCN model gave and your manually revised version?\n\nThe mean IoU I think was around ~.86, one would have to mention a few things though.\n- This is a biased metric here, as for a large part I corrected the preexisting automated masks rather than drawing from scratch. \n- There is no reliable \"true\" border but only an approximation from dermatoscopic images. Following, any human, and automated, segmentation will be biased on what one believes to be the true lesion border. Especially delimiting lesion area for groups BKL and AKIEC surrounded by sun-damaged skin are sometimes problematic where I often made a \"best guess\". \n- As a side note, even with less training data I guess most of today's segmentation networks show an IoU comparable to human raters among themselves already - for pigmented skin lesions. In fact we have shown it for one FCN [here](https://doi.org/10.1016/j.compbiomed.2018.11.010) quite a while ago.\n\n\n> - Did you look for different softwares than FIJI to perform your manual annotation? Any tips on how to do this as quickly as possible?\n\nI did, but I can hardly imagine anything being quicker for offline single-instance binary segmentation than what I ended up with: I made a FIJI-macro (triggered by one keystroke) to automatically do the following sequence, after I drawing the lesion outline: 1) convert selection outline to a binary segmentation mask - 2) store segmentation mask as \"verified\" in new folder - 3) close mask and image - 4) load the next image with an \"unverified\" segmentation mask - 5) load automated segmentation and convert to selection outline.", "votes": null }, { "id": "916246", "postDate": "07/05/2020 14:00:55", "content": "Thank you for your detailed answer, really appreciate it.
\n\nCongratulations for this work!
", "rawMarkdown": "Thank you for your detailed answer, really appreciate it.\n\nCongratulations for this work!", "votes": null }, { "id": "916602", "postDate": "07/05/2020 20:34:48", "content": "Thanks :)
", "rawMarkdown": "Thanks :)", "votes": null } ], "comments": [ { "id": 915331, "author_name": "optimo", "author_url": "", "post_date": "07/04/2020 16:43:13", "content": "Hi @tschandl , thanks a lot for sharing this!\nIt looks like a loooot of work!
\n\nI have a few questions : \n- How much time would you say it took you to go through those 10K images?\n- In the end, do you know what was the Intersection Over Union (or any other metric) between what the FCN model gave and your manually revised version?\n- Did you look for different softwares than FIJI to perform your manual annotation? Any tips on how to do this as quickly as possible?
\n\nThank you very much!
", "votes": null, "replies": [ { "id": 915983, "author_name": "tschandl", "author_url": "", "post_date": "07/05/2020 09:04:41", "content": "\n\n\n\n
\n- How much time would you say it took you to go through those 10K images?
\n
Net-time I can't estimate easily as I did it in numerous sessions whenever I had time to spare. I can tell you the timediff between first and last image-segmentation: 496 days.
\n\n\n\n\n\n
\n- In the end, do you know what was the Intersection Over Union (or any other metric) between what the FCN model gave and your manually revised version?
\n
The mean IoU I think was around ~.86, one would have to mention a few things though.\n- This is a biased metric here, as for a large part I corrected the preexisting automated masks rather than drawing from scratch. \n- There is no reliable \"true\" border but only an approximation from dermatoscopic images. Following, any human, and automated, segmentation will be biased on what one believes to be the true lesion border. Especially delimiting lesion area for groups BKL and AKIEC surrounded by sun-damaged skin are sometimes problematic where I often made a \"best guess\". \n- As a side note, even with less training data I guess most of today's segmentation networks show an IoU comparable to human raters among themselves already - for pigmented skin lesions. In fact we have shown it for one FCN here quite a while ago.
\n\n\n\n\n\n
\n- Did you look for different softwares than FIJI to perform your manual annotation? Any tips on how to do this as quickly as possible?
\n
I did, but I can hardly imagine anything being quicker for offline single-instance binary segmentation than what I ended up with: I made a FIJI-macro (triggered by one keystroke) to automatically do the following sequence, after I drawing the lesion outline: 1) convert selection outline to a binary segmentation mask - 2) store segmentation mask as \"verified\" in new folder - 3) close mask and image - 4) load the next image with an \"unverified\" segmentation mask - 5) load automated segmentation and convert to selection outline.
", "votes": null, "replies": [] }, { "id": 916246, "author_name": "optimo", "author_url": "", "post_date": "07/05/2020 14:00:55", "content": "Thank you for your detailed answer, really appreciate it.
\n\nCongratulations for this work!
", "votes": null, "replies": [] }, { "id": 916602, "author_name": "tschandl", "author_url": "", "post_date": "07/05/2020 20:34:48", "content": "Thanks :)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "913004": "Hi everyone, \n\none of the creators of the [HAM10000 dataset](https://www.nature.com/articles/sdata2018161) (= ISIC 2018 Task 3 Data) here. You might be interested to hear that I created lesion segmentations for all 10015 images of before-mentioned data. They are free to use for non-commercial purposes: \n- Outside kaggle: https://dataverse.harvard.edu/file.xhtml?persistentId=doi:10.7910/DVN/DBW86T/EGDUDF&version=2.0\n- Inside kaggle: https://www.kaggle.com/tschandl/ham10000-lesion-segmentations\n\nIt was quite the work to produce them, so I wouldn't be unhappy if you acknowledge the publication I made them for :) --> https://www.nature.com/articles/s41591-020-0942-0\n\nLooking forward to see whether this can help making techniques like CutMix work?\n\nphil", "915331": "Hi @tschandl , thanks a lot for sharing this!\nIt looks like a loooot of work!\n\nI have a few questions : \n- How much time would you say it took you to go through those 10K images?\n- In the end, do you know what was the Intersection Over Union (or any other metric) between what the FCN model gave and your manually revised version?\n- Did you look for different softwares than FIJI to perform your manual annotation? Any tips on how to do this as quickly as possible?\n\nThank you very much!", "915983": "> - How much time would you say it took you to go through those 10K images?\n\nNet-time I can't estimate easily as I did it in numerous sessions whenever I had time to spare. I can tell you the timediff between first and last image-segmentation: 496 days.\n\n\n> - In the end, do you know what was the Intersection Over Union (or any other metric) between what the FCN model gave and your manually revised version?\n\nThe mean IoU I think was around ~.86, one would have to mention a few things though.\n- This is a biased metric here, as for a large part I corrected the preexisting automated masks rather than drawing from scratch. \n- There is no reliable \"true\" border but only an approximation from dermatoscopic images. Following, any human, and automated, segmentation will be biased on what one believes to be the true lesion border. Especially delimiting lesion area for groups BKL and AKIEC surrounded by sun-damaged skin are sometimes problematic where I often made a \"best guess\". \n- As a side note, even with less training data I guess most of today's segmentation networks show an IoU comparable to human raters among themselves already - for pigmented skin lesions. In fact we have shown it for one FCN [here](https://doi.org/10.1016/j.compbiomed.2018.11.010) quite a while ago.\n\n\n> - Did you look for different softwares than FIJI to perform your manual annotation? Any tips on how to do this as quickly as possible?\n\nI did, but I can hardly imagine anything being quicker for offline single-instance binary segmentation than what I ended up with: I made a FIJI-macro (triggered by one keystroke) to automatically do the following sequence, after I drawing the lesion outline: 1) convert selection outline to a binary segmentation mask - 2) store segmentation mask as \"verified\" in new folder - 3) close mask and image - 4) load the next image with an \"unverified\" segmentation mask - 5) load automated segmentation and convert to selection outline.", "916246": "Thank you for your detailed answer, really appreciate it.\n\nCongratulations for this work!", "916602": "Thanks :)" }, "source": "meta" }