{"id":7466,"date":"2022-09-10T22:00:00","date_gmt":"2022-09-10T14:00:00","guid":{"rendered":"http:\/\/139.9.1.231\/?p=7466"},"modified":"2022-09-07T22:01:31","modified_gmt":"2022-09-07T14:01:31","slug":"3d-u-net","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/09\/10\/3d-u-net\/","title":{"rendered":"3D U-Net"},"content":{"rendered":"\n<p class=\"has-light-pink-background-color has-background\"><strong><em>\u8bba\u6587\uff1a<a rel=\"noreferrer noopener\" href=\"https:\/\/arxiv.org\/abs\/1606.06650\" target=\"_blank\">3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation<\/a><\/em><\/strong><\/p>\n\n\n\n<p class=\"has-light-pink-background-color has-background\"><strong><em>github: <a href=\"https:\/\/github.com\/wolny\/pytorch-3dunet\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/wolny\/pytorch-3dunet<\/a><\/em><\/strong><\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"695\" height=\"362\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-63.png\" alt=\"\" class=\"wp-image-7470\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-63.png 695w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-63-300x156.png 300w\" sizes=\"(max-width: 695px) 100vw, 695px\" \/><\/figure>\n\n\n\n<p>\u8bba\u6587\u6700\u65e9\u7248\u672carXiv\u4e0a\u7684\u53d1\u8868\u65f6\u95f4\u662f2016.06\uff0c\u672c\u6587\u662f\u8bba\u6587v1\u7248\u672c\u7b14\u8bb0  MICCAI 2016\u6536\u5f55<\/p>\n\n\n\n<p>    \u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u4ece\u7a00\u758f\u6ce8\u91ca\u7684\u7acb\u4f53\u6570\u636e\u4e2d\u5b66\u4e60\u4e09\u7ef4\u5206\u5272\u7684\u7f51\u7edc\u30023D U-Net\u8fd9\u7bc7\u8bba\u6587\u7684\u8bde\u751f\u4e3b\u8981\u662f\u4e3a\u4e86\u5904\u7406\u4e00\u4e9b\u5757\u72b6\u56fe\uff08volumetric images\uff09\uff0c\u57fa\u672c\u7684\u539f\u7406\u8ddfU-Net\u5176\u5b9e\u5e76\u65e0\u5927\u5dee\uff0c\u56e0\u4e3a3D U-Net\u5c31\u662f\u75283D\u5377\u79ef\u64cd\u4f5c\u66ff\u6362\u4e862D\u7684<\/p>\n\n\n\n<p><strong>3D\u6570\u636e\u5bf9\u4e8e\u751f\u7269\u533b\u5b66\u6570\u636e\u5206\u6790\u6765\u8bf4\u663e\u5f97\u975e\u5e38\u5197\u4f59<\/strong><\/p>\n\n\n\n<ul><li>\u5728\u4e09\u7ef4\u5c42\u9762\u4e0a\u6807\u6ce8\u5206\u5272label\u6bd4\u8f83\u56f0\u96be\uff0c\u56e0\u4e3a\u7535\u8111\u5c4f\u5e55\u4e0a\u53ea\u80fd\u5c55\u793a2D\u7684\u5207\u7247<\/li><li>\u540c\u65f6\uff0c\u9010\u5c42\u6807\u6ce8\u5927\u91cf\u7684\u5207\u7247\u53c8\u5f88\u7e41\u7410\uff0c\u4e14\u76f8\u90bb\u5c42\u7684\u4fe1\u606f\u51e0\u4e4e\u662f\u76f8\u540c\u7684<\/li><li>\u56e0\u6b64\uff0c\u5b8c\u6574\u6ce8\u91ca3D\u6570\u636e\u5e76\u4e0d\u662f\u521b\u5efa\u5927\u800c\u4e30\u5bcc\u7684\u8bad\u7ec3\u6570\u636e\u96c6\u7684\u6709\u6548\u65b9\u6cd5\uff0c\u5c24\u5176\u662f\u5bf9\u4e8e\u9700\u8981\u5927\u91cf\u6807\u7b7e\u6570\u636e\u7684\u5b66\u4e60\u7c7b\u7b97\u6cd5<\/li><\/ul>\n\n\n\n<p class=\"has-light-blue-background-color has-background\">    \u751f\u7269\u533b\u5b66\u5f71\u50cf\uff08biomedical images\uff09\u5f88\u591a\u65f6\u5019\u90fd\u662f\u5757\u72b6\u7684\uff0c\u4e5f\u5c31\u662f\u8bf4\u662f\u7531\u5f88\u591a\u4e2a\u5207\u7247\u6784\u6210\u4e00\u6574\u5f20\u56fe\u7684\u5b58\u5728\u3002\u5982\u679c\u662f\u75282D\u7684\u56fe\u50cf\u5904\u7406\u6a21\u578b\u53bb\u5904\u74063D\u672c\u8eab\u4e0d\u662f\u4e0d\u53ef\u4ee5\uff0c\u4f46\u662f\u4f1a\u5b58\u5728\u4e00\u4e2a\u95ee\u9898\uff0c\u5c31\u662f\u4e0d\u5f97\u4e0d\u5c06\u751f\u7269\u533b\u5b66\u5f71\u50cf\u7684\u56fe\u7247\u4e00\u4e2aslice\u4e00\u4e2aslice\u6210\u7ec4\u7684\uff08\u5305\u542b\u8bad\u7ec3\u6570\u636e\u548c\u6807\u6ce8\u597d\u7684\u6570\u636e\uff09\u7684\u9001\u8fdb\u53bb\u8bbe\u8ba1\u7684\u6a21\u578b\u8fdb\u884c\u8bad\u7ec3\uff0c\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\u4f1a\u5b58\u5728\u4e00\u4e2a\u6548\u7387\u95ee\u9898\uff0c\u56e0\u800c\u5f88\u591a\u65f6\u5019\u5904\u7406\u5757\u72b6\u56fe\u7684\u65f6\u5019\u4f1a\u8ba9\u4efb\u611f\u5230\u4e0d\u9002\uff0c\u5e76\u4e14\u6570\u636e\u9884\u5904\u7406\u7684\u65b9\u5f0f\u4e5f\u76f8\u5bf9\u6bd4\u8f83\u7e41\u7410\uff08tedious\uff09\u3002<\/p>\n\n\n\n<p class=\"has-light-blue-background-color has-background\">   \u6240\u4ee5\uff0c\u8bba\u6587\u7684\u4f5c\u8005\u5c31\u63d0\u51fa\u6765\u4e863D -Net\u6a21\u578b\uff0c\u6a21\u578b\u4e0d\u4ec5\u89e3\u51b3\u4e86\u6548\u7387\u7684\u95ee\u9898\uff0c\u5e76\u4e14\u5bf9\u4e8e\u5757\u72b6\u56fe\u7684\u5207\u5272\u53ea\u8981\u6c42\u6570\u636e\u4e2d<strong>\u90e8\u5206\u5207\u7247\u88ab\u6807\u6ce8<\/strong>\u5373\u53ef\uff08\u53ef\u53c2\u8003\u4e0b\u56fe\u8bf4\u660e\uff09\u3002<\/p>\n\n\n\n<p><strong>\u6a21\u578b\u7ed3\u6784\uff08Network Architecture\uff09<\/strong><\/p>\n\n\n\n<p>  \u6574\u4e2a3D U-Net\u7684\u6a21\u578b\u662f\u57fa\u4e8e\u4e4b\u524dU-Net\uff082D\uff09\u521b\u5efa\u800c\u6765\uff0c\u540c\u6837\u5305\u542b\u4e86\u4e00\u4e2aencoder\u90e8\u5206\u548c\u4e00\u4e2adecoder\u90e8\u5206\uff0cencoder\u90e8\u5206\u662f\u7528\u6765\u5206\u6790\u6574\u5f20\u56fe\u7247\u5e76\u4e14\u8fdb\u884c\u7279\u5f81\u63d0\u53d6\u4e0e\u5206\u6790\uff0c\u800c\u4e0e\u4e4b\u76f8\u5bf9\u5e94\u7684decoder\u90e8\u5206\u662f\u751f\u6210\u4e00\u5f20\u5206\u5272\u597d\u7684\u5757\u72b6\u56fe\u3002\u8bba\u6587\u4e2d\u4f7f\u7528\u7684\u8f93\u5165\u56fe\u50cf\u7684\u5927\u5c0f\u662f132 * 132 * 116\uff0c\u6574\u4e2a\u7f51\u7edc\u7684\u7ed3\u6784\u524d\u534a\u90e8\u5206\uff08analysis path\uff09\u5305\u542b\u53ca\u4f7f\u7528\u5982\u4e0b\u5377\u79ef\u64cd\u4f5c\uff1a<\/p>\n\n\n\n<p>a. \u6bcf\u4e00\u5c42\u795e\u7ecf\u7f51\u7edc\u90fd\u5305\u542b\u4e86\u4e24\u4e2a 3 * 3 * 3\u7684\u5377\u79ef(convolution)<\/p>\n\n\n\n<p>b. Batch Normalization\uff08\u4e3a\u4e86\u8ba9\u7f51\u7edc\u80fd\u66f4\u597d\u7684\u6536\u655bconvergence\uff09<\/p>\n\n\n\n<p>c. ReLU<\/p>\n\n\n\n<p>d. Downsampling\uff1a2 * 2 * 2\u7684max_polling\uff0c\u6b65\u957fstride = 2<\/p>\n\n\n\n<p>\u800c\u4e0e\u4e4b\u76f8\u5bf9\u5e94\u7684\u5408\u6210\u8def\u5f84\uff08synthesis path\uff09\u5219\u6267\u884c\u4e0b\u9762\u7684\u64cd\u4f5c\uff1a<\/p>\n\n\n\n<p>a. upconvolution: 2 * 2 * 2\uff0c\u6b65\u957f=2<\/p>\n\n\n\n<p>b. \u4e24\u4e2a\u6b63\u5e38\u7684\u5377\u79ef\u64cd\u4f5c\uff1a3 * 3 * 3<\/p>\n\n\n\n<p>c. Batch Normalization<\/p>\n\n\n\n<p>d. ReLU<\/p>\n\n\n\n<p>   \u4e8e\u6b64\u540c\u65f6\uff0c\u9700\u8981\u628a\u5728analysis path\u4e0a\u76f8\u5bf9\u5e94\u7684\u7f51\u7edc\u5c42\u7684\u7ed3\u679c\u4f5c\u4e3adecoder\u7684\u90e8\u5206\u8f93\u5165\uff0c\u8fd9\u6837\u5b50\u505a\u7684\u539f\u56e0\u8ddfU-Net\u535a\u6587\u63d0\u5230\u7684\u4e00\u6837\uff0c\u662f\u4e3a\u4e86\u80fd\u91c7\u96c6\u5230\u7279\u5f81\u5206\u6790\u4e2d\u4fdd\u7559\u4e0b\u6765\u7684\u9ad8\u50cf\u7d20\u7279\u5f81\u4fe1\u606f\uff0c\u4ee5\u4fbf\u56fe\u50cf\u53ef\u4ee5\u66f4\u597d\u7684\u5408\u6210\u3002<\/p>\n\n\n\n<p>   \u6574\u4f53\u7684\u4e00\u4e2a\u7f51\u7edc\u7ed3\u6784\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u5176\u5b9e\u53ef\u4ee5\u770b\u51fa\u6765\u8ddf2D\u7ed3\u6784\u7684U-Net\u662f\u57fa\u672c\u4e00\u6837\uff0c\u552f\u4e00\u4e0d\u540c\u7684\u5c31\u662f\u5168\u90e82D\u64cd\u4f5c\u6362\u6210\u4e863D\uff0c\u8fd9\u6837\u5b50\u505a\u4e86\u4e4b\u540e\uff0c\u5bf9\u4e8evolumetric image\u5c31\u4e0d\u9700\u8981\u5355\u72ec\u8f93\u5165\u6bcf\u4e2a\u5207\u7247\u8fdb\u884c\u8bad\u7ec3\uff0c\u800c\u662f\u53ef\u4ee5\u91c7\u53d6\u56fe\u7247\u6574\u5f20\u4f5c\u4e3a\u8f93\u5165\u5230\u6a21\u578b\u4e2d\uff08PS\uff1a\u4f46\u662f\u5f53\u56fe\u50cf\u592a\u5927\u7684\u65f6\u5019\uff0c\u6b64\u65f6\u9700\u8981\u8fd0\u7528random crop\u7684\u6280\u5de7\u5c06\u56fe\u7247\u968f\u673a\u88c1\u5207\u6210\u56fa\u5b9a\u5927\u5c0f\u6a21\u5757\u7684\u56fe\u7247\u653e\u5165\u642d\u5efa\u7684\u6a21\u578b\u8fdb\u884c\u8bad\u7ec3\uff0c\u5f53\u7136\u8fd9\u662f\u540e\u8bdd\uff0c\u4e4b\u540e\u5c06\u4f1a\u5728\u5176\u4ed6\u6587\u7ae0\u4e2d\u8fdb\u884c\u4ecb\u7ecd\uff09\u3002\u9664\u6b64\u4e4b\u5916\uff0c\u8bba\u6587\u4e2d\u63d0\u5230\u7684\u4e00\u4e2a\u4eae\u70b9\u5c31\u662f\uff0c3D U-Net\u4f7f\u7528\u4e86weighted softmax loss function\u5c06\u672a\u6807\u8bb0\u7684\u50cf\u7d20\u70b9\u8bbe\u7f6e\u4e3a0\u4ee5\u81f3\u4e8e\u53ef\u4ee5\u8ba9\u7f51\u7edc\u53ef\u4ee5\u66f4\u591a\u5730\u4ec5\u4ec5\u5b66\u4e60\u6807\u6ce8\u5230\u7684\u50cf\u7d20\u70b9\uff0c\u4ece\u800c\u8fbe\u5230\u666e\u9002\u6027\u5730\u7279\u70b9\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"579\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-64-1024x579.png\" alt=\"\" class=\"wp-image-7480\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-64-1024x579.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-64-300x170.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-64-768x434.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/09\/image-64.png 1074w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>\u8bad\u7ec3\u7ec6\u8282\uff08Training\uff09<\/strong><\/p>\n\n\n\n<p>   3D U-Net\u540c\u6837\u91c7\u7528\u4e86\u6570\u636e\u589e\u5f3a\uff08data augmentation\uff09\u5730\u624b\u6bb5\uff0c\u4e3b\u8981\u7531rotation\u3001scaling\u548c\u5c06\u56fe\u50cf\u8bbe\u7f6e\u4e3agray\uff0c\u4e8e\u6b64\u540c\u65f6\u5728\u8bad\u7ec3\u6570\u636e\u4e0a\u548c\u771f\u5b9e\u6807\u6ce8\u7684\u6570\u636e\u4e0a\u8fd0\u7528\u5e73\u6ed1\u7684\u5bc6\u96c6\u53d8\u5f62\u573a(smooth dense deformation field)\uff0c\u4e3b\u8981\u662f\u901a\u8fc7\u4ece\u4e00\u4e2a\u6b63\u6001\u5206\u5e03\u7684\u968f\u673a\u5411\u91cf\u6837\u672c\u4e2d\u9009\u53d6\u6807\u51c6\u504f\u5dee\u4e3a4\u7684\u7f51\u683c\uff0c\u5728\u6bcf\u4e2a\u65b9\u5411\u4e0a\u5177\u670932\u4e2a\u4f53\u7d20\u7684\u95f4\u8ddd\uff0c\u7136\u540e\u5e94\u7528<a rel=\"noreferrer noopener\" href=\"http:\/\/www.cad.zju.edu.cn\/home\/zhx\/GM\/009\/00-bsia.pdf\" target=\"_blank\">B\u6837\u6761\u63d2\u503c<\/a>(B-Spline Interpolation\uff0c\u4e0d\u77e5\u9053\u4ec0\u4e48\u662fB\u6837\u6761\u63d2\u503c\u6cd5\u7684\u53ef\u4ee5\u70b9\u8fde\u63a5\u8fdb\u884c\u67e5\u770b\uff0c\u5728\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u7684\u521b\u5efa\u4e2d\u6709\u65f6\u5019\u4e5f\u4e0d\u9700\u8981\u90a3\u4e48\u590d\u6742\uff0c\u6240\u4ee5\u8fd9\u91cc\u4ec5\u9650\u4e86\u89e3\uff0c\u9664\u975e\u672c\u8eab\u6570\u5b66\u5e95\u5b50\u5f88\u597d\u5df2\u7ecf\u6709\u6240\u4e86\u89e3)\uff0cB\u6837\u6761\u63d2\u503c\u6cd5\u6bd4\u8f83\u7b3c\u7edf\u5730\u8bf4\u6cd5\u5c31\u662f\u5728\u539f\u672c\u5730\u5f62\u72b6\u4e0a\u627e\u5230\u4e00\u4e2a\u7c7b\u4f3c\u5730\u5f62\u72b6\u6765\u8fd1\u4f3c\uff08approximation\uff09\u3002\u4e4b\u540e\u5c31\u5bf9\u6570\u636e\u5f00\u59cb\u8fdb\u884c\u8bad\u7ec3\uff0c\u8bad\u7ec3\u91c7\u7528\u7684\u662f\u52a0\u6743\u4ea4\u53c9\u71b5\u635f\u5931\uff08weighted cross-entropy loss function\uff09\u4ee5\u81f3\u4e8e\u51cf\u5c11\u80cc\u666f\u7684\u6743\u91cd\u5e76\u589e\u52a0\u6807\u6ce8\u5230\u7684\u56fe\u50cf\u6570\u636e\u90e8\u5206\u7684\u6743\u91cd\u4ee5\u8fbe\u5230\u5e73\u8861\u7684\u5f71\u54cd\u5c0f\u7ba1\u548c\u80cc\u666f\u4f53\u7d20\u4e0a\u7684\u635f\u5931\u3002<\/p>\n\n\n\n<p>   \u5b9e\u9a8c\u7684\u7ed3\u679c\u662f\u7528IoU\uff08intersection over union\uff09\u8fdb\u884c\u8861\u91cf\u7684\uff0c\u5373\u6bd4\u8f83\u751f\u6210\u56fe\u50cf\u4e0e\u771f\u5b9e\u88ab\u6807\u6ce8\u90e8\u5206\u7684\u91cd\u53e0\u90e8\u5206\u3002<\/p>\n\n\n\n<p>   \u8bba\u6587\u9488\u5bf9\u80be\u810f\u7684\u751f\u7269\u533b\u5b66\u5f71\u50cf\u7684\u5206\u5272\u7ed3\u679c\u8fbe\u5230\u4e86IoU=86.3%\u7684\u7ed3\u679c\u30023D U-Net\u7684\u8bde\u751f\u5728\u533b\u5b66\u5f71\u50cf\u5206\u5272\uff0c\u7279\u522b\u662f\u90a3\u4e9bvolumetric images\u90fd\u662f\u7531\u5f88\u5927\u5e2e\u52a9\u7684\uff0c\u56e0\u4e3a\u5b83\u5f88\u5927\u7a0b\u5ea6\u4e0a\u89e3\u51b3\u4e863D\u56fe\u50cf\u4e00\u4e2a\u4e2aslice\u9001\u5165\u6a21\u578b\u8fdb\u884c\u8bad\u7ec3\u7684\u5c34\u5c2c\u5c40\u9762\uff0c\u4e5f\u5927\u5e45\u5ea6\u7684\u63d0\u5347\u8bad\u7ec3\u6548\u7387\uff0c\u5e76\u4e14\u4fdd\u7559\u4e86FCN\u548cU-Net\u672c\u6765\u5177\u5907\u7684\u4f18\u79c0\u7279\u5f81\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587\uff1a3D U-Net: Learning Dense Volumetric Segmentation fro &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2022\/09\/10\/3d-u-net\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">3D U-Net<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[24,9],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/7466"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=7466"}],"version-history":[{"count":15,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/7466\/revisions"}],"predecessor-version":[{"id":7483,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/7466\/revisions\/7483"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=7466"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=7466"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=7466"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}