I don't do much in the way of image processing, so i thought i'd ask you all. (someone here is gonna be in the know).
\n\nIt seems like this is kind of a no brainer for modern edge detection techniques. especially if you then feed that in to a neural net and let it chew on it for a while. (with the source image available too.) it should just build your model/prediction without much thought.
\n\nIs it really that straightforward a solution here or is there something tricky that makes this real hard? I don't want to tackle this and like 2 hours in someone, who has the tools just laying around, already has a 99.95% perfect nuclei detection up... So yeah, any opinions are welcome!
", "messages": [ { "id": "269483", "postDate": "01/16/2018 22:58:43", "content": "I don't do much in the way of image processing, so i thought i'd ask you all. (someone here is gonna be in the know).
\n\nIt seems like this is kind of a no brainer for modern edge detection techniques. especially if you then feed that in to a neural net and let it chew on it for a while. (with the source image available too.) it should just build your model/prediction without much thought.
\n\nIs it really that straightforward a solution here or is there something tricky that makes this real hard? I don't want to tackle this and like 2 hours in someone, who has the tools just laying around, already has a 99.95% perfect nuclei detection up... So yeah, any opinions are welcome!
", "rawMarkdown": "I don't do much in the way of image processing, so i thought i'd ask you all. (someone here is gonna be in the know).\n\nIt seems like this is kind of a no brainer for modern edge detection techniques. especially if you then feed that in to a neural net and let it chew on it for a while. (with the source image available too.) it should just build your model/prediction without much thought.\n\nIs it really that straightforward a solution here or is there something tricky that makes this real hard? I don't want to tackle this and like 2 hours in someone, who has the tools just laying around, already has a 99.95% perfect nuclei detection up... So yeah, any opinions are welcome!", "votes": null }, { "id": "269500", "postDate": "01/16/2018 23:38:49", "content": "I've been working on Computer Vision problems for a while and I doubt there is an off-the-shelf tool, which achieves 99.95% nucleus detection accuracy. Not even close.
", "rawMarkdown": "I've been working on Computer Vision problems for a while and I doubt there is an off-the-shelf tool, which achieves 99.95% nucleus detection accuracy. Not even close.", "votes": null }, { "id": "269634", "postDate": "01/17/2018 03:30:12", "content": "You can learn a lot from reading comments, kernels, and solutions from past segmentation competitions.
\n\nHere are a few:
\n\nIt seems like most of the time the winning models are some variation of a fully-convolutional encoder-decoder architecture. UNet is typically a strong place to start: https://arxiv.org/abs/1505.04597. Here is an implementation of UNet in python/keras for the ultrasound segmentation problem: https://github.com/jocicmarko/ultrasound-nerve-segmentation
\n\nOther models have been developed since UNet. One interesting example is Adversarial Segmentation Networks: https://arxiv.org/pdf/1611.08408.pdf. I took a pass at implementing this in Keras/Python, written up here: http://alex.klibisz.com/posts/2017-10-13-adversarial-segmentation-quick-first-pass/
\n\nI suspect top solutions will involve clever combinations of pre/post-processing and ensembles of encoder-decoder fully-convolutional networks. Still in my experience a rather simple UNet in Keras will get you pretty far.
", "rawMarkdown": "You can learn a lot from reading comments, kernels, and solutions from past segmentation competitions. \n\nHere are a few:\n\n 1. https://www.kaggle.com/c/carvana-image-masking-challenge\n 2. https://www.kaggle.com/c/dstl-satellite-imagery-feature-detection\n 3. https://www.kaggle.com/c/ultrasound-nerve-segmentation\n\nIt seems like most of the time the winning models are some variation of a fully-convolutional encoder-decoder architecture. UNet is typically a strong place to start: https://arxiv.org/abs/1505.04597. Here is an implementation of UNet in python/keras for the ultrasound segmentation problem: https://github.com/jocicmarko/ultrasound-nerve-segmentation\n\nOther models have been developed since UNet. One interesting example is Adversarial Segmentation Networks: https://arxiv.org/pdf/1611.08408.pdf. I took a pass at implementing this in Keras/Python, written up here: http://alex.klibisz.com/posts/2017-10-13-adversarial-segmentation-quick-first-pass/\n\nI suspect top solutions will involve clever combinations of pre/post-processing and ensembles of encoder-decoder fully-convolutional networks. Still in my experience a rather simple UNet in Keras will get you pretty far.", "votes": null }, { "id": "269861", "postDate": "01/17/2018 12:18:42", "content": "Agree! Besides, the problem here is generalizability across a variety of image/cell/microscope types. Solving any one image type, one might achieve high accuracy but the variation is what makes this challenge a challenge.
", "rawMarkdown": "Agree! Besides, the problem here is generalizability across a variety of image/cell/microscope types. Solving any one image type, one might achieve high accuracy but the variation is what makes this challenge a challenge.", "votes": null }, { "id": "277713", "postDate": "02/04/2018 01:56:35", "content": "I have a PhD in Materials Science and this very recent open access paper from Materials and Design journal, Elsevier may possibly help Dr Carpenter and you all competitors with generalising the latest digital image processing techniques / algorithms, which may be possibly transferred from the microstructural features of inorganic materials to biological samples: New methods for automatic quantification of microstructural features using digital image processing . Good luck.
", "rawMarkdown": "I have a PhD in Materials Science and this very recent open access paper from Materials and Design journal, Elsevier may possibly help Dr Carpenter and you all competitors with generalising the latest digital image processing techniques / algorithms, which may be possibly transferred from the microstructural features of inorganic materials to biological samples: [New methods for automatic quantification of microstructural features using digital image processing][1] . Good luck.\n\n\n [1]: https://doi.org/10.1016/j.matdes.2017.12.049", "votes": null }, { "id": "277714", "postDate": "02/04/2018 01:57:55", "content": ".
", "rawMarkdown": ".", "votes": null } ], "comments": [ { "id": 269500, "author_name": "pauljurczak", "author_url": "", "post_date": "01/16/2018 23:38:49", "content": "I've been working on Computer Vision problems for a while and I doubt there is an off-the-shelf tool, which achieves 99.95% nucleus detection accuracy. Not even close.
", "votes": null, "replies": [ { "id": 269861, "author_name": "drannecarpenter", "author_url": "", "post_date": "01/17/2018 12:18:42", "content": "Agree! Besides, the problem here is generalizability across a variety of image/cell/microscope types. Solving any one image type, one might achieve high accuracy but the variation is what makes this challenge a challenge.
", "votes": null, "replies": [] } ] }, { "id": 269634, "author_name": "alexklibisz", "author_url": "", "post_date": "01/17/2018 03:30:12", "content": "You can learn a lot from reading comments, kernels, and solutions from past segmentation competitions.
\n\nHere are a few:
\n\nIt seems like most of the time the winning models are some variation of a fully-convolutional encoder-decoder architecture. UNet is typically a strong place to start: https://arxiv.org/abs/1505.04597. Here is an implementation of UNet in python/keras for the ultrasound segmentation problem: https://github.com/jocicmarko/ultrasound-nerve-segmentation
\n\nOther models have been developed since UNet. One interesting example is Adversarial Segmentation Networks: https://arxiv.org/pdf/1611.08408.pdf. I took a pass at implementing this in Keras/Python, written up here: http://alex.klibisz.com/posts/2017-10-13-adversarial-segmentation-quick-first-pass/
\n\nI suspect top solutions will involve clever combinations of pre/post-processing and ensembles of encoder-decoder fully-convolutional networks. Still in my experience a rather simple UNet in Keras will get you pretty far.
", "votes": null, "replies": [] }, { "id": 277713, "author_name": "giusco", "author_url": "", "post_date": "02/04/2018 01:56:35", "content": "I have a PhD in Materials Science and this very recent open access paper from Materials and Design journal, Elsevier may possibly help Dr Carpenter and you all competitors with generalising the latest digital image processing techniques / algorithms, which may be possibly transferred from the microstructural features of inorganic materials to biological samples: New methods for automatic quantification of microstructural features using digital image processing . Good luck.
", "votes": null, "replies": [] }, { "id": 277714, "author_name": "giusco", "author_url": "", "post_date": "02/04/2018 01:57:55", "content": ".
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "269483": "I don't do much in the way of image processing, so i thought i'd ask you all. (someone here is gonna be in the know).\n\nIt seems like this is kind of a no brainer for modern edge detection techniques. especially if you then feed that in to a neural net and let it chew on it for a while. (with the source image available too.) it should just build your model/prediction without much thought.\n\nIs it really that straightforward a solution here or is there something tricky that makes this real hard? I don't want to tackle this and like 2 hours in someone, who has the tools just laying around, already has a 99.95% perfect nuclei detection up... So yeah, any opinions are welcome!", "269500": "I've been working on Computer Vision problems for a while and I doubt there is an off-the-shelf tool, which achieves 99.95% nucleus detection accuracy. Not even close.", "269634": "You can learn a lot from reading comments, kernels, and solutions from past segmentation competitions. \n\nHere are a few:\n\n 1. https://www.kaggle.com/c/carvana-image-masking-challenge\n 2. https://www.kaggle.com/c/dstl-satellite-imagery-feature-detection\n 3. https://www.kaggle.com/c/ultrasound-nerve-segmentation\n\nIt seems like most of the time the winning models are some variation of a fully-convolutional encoder-decoder architecture. UNet is typically a strong place to start: https://arxiv.org/abs/1505.04597. Here is an implementation of UNet in python/keras for the ultrasound segmentation problem: https://github.com/jocicmarko/ultrasound-nerve-segmentation\n\nOther models have been developed since UNet. One interesting example is Adversarial Segmentation Networks: https://arxiv.org/pdf/1611.08408.pdf. I took a pass at implementing this in Keras/Python, written up here: http://alex.klibisz.com/posts/2017-10-13-adversarial-segmentation-quick-first-pass/\n\nI suspect top solutions will involve clever combinations of pre/post-processing and ensembles of encoder-decoder fully-convolutional networks. Still in my experience a rather simple UNet in Keras will get you pretty far.", "269861": "Agree! Besides, the problem here is generalizability across a variety of image/cell/microscope types. Solving any one image type, one might achieve high accuracy but the variation is what makes this challenge a challenge.", "277713": "I have a PhD in Materials Science and this very recent open access paper from Materials and Design journal, Elsevier may possibly help Dr Carpenter and you all competitors with generalising the latest digital image processing techniques / algorithms, which may be possibly transferred from the microstructural features of inorganic materials to biological samples: [New methods for automatic quantification of microstructural features using digital image processing][1] . Good luck.\n\n\n [1]: https://doi.org/10.1016/j.matdes.2017.12.049", "277714": "." }, "source": "meta" }