![\"\"](\"https://github.com/alceubissoto/gan-skin-lesion/raw/images_examples/image_examples/semantic_ISIC_0000031_synthesized_image.jpg?raw=true\")
As seen , the dataset is highly imbalanced , and generating new data has always been one way to deal with it. Has anyone tried using GAN's , or some other approaches to generate more data? \nWould be great to have some insights on it. Thank you.
", "messages": [ { "id": "871795", "postDate": "06/02/2020 16:30:23", "content": "As seen , the dataset is highly imbalanced , and generating new data has always been one way to deal with it. Has anyone tried using GAN's , or some other approaches to generate more data? \nWould be great to have some insights on it. Thank you.
", "rawMarkdown": "As seen , the dataset is highly imbalanced , and generating new data has always been one way to deal with it. Has anyone tried using GAN's , or some other approaches to generate more data? \nWould be great to have some insights on it. Thank you.", "votes": null }, { "id": "871878", "postDate": "06/02/2020 17:43:55", "content": "there are many papers and opensource code on that. you may want to read about it and do experiments on it. e.g.\nhttps://github.com/alceubissoto/gan-skin-lesion
\n\nExample Synthetic Images\n
here is google experiments on it: https://ai.googleblog.com/2020/02/generating-diverse-synthetic-medical.html
\n\nhttps://arxiv.org/pdf/1911.08716.pdf
\n\n\"Figure 8: Augmenting training data with synthetic images. We added 20000 synthetic images to the original training data of size 49920. The overall performance is comparable to the baseline, but the performance on rare conditions like Melanoma and Basal cell carcinoma has noticeable improvement\"
", "rawMarkdown": "here is google experiments on it: https://ai.googleblog.com/2020/02/generating-diverse-synthetic-medical.html\n\nhttps://arxiv.org/pdf/1911.08716.pdf\n\n\"Figure 8: Augmenting training data with synthetic images. We added 20000 synthetic images to the original training data of size 49920. The overall performance is comparable to the baseline, but the performance on rare conditions like Melanoma and Basal cell carcinoma has noticeable improvement\"", "votes": null }, { "id": "871894", "postDate": "06/02/2020 17:59:24", "content": "Thanks for the leads. I'm also curious if anyone has done such experiments for this specific comp so far , since I don't have any GPU/resources I can use to work on / experiment with GAN's . I'll try to learn more about them.
", "rawMarkdown": "Thanks for the leads. I'm also curious if anyone has done such experiments for this specific comp so far , since I don't have any GPU/resources I can use to work on / experiment with GAN's . I'll try to learn more about them.", "votes": null } ], "comments": [ { "id": 871878, "author_name": "hengck23", "author_url": "", "post_date": "06/02/2020 17:43:55", "content": "there are many papers and opensource code on that. you may want to read about it and do experiments on it. e.g.\nhttps://github.com/alceubissoto/gan-skin-lesion
\n\nExample Synthetic Images\n
here is google experiments on it: https://ai.googleblog.com/2020/02/generating-diverse-synthetic-medical.html
\n\nhttps://arxiv.org/pdf/1911.08716.pdf
\n\n\"Figure 8: Augmenting training data with synthetic images. We added 20000 synthetic images to the original training data of size 49920. The overall performance is comparable to the baseline, but the performance on rare conditions like Melanoma and Basal cell carcinoma has noticeable improvement\"
", "votes": null, "replies": [] }, { "id": 871894, "author_name": "p4rallax", "author_url": "", "post_date": "06/02/2020 17:59:24", "content": "Thanks for the leads. I'm also curious if anyone has done such experiments for this specific comp so far , since I don't have any GPU/resources I can use to work on / experiment with GAN's . I'll try to learn more about them.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "871795": "As seen , the dataset is highly imbalanced , and generating new data has always been one way to deal with it. Has anyone tried using GAN's , or some other approaches to generate more data? \nWould be great to have some insights on it. Thank you.", "871878": "there are many papers and opensource code on that. you may want to read about it and do experiments on it. e.g.\nhttps://github.com/alceubissoto/gan-skin-lesion\n\nExample Synthetic Images\n", "871882": "here is google experiments on it: https://ai.googleblog.com/2020/02/generating-diverse-synthetic-medical.html\n\nhttps://arxiv.org/pdf/1911.08716.pdf\n\n\"Figure 8: Augmenting training data with synthetic images. We added 20000 synthetic images to the original training data of size 49920. The overall performance is comparable to the baseline, but the performance on rare conditions like Melanoma and Basal cell carcinoma has noticeable improvement\"", "871894": "Thanks for the leads. I'm also curious if anyone has done such experiments for this specific comp so far , since I don't have any GPU/resources I can use to work on / experiment with GAN's . I'll try to learn more about them." }, "source": "meta" }