{"id":1064,"date":"2026-05-17T08:36:36","date_gmt":"2026-05-17T00:36:36","guid":{"rendered":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/17\/2026-05-17-%e5%8c%bb%e5%ad%a6%e5%9b%be%e5%83%8f%e5%88%86%e5%89%b2%e8%ae%ba%e6%96%87%e7%b2%be%e8%af%bb%ef%bc%9asplitfed-cl-%e4%b8%8e-duetfair\/"},"modified":"2026-05-17T08:36:36","modified_gmt":"2026-05-17T00:36:36","slug":"2026-05-17-%e5%8c%bb%e5%ad%a6%e5%9b%be%e5%83%8f%e5%88%86%e5%89%b2%e8%ae%ba%e6%96%87%e7%b2%be%e8%af%bb%ef%bc%9asplitfed-cl-%e4%b8%8e-duetfair","status":"publish","type":"post","link":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/17\/2026-05-17-%e5%8c%bb%e5%ad%a6%e5%9b%be%e5%83%8f%e5%88%86%e5%89%b2%e8%ae%ba%e6%96%87%e7%b2%be%e8%af%bb%ef%bc%9asplitfed-cl-%e4%b8%8e-duetfair\/","title":{"rendered":"2026-05-17 \u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\u7cbe\u8bfb\uff1aSplitFed-CL \u4e0e DuetFair"},"content":{"rendered":"<h1>\u4eca\u65e5\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6700\u65b0\u8bba\u6587\u7cbe\u8bfb\u8ffd\u8e2a<\/h1>\n<h2>\u4eca\u65e5\u7ed3\u8bba<\/h2>\n<p>\u4eca\u5929\u68c0\u7d22\u5230\u7684\u6700\u65b0\u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\u4e2d\uff0c\u65b9\u6cd5\u521b\u65b0\u6700\u503c\u5f97\u5173\u6ce8\u7684\u4e0d\u662f\u53c8\u4e00\u4e2a U-Net \u6a21\u5757\u5806\u53e0\uff0c\u800c\u662f\u4e24\u4e2a\u66f4\u504f\u201c\u771f\u5b9e\u90e8\u7f72\u95ee\u9898\u201d\u7684\u65b9\u5411\uff1a\u4e00\u7bc7\u5904\u7406 Split Federated Learning \u573a\u666f\u4e0b\u591a\u4e2d\u5fc3\u4e0d\u51c6\u786e\u6807\u6ce8\uff0c\u53e6\u4e00\u7bc7\u5904\u7406\u533b\u5b66\u5206\u5272\u516c\u5e73\u6027\u4e2d\u201c\u7ec4\u5185\u56f0\u96be\u6837\u672c\u88ab\u5e73\u5747\u503c\u63a9\u76d6\u201d\u7684\u95ee\u9898\u3002\u4e24\u7bc7\u5747\u4e3a 2026 \u5e74 arXiv \u9884\u5370\u672c\uff0c\u5747\u665a\u4e8e 2025 \u5e74\uff0c\u4e14 PDF \u53ef\u76f4\u63a5\u83b7\u53d6\uff1b\u5b83\u4eec\u5bf9 polyp \/ DAMamba \u4e3b\u5e72\u8bbe\u8ba1\u4e0d\u662f\u76f4\u63a5 backbone \u66ff\u6362\uff0c\u4f46\u5bf9\u5f31\u6807\u6ce8\u3001\u591a\u4e2d\u5fc3\u6cdb\u5316\u3001\u4e34\u5e8a\u90e8\u7f72\u53ef\u4fe1\u6027\u76f8\u5173\u7ae0\u8282\u548c\u5b9e\u9a8c\u8bbe\u8ba1\u5f88\u6709\u53c2\u8003\u4ef7\u503c\u3002<\/p>\n<h2>\u68c0\u7d22\u8bf4\u660e<\/h2>\n<p>\u672c\u6b21\u4f18\u5148\u68c0\u7d22 arXiv 2026-05-11 \u81f3 2026-05-14 \u65b0\u8fd1\u63d0\u4ea4\u7684 medical image segmentation \/ federated segmentation \/ fairness segmentation \/ 3D medical segmentation \u8bba\u6587\uff0c\u5e76\u5bf9\u5386\u53f2 cron \u8f93\u51fa\u4e2d\u7684\u5df2\u63a8\u8350\u6807\u9898\u4e0e arXiv ID \u505a\u4e86\u53bb\u91cd\u3002\u6240\u6709\u5165\u9009\u8bba\u6587\u5747\u4e3a 2025 \u5e74\u53ca\u4ee5\u540e\u8bba\u6587\uff1b\u7531\u4e8e\u4eca\u5929\u672a\u53d1\u73b0\u65b0\u7684\u5df2\u63a5\u6536 MICCAI\/CVPR\/ICCV\/ECCV\/NeurIPS\/ICLR \u9876\u4f1a\u6b63\u5f0f\u5206\u5272\u8bba\u6587 PDF\uff0c\u6700\u7ec8\u9009\u62e9\u4e86\u4e24\u7bc7\u6700\u65b0\u4e14\u5b9e\u9a8c\u8f83\u5b8c\u6574\u3001\u95ee\u9898\u8bbe\u7f6e\u66f4\u6709\u7814\u7a76\u4ef7\u503c\u7684 arXiv preprint\u3002\u5df2\u68c0\u67e5\u5386\u53f2\u63a8\u8350\u8bb0\u5f55\u5e76\u6392\u9664\u4e86\u91cd\u590d\u8bba\u6587\uff1b\u672c\u6b21\u8df3\u8fc7\u7684\u5386\u53f2\u63a8\u8350\u5019\u9009\u5305\u62ec Med-DisSeg\u3001SpectraFlow\u3001FEFormer\u3001USEMA\u3001GeoProto\u3001XTinyU-Net\u3001MedCore\u3001TopoMamba\u3001ESICA\u3001SemiSAM-O1 \u7b49\u3002<\/p>\n<h2>WordPress \u53d1\u5e03<\/h2>\n<ul>\n<li>WordPress \u6587\u7ae0\u94fe\u63a5\uff1a\u5f85\u53d1\u5e03\u540e\u586b\u5199<\/li>\n<li>WordPress Post ID\uff1a\u5f85\u53d1\u5e03\u540e\u586b\u5199<\/li>\n<\/ul>\n<hr \/>\n<h2>\u8bba\u6587 1\uff1aSplitFed-CL: A Split Federated Co-Learning Framework for Medical Image Segmentation with Inaccurate Labels<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aSplitFed-CL: A Split Federated Co-Learning Framework for Medical Image Segmentation with Inaccurate Labels<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aZahra Hafezi Kafshgari, Hadi Hadizadeh, Parvaneh Saeedi \/ Zahra Hafezi Kafshgari<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-11 arXiv v1\uff1b\u8bba\u6587\u9996\u9875\u6807\u6ce8\u5c06\u51fa\u73b0\u5728 ICIP 2026 proceedings<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint \/ ICIP 2026 proceedings \u6807\u6ce8\u7248\u672c<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.11060<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.11060<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1a\u672a\u83b7\u53d6 \/ \u8bba\u6587\u6b63\u6587\u672a\u63d0\u4f9b\u660e\u786e\u4ee3\u7801\u94fe\u63a5<\/li>\n<li>\u4efb\u52a1\uff1aSplit Federated Learning \u4e0b\u7684\u533b\u5b66\u56fe\u50cf\u5206\u5272\uff0c\u91cd\u70b9\u662f\u4e0d\u51c6\u786e \/ \u5f02\u8d28\u6807\u6ce8\u8d28\u91cf\u4e0b\u7684\u9c81\u68d2\u8bad\u7ec3<\/li>\n<li>\u6570\u636e\u96c6\uff1aHuman Embryo \u56db\u7c7b\u663e\u5fae\u56fe\u50cf\u5206\u5272\uff1bPSFHS \u4ea7\u65f6\u7ecf\u4f1a\u9634\u8d85\u58f0 pubic symphysis \/ fetal head \u5206\u5272\uff1bISIC \/ ISIC MultiAnnot++ \u76ae\u80a4\u75c5\u7076\u5206\u5272<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1aSplitFed + student-teacher co-learning + label reliability detection + label correction + consistency regularization + reliability-aware aggregation<\/li>\n<\/ul>\n<h3>paper-deep-reader \u7cbe\u8bfb\u7ed3\u679c<\/h3>\n<h4>1. \u4e00\u53e5\u8bdd\u7ed3\u8bba<\/h4>\n<p>\u8fd9\u7bc7\u8bba\u6587\u6700\u6709\u4ef7\u503c\u7684\u5730\u65b9\uff0c\u662f\u628a\u591a\u4e2d\u5fc3\u533b\u5b66\u5206\u5272\u4e2d\u5404\u5ba2\u6237\u7aef\u6807\u6ce8\u8d28\u91cf\u4e0d\u540c\u7684\u95ee\u9898\u5efa\u6a21\u4e3a SplitFed \u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u53ef\u9760\u6837\u672c\u8bc6\u522b\u3001\u4f2a\u6807\u7b7e\u4fee\u6b63\u4e0e\u53ef\u9760\u6027\u52a0\u6743\u805a\u5408\u95ee\u9898\uff0c\u800c\u4e0d\u662f\u53ea\u505a\u5e38\u89c4 FedAvg \u6216\u7b80\u5355 noisy-label loss\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u8bba\u6587\u7814\u7a76\u7684\u662f Split Federated Learning \u573a\u666f\u4e0b\u7684\u533b\u5b66\u56fe\u50cf\u5206\u5272\uff1a\u6bcf\u4e2a\u5ba2\u6237\u7aef\u4fdd\u7559\u6570\u636e\u672c\u5730\uff0c\u6a21\u578b\u88ab\u5207\u5206\u4e3a client front-end\u3001server-side heavy computation \u548c client back-end\uff0c\u4ee5\u517c\u987e\u9690\u79c1\u548c\u8fb9\u7f18\u8bbe\u5907\u7b97\u529b\u3002\u533b\u5b66\u5206\u5272\u7684\u73b0\u5b9e\u95ee\u9898\u662f\u4e0d\u540c\u533b\u9662\u3001\u8bbe\u5907\u3001\u6807\u6ce8\u8005\u7ecf\u9a8c\u4f1a\u5bfc\u81f4\u6807\u7b7e\u8d28\u91cf\u5dee\u5f02\uff0c\u5c24\u5176\u8fb9\u754c\u533a\u57df\u5e38\u6709\u4e0d\u4e00\u81f4\uff1b\u5982\u679c\u76f4\u63a5\u628a\u8fd9\u4e9b\u5ba2\u6237\u7aef\u4e00\u8d77\u8bad\u7ec3\uff0c\u5168\u5c40\u6a21\u578b\u4f1a\u88ab\u4f4e\u8d28\u91cf\u6807\u7b7e\u62d6\u504f\u3002<\/p>\n<p>paper map \u53ef\u6982\u62ec\u4e3a\uff1a\u8bba\u6587\u7814\u7a76 SplitFed \u533b\u5b66\u56fe\u50cf\u5206\u5272\u4e2d\u5f02\u8d28\u4e0d\u51c6\u786e\u6807\u6ce8\u95ee\u9898\uff1b\u4e3b\u7ebf\u662f\u5728 DeepLabV3+ SplitFed \u67b6\u6784\u4e2d\u5f15\u5165\u5168\u5c40 teacher\u3001\u53ef\u9760\u6837\u672c\u9608\u503c\u3001student-teacher \u6807\u7b7e\u4fee\u6b63\u3001consistency loss \u548c\u53ef\u9760\u6027\u805a\u5408\uff1b\u58f0\u79f0\u5728\u4e24\u4e2a\u5408\u6210\u566a\u58f0\u591a\u7c7b\u5206\u5272\u6570\u636e\u96c6\u548c\u4e00\u4e2a\u771f\u5b9e\u6807\u6ce8\u566a\u58f0 ISIC \u6570\u636e\u96c6\u4e0a\u4f18\u4e8e FedAvg\u3001FedMix\u3001FedNCL\u3001ARFL\u3001QA-SplitFed\u3001CELC\u3001DHLC\uff1b\u5173\u952e\u6280\u672f\u5bf9\u8c61\u662f difficulty-guided annotation error simulation\u3001tau \u9608\u503c\u3001\u53ef\u9760\/\u4e0d\u53ef\u9760\u6837\u672c\u96c6\u5408\u3001\u4fee\u6b63\u6807\u7b7e\u3001\u53ef\u9760\u6027\u6743\u91cd\u3001\u53ef\u5b66\u4e60 loss \u6743\u91cd\uff1b\u4e3b\u8981\u98ce\u9669\u662f\u566a\u58f0\u6a21\u578b\u548c SplitFed \u5b9e\u9a8c\u89c4\u6a21\u4ecd\u504f\u53d7\u63a7\uff0c\u672a\u5145\u5206\u8bc1\u660e\u771f\u5b9e\u591a\u4e2d\u5fc3\u5927\u89c4\u6a21\u90e8\u7f72\u4e0b\u7684\u901a\u4fe1\u3001\u9690\u79c1\u6cc4\u9732\u548c\u6cdb\u5316\u8868\u73b0\u3002<\/p>\n<p>\u8def\u7ebf\u8bb0\u5f55\uff1aPrimary adapter = method-algorithm\uff1bSecondary adapter = benchmark-evaluation\uff1bEvidence packs = general\u3001experimental-eval\u3001robustness-and-ood\u3001reproducibility-and-compute\uff1bRoute confidence = \u9ad8\u3002\u539f\u56e0\uff1a\u8bba\u6587\u6838\u5fc3\u662f\u4e00\u4e2a\u8bad\u7ec3\u6846\u67b6 \/ \u7b97\u6cd5\uff0c\u8bc1\u636e\u4e3b\u8981\u6765\u81ea\u591a\u6570\u636e\u96c6\u9c81\u68d2\u6027\u6bd4\u8f83\u4e0e\u6d88\u878d\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u6307\u51fa\uff0c\u5df2\u6709 noisy-label federated learning \u591a\u6570\u9762\u5411\u5206\u7c7b\uff0c\u96be\u4ee5\u76f4\u63a5\u8fc1\u79fb\u5230\u533b\u5b66\u5206\u5272\uff0c\u56e0\u4e3a\u5206\u5272\u6807\u7b7e\u9519\u8bef\u5f80\u5f80\u96c6\u4e2d\u5728\u6a21\u7cca\u8fb9\u754c\u3001\u66f2\u7387\u9ad8\u3001\u5f31\u8fb9\u7f18\u533a\u57df\uff0c\u800c\u4e0d\u662f\u968f\u673a label flip\u3002\u5df2\u6709\u533b\u5b66 FL \/ SplitFed \u65b9\u6cd5\u5982 FedMix\u3001FedA3I\u3001FedDM\u3001QA-SplitFed\u3001DHLC \/ CELC \u8981\u4e48\u5904\u7406\u6709\u9650\u76d1\u7763\u548c\u5f31\u76d1\u7763\uff0c\u8981\u4e48\u505a client quality weighting \u6216 lesion label refinement\uff0c\u4f46\u5bf9 SplitFed \u4e0b\u6837\u672c\u7ea7\u53ef\u9760\u6027\u8bc6\u522b\u3001\u6807\u7b7e\u5c40\u90e8\u4fee\u6b63\u548c\u5ba2\u6237\u7aef\u805a\u5408\u8d28\u91cf\u540c\u65f6\u5efa\u6a21\u4e0d\u8db3\u3002\u4f5c\u8005\u8fd8\u6279\u8bc4\u5e38\u89c1\u566a\u58f0\u6a21\u62df\u8fc7\u4e8e\u7b80\u5355\uff0c\u5982\u968f\u673a\u6807\u7b7e\u6253\u4e71\u3001\u57fa\u672c dilation \/ erosion\uff0c\u4e0d\u80fd\u53cd\u6620\u4eba\u7c7b\u8fb9\u754c\u6807\u6ce8\u9519\u8bef\u3002<\/p>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>\u6574\u4f53\u6846\u67b6\u57fa\u4e8e DeepLabV3+\uff0c\u4f7f\u7528 ResNet50 encoder\u3002\u7f51\u7edc\u88ab\u5207\u5206\u4e3a\u4e09\u6bb5\uff1a\u5ba2\u6237\u7aef front-end \u7f16\u7801\u8f93\u5165\u5e76\u4e0a\u4f20\u4e2d\u95f4\u6fc0\u6d3b\uff1b\u670d\u52a1\u5668\u5b50\u6a21\u578b\u6267\u884c\u91cd\u8ba1\u7b97\uff1b\u5ba2\u6237\u7aef back-end \u89e3\u7801\u8f93\u51fa\u5206\u5272 mask\u3002\u6bcf\u8f6e\u8bad\u7ec3\u4e2d\uff0c\u5ba2\u6237\u7aef\u5b66\u751f\u6a21\u578b\u66f4\u65b0\uff0c\u5168\u7403\u5e73\u5747\u6a21\u578b\u4f5c\u4e3a teacher \u63d0\u4f9b\u7a33\u5b9a\u9884\u6d4b\u3002<\/p>\n<p>\u6838\u5fc3\u8bad\u7ec3\u76ee\u6807\u5305\u542b\u4e09\u90e8\u5206\uff1a\u53ef\u9760\u6837\u672c\u7684 region loss\uff1b\u4e0d\u53ef\u9760\u6837\u672c\u7ecf student-teacher \u4fee\u6b63\u540e\u7684 region loss\uff1b\u6270\u52a8\u8f93\u5165\u4e0a\u7684 consistency loss\u3002\u53ef\u9760\u6837\u672c\u76f4\u63a5\u7528\u539f\u6807\u7b7e\u8bad\u7ec3\uff1b\u4e0d\u53ef\u9760\u6837\u672c\u5148\u7531 student-teacher \u4fee\u6b63\uff1b\u6270\u52a8\u8f93\u5165\u4e0a\u7684 student \u9884\u6d4b\u8981\u4e0e teacher \u5728\u539f\u8f93\u5165\u4e0a\u7684\u9884\u6d4b\u4fdd\u6301\u4e00\u81f4\u3002\u540e\u7eed\u4f5c\u8005\u628a\u56fa\u5b9a alpha\u3001beta \u6362\u6210\u53ef\u5b66\u4e60\u6743\u91cd\uff0c\u5e76\u7528 warm-up \u907f\u514d\u65e9\u671f\u4f2a\u6807\u7b7e\u8fc7\u5f3a\u5f71\u54cd\u3002<\/p>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<ol>\n<li>\n<p><strong>Difficulty-guided \u6807\u6ce8\u9519\u8bef\u6a21\u62df<\/strong>\uff1a\u4f5c\u8005\u5148\u7528 signed distance function \u5b9a\u4e49\u8fb9\u754c\u7a84\u5e26\uff0c\u4f7f\u8fb9\u754c\u5e26\u968f\u76ee\u6807\u5c3a\u5ea6\u81ea\u9002\u5e94\u3002\u7136\u540e\u5728\u8fb9\u754c\u5e26\u5185\u8ba1\u7b97\u4e09\u7c7b\u56f0\u96be\u5ea6\uff1a\u5f31\u8fb9\u7f18\u3001\u6a21\u7cca\u3001\u66f2\u7387\uff0c\u7ec4\u5408\u4e3a difficulty map\u3002\u518d\u628a\u56f0\u96be\u5ea6\u6620\u5c04\u4e3a\u5f62\u53d8\u5e45\u5ea6\uff0c\u65b9\u5411\u7531\u8fb9\u754c\u5185\u5916 edge evidence \u7684\u5dee\u5f02\u51b3\u5b9a\uff0c\u6700\u540e\u5f62\u53d8 SDF \u5e76\u9608\u503c\u5f97\u5230 noisy mask\u3002\u8fd9\u4e2a\u6a21\u5757\u7684\u610f\u4e49\u662f\u8ba9\u5408\u6210\u566a\u58f0\u66f4\u50cf\u771f\u5b9e\u5206\u5272\u8fb9\u754c\u9519\u8bef\uff1b\u521b\u65b0\u6027\u6bd4\u968f\u673a\u8150\u8680\/\u81a8\u80c0\u66f4\u5f3a\uff0c\u4f46\u4ecd\u662f hand-crafted noise model\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u53ef\u9760 \/ \u4e0d\u53ef\u9760\u6837\u672c\u8bc6\u522b<\/strong>\uff1a\u5bf9 batch \u6837\u672c\u8ba1\u7b97 student \u4e0e teacher \u7684 per-sample region loss\uff0c\u82e5\u4e8c\u8005\u5747\u4f4e\u4e8e\u5168\u5c40\u9608\u503c tau\uff0c\u5212\u4e3a\u53ef\u9760\u96c6\u5408\uff1b\u82e5\u4e8c\u8005\u5747\u9ad8\u4e8e tau\uff0c\u5212\u4e3a\u4e0d\u53ef\u9760\u96c6\u5408\u3002tau \u521d\u59cb\u8f83\u9ad8\uff0c\u8ba9\u8bad\u7ec3\u65e9\u671f\u5c3d\u91cf\u5229\u7528\u6570\u636e\uff1b\u968f\u540e\u7531\u8de8\u5ba2\u6237\u7aef loss \u5747\u503c\u548c\u65b9\u5dee\u7ec4\u6210\u7edf\u8ba1\u91cf\uff0c\u518d\u6309 performance ratio \u52a0\u6743\u5f97\u5230\uff0c\u5e76\u968f\u8bad\u7ec3\u9010\u6e10\u4e25\u683c\u3002<\/p>\n<\/li>\n<li>\n<p><strong>Student-teacher \u6807\u7b7e\u4fee\u6b63<\/strong>\uff1a\u5bf9\u4e0d\u53ef\u9760\u6837\u672c\uff0c\u6784\u9020 student hard mask\u3001teacher hard mask \u4e0e\u539f\u6807\u7b7e\u4e4b\u95f4\u7684 symmetric difference \u533a\u57df\uff0c\u53ea\u5728\u4e0d\u4e00\u81f4\u533a\u57df\u505a\u4fee\u6b63\u3002\u82e5 student \u6216 teacher \u5728\u533a\u57df\u5185\u7f6e\u4fe1\u5ea6\u8d85\u8fc7\u9608\u503c T=0.9\uff0c\u7528\u76f8\u5e94\u9884\u6d4b\u66ff\u6362\u539f\u6807\u7b7e\uff1b\u5426\u5219\u4fdd\u7559\u539f\u6807\u7b7e\u3002\u8fd9\u4e2a\u8bbe\u8ba1\u9002\u5408\u8fb9\u754c\u4fee\u6b63\uff0c\u4f46\u5bf9\u7cfb\u7edf\u6027\u9519\u8bef\u6216 teacher \u65e9\u671f\u9519\u8bef\u6709\u4f9d\u8d56\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u53ef\u9760\u6027\u52a0\u6743 SplitFed \u805a\u5408<\/strong>\uff1a\u5ba2\u6237\u7aef\u4e0a\u4f20\u53ef\u9760\u6837\u672c\u6570\u4e0e\u53ef\u9760\u5b50\u96c6 loss\u3002\u670d\u52a1\u5668\u8ba1\u7b97 data ratio \u548c performance ratio\uff0c\u518d\u5bf9 FE\/S\/BE \u4e09\u6bb5\u53c2\u6570\u52a0\u6743\u5e73\u5747\u3002\u76f8\u6bd4 FedAvg \u6309\u6837\u672c\u91cf\u805a\u5408\uff0c\u4f4e\u635f\u5931\u4e14\u53ef\u9760\u6837\u672c\u591a\u7684\u5ba2\u6237\u7aef\u6743\u91cd\u66f4\u9ad8\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u53ef\u5b66\u4e60 loss \u6743\u91cd<\/strong>\uff1a\u8bba\u6587\u4f7f\u7528 sigmoid \u53c2\u6570\u5316\u7684 unreliable-label loss \u6743\u91cd\u4e0e consistency loss \u6743\u91cd\uff0c\u914d\u5408 warm-up \u548c EMA normalization\uff0c\u907f\u514d\u4eba\u5de5\u8c03 alpha\u3001beta\u3002\u8fd9\u5bf9\u5de5\u7a0b\u590d\u73b0\u6709\u610f\u4e49\uff0c\u4f46\u4e5f\u589e\u52a0\u4e86\u8c03\u53c2\u7ec6\u8282\u3002<\/p>\n<\/li>\n<\/ol>\n<p>\u8fc1\u79fb\u4ef7\u503c\uff1a\u8fd9\u4e9b\u6a21\u5757\u4e0d\u662f\u76f4\u63a5\u7528\u4e8e polyp backbone \u7684\u7ed3\u6784\u521b\u65b0\uff0c\u4f46 difficulty-guided boundary-noise simulation \u4e0e student-teacher \u4fee\u6b63\u5f88\u9002\u5408\u8fc1\u79fb\u5230 polyp segmentation \u7684\u5f31\u6807\u6ce8 \/ noisy annotation \u5b9e\u9a8c\uff1bSplitFed \u805a\u5408\u90e8\u5206\u9002\u5408\u591a\u4e2d\u5fc3\u534f\u4f5c\u8bad\u7ec3\uff0c\u4e0d\u9002\u5408\u5355\u673a DAMamba \u4e3b\u5e72\u6539\u9020\u3002<\/p>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p>\u5b9e\u9a8c\u5305\u542b\u4e09\u7c7b\u6570\u636e\uff1aHuman Embryo 594 \u5f20\u663e\u5fae\u56fe\u50cf\uff0c\u56db\u7c7b\u7ed3\u6784\u5206\u5272\uff0c\u56db\u5ba2\u6237\u7aef\u5f02\u8d28\u5212\u5206\uff1bPSFHS 1358 \u5f20\u8d85\u58f0\u56fe\u50cf\uff0c\u5206\u5272 pubic symphysis \u548c fetal head\uff0c\u5ba2\u6237\u7aef corruption ratios \u4e3a 20%\u300150%\u300180%\u30010%\uff1bISIC \/ IMA++ \u4e2d\u9009\u53d6 T2\/S2 semi-automated + novice verified \u4f5c\u4e3a\u771f\u5b9e\u4e0d\u53ef\u9760\u6807\u6ce8\uff0cT1\/S1 expert mask \u4f5c\u4e3a\u8f83\u51c6\u786e\u6807\u6ce8\uff0c\u5e76\u6784\u5efa\u4e94\u5ba2\u6237\u7aef 0%\u300150%\u300150%\u300180%\u300160% \u566a\u58f0\u573a\u666f\u3002\u6a21\u578b\u8f93\u5165 resize \u5230 352x352\uff0cAdam \u5b66\u4e60\u7387 1e-4\uff0c5 local epochs \/ round\uff0c100 global epochs\uff0c\u9608\u503c T=0.9\u3002<\/p>\n<p>\u4e3b\u8981\u7ed3\u679c\uff1aPSFHS \u4e0a SplitFed-CL full \u7684 accuracy 0.9788\u3001Dice loss 0.0363\u3001mean IoU 0.9592\u3001PS IoU 0.8805\u3001FH IoU 0.9424\uff0c\u4f18\u4e8e QA-SplitFed\u3001FedAvg\u3001FedMix\u3001FedNCL-V2\u3001ARFL\u3001CELC\u3001DHLC\u3002Human Embryo \u4e0a full method accuracy 0.9451\u3001Dice loss 0.0725\u3001mean IoU 0.8995\uff0c\u9ad8\u4e8e\u591a\u79cd baseline\uff0c\u5e76\u4f18\u4e8e no label correction \u548c no consistency loss \u6d88\u878d\u3002ISIC \u4e0a full method accuracy 0.9830\u3001Dice loss 0.1320\u3001FG IoU 0.7641\u3001precision 0.9074\u3001recall 0.9157\uff0c\u63a5\u8fd1 clean-label FedAvg \u7684 FG IoU 0.7641\uff0c\u5e76\u4f18\u4e8e QA-SplitFed 0.7432\u3002Figure 3 \u663e\u793a PSFHS \u4e0a\u4f30\u8ba1\u8bad\u7ec3\u566a\u58f0\u6bd4\u4f8b\u63a5\u8fd1\u6ce8\u5165\u6bd4\u4f8b\uff1a\u771f\u5b9e 20%\u300150%\u300180%\u30010%\uff0c\u68c0\u6d4b\u4e3a 18.8%\u300145.6%\u300174.8%\u30010.4%\u3002<\/p>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u53ef\u4fe1\u70b9\uff1a\u5b9e\u9a8c\u8986\u76d6\u591a\u7c7b\u663e\u5fae\u56fe\u50cf\u3001\u8d85\u58f0\u548c\u76ae\u80a4\u75c5\u7076\uff1b\u5305\u542b\u5408\u6210\u566a\u58f0\u4e0e\u771f\u5b9e\u6807\u6ce8\u566a\u58f0\uff1bbaseline \u4e0d\u5c11\uff0c\u5e76\u5305\u62ec QA-SplitFed \u8fd9\u4e00\u76f4\u63a5\u76f8\u5173 SplitFed \u65b9\u6cd5\uff1b\u6d88\u878d\u663e\u793a label correction \u4e0e consistency loss \u5747\u6709\u8d21\u732e\uff1b\u566a\u58f0\u6bd4\u4f8b\u4f30\u8ba1\u548c label correction \u53ef\u89c6\u5316\u652f\u6301\u673a\u5236\u89e3\u91ca\u3002<\/p>\n<p>\u4e0d\u8db3\u4e5f\u5f88\u660e\u663e\uff1a\u7b2c\u4e00\uff0c\u6b63\u6587\u6ca1\u6709\u62a5\u544a\u968f\u673a\u79cd\u5b50\u65b9\u5dee \/ \u7f6e\u4fe1\u533a\u95f4\uff0c\u7edf\u8ba1\u663e\u8457\u6027\u4e0d\u6e05\u695a\u3002\u7b2c\u4e8c\uff0c\u5927\u591a\u6570\u566a\u58f0\u4ecd\u6765\u81ea\u4f5c\u8005\u81ea\u5b9a\u4e49\u7684 difficulty-guided \u5f62\u53d8\uff0c\u771f\u5b9e\u4e34\u5e8a\u591a\u4e2d\u5fc3\u6807\u6ce8\u8bef\u5dee\u53ef\u80fd\u5305\u542b\u6f0f\u6807\u3001\u7c7b\u522b\u6df7\u6dc6\u3001protocol \u5dee\u5f02\uff0c\u4e0d\u53ea\u662f\u8fb9\u754c\u5f62\u53d8\u3002\u7b2c\u4e09\uff0cSplitFed \u7684\u901a\u4fe1\u91cf\u3001\u9690\u79c1\u6cc4\u9732\u98ce\u9669\u3001server activation inversion \u7b49\u6ca1\u6709\u5c55\u5f00\u3002\u7b2c\u56db\uff0c\u4ee3\u7801\u94fe\u63a5\u672a\u63d0\u4f9b\uff0c\u590d\u73b0\u96be\u5ea6\u8f83\u9ad8\u3002\u7b2c\u4e94\uff0cDeepLabV3+ \/ ResNet50 \u4e3a\u57fa\u7840\u6a21\u578b\uff0c\u672a\u9a8c\u8bc1 nnU-Net\u3001UNetR\u3001Swin-UNet\u3001MedNeXt\u3001Mamba \u7cfb\u5217\u5206\u5272\u6846\u67b6\u4e0b\u662f\u5426\u540c\u6837\u6709\u6548\u3002<\/p>\n<h4>8. \u4e0e\u4e3b\u6d41\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u7684\u5173\u7cfb<\/h4>\n<p>\u5b83\u4e0d\u662f U-Net \/ nnU-Net \/ Transformer \/ Mamba \u7684\u65b0\u4e3b\u5e72\uff0c\u800c\u662f\u8bad\u7ec3\u8303\u5f0f\u548c\u9c81\u68d2\u5b66\u4e60\u6846\u67b6\u3002\u4e0e U-Net \/ nnU-Net \u7684\u5173\u7cfb\uff1a\u53ef\u628a SplitFed-CL \u7684\u53ef\u9760\u6837\u672c\u8bc6\u522b\u3001\u6807\u7b7e\u4fee\u6b63\u548c\u805a\u5408\u7b56\u7565\u6302\u5230 nnU-Net \u6216 3D U-Net \u7684 federated \/ split-federated \u8bad\u7ec3\u4e0a\u3002\u4e0e TransUNet \/ Swin-UNet \/ UNetR \u7684\u5173\u7cfb\uff1a\u5982\u679c server-side \u80fd\u627f\u8f7d Transformer bottleneck\uff0c\u5b83\u4e5f\u53ef\u7528\u4e8e\u8fd9\u4e9b\u6a21\u578b\uff0c\u4f46\u8bba\u6587\u672a\u9a8c\u8bc1\u3002\u4e0e DAMamba \/ SegMamba \u7684\u5173\u7cfb\uff1a\u6ca1\u6709\u72b6\u6001\u7a7a\u95f4\u6a21\u5757\uff1b\u53ef\u501f\u9274\u5176 noisy-label \u548c consistency training \u903b\u8f91\uff0c\u4e3a Mamba-based medical segmentation \u8bbe\u8ba1\u591a\u4e2d\u5fc3\u9c81\u68d2\u8bad\u7ec3\u5b9e\u9a8c\u3002\u4e0e foundation model \u7684\u5173\u7cfb\uff1a\u4e0d\u4f9d\u8d56 SAM \/ MedSAM\uff1b\u4f46 teacher-student label correction \u548c\u4e0d\u53ef\u9760\u6807\u6ce8\u4fee\u6b63\u601d\u8def\u53ef\u4e0e foundation model pseudo-label \u7ed3\u5408\u3002<\/p>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u5bf9 polyp segmentation\uff1a\u4e2d\u7b49\u504f\u9ad8\u3002\u5b83\u7684 boundary-centric error simulation \u5f88\u9002\u5408\u6a21\u62df\u606f\u8089\u8fb9\u754c\u6a21\u7cca\u6807\u6ce8\uff1bstudent-teacher correction \u53ef\u7528\u4e8e Kvasir-SEG \/ CVC-ClinicDB \/ ETIS \u7b49\u6570\u636e\u7684 noisy-label robustness \u5b9e\u9a8c\u3002\u5bf9 DAMamba \u6539\u9020\uff1a\u4e0d\u662f\u7ed3\u6784\u6a21\u5757\uff0c\u4e0d\u80fd\u76f4\u63a5\u4f5c\u4e3a DAMamba block\uff0c\u4f46\u53ef\u4f5c\u4e3a\u8bad\u7ec3\u9c81\u68d2\u6027 \/ \u591a\u4e2d\u5fc3\u6cdb\u5316\u6269\u5c55\u5b9e\u9a8c\u3002\u5bf9 introduction \/ related work\uff1a\u5982\u679c\u8981\u8ba8\u8bba\u4e34\u5e8a\u90e8\u7f72\u3001\u591a\u4e2d\u5fc3\u9690\u79c1\u8bad\u7ec3\u3001\u4e0d\u51c6\u786e\u6807\u6ce8\uff0c\u8fd9\u662f\u5f88\u597d\u7684\u8fd1\u671f\u5f15\u7528\u3002\u5bf9 baseline\uff1a\u53ea\u6709\u5728\u505a federated \/ split-federated polyp segmentation \u65f6\u624d\u9002\u5408\u4f5c\u4e3a baseline\uff1b\u666e\u901a\u5355\u4e2d\u5fc3 polyp segmentation \u4e0d\u5fc5\u5f3a\u884c\u6bd4\u8f83\u3002<\/p>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5efa\u8bae\u7cbe\u8bfb<\/strong>\u3002\u5982\u679c\u8fd1\u671f\u76ee\u6807\u662f\u4e3b\u5e72\u7ed3\u6784\u521b\u65b0\uff0c\u53ef\u5148\u8bfb\u65b9\u6cd5\u548c\u5b9e\u9a8c\u8868\uff1b\u5982\u679c\u8ba1\u5212\u5199 noisy-label\u3001\u5f31\u6807\u6ce8\u3001\u591a\u4e2d\u5fc3 polyp segmentation \u6216 clinical deployment \u7ae0\u8282\uff0c\u5efa\u8bae\u5168\u6587\u7cbe\u8bfb\u5e76\u590d\u73b0 difficulty-guided noise simulation\u3002<\/p>\n<hr \/>\n<h2>\u8bba\u6587 2\uff1aDuetFair: Coupling Inter- and Intra-Subgroup Robustness for Fair Medical Image Segmentation<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aDuetFair: Coupling Inter- and Intra-Subgroup Robustness for Fair Medical Image Segmentation<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aYiqi Tian, Sangjoon Park, Bo Zeng, Pengfei Jin, Yujin Oh, Quanzheng Li \/ Yiqi Tian<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-11 arXiv v1<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.10521<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.10521<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1a\u672a\u83b7\u53d6 \/ \u8bba\u6587\u6b63\u6587\u672a\u63d0\u4f9b\u660e\u786e\u4ee3\u7801\u94fe\u63a5<\/li>\n<li>\u4efb\u52a1\uff1a\u516c\u5e73\u6027\u533b\u5b66\u56fe\u50cf\u5206\u5272\uff1b2D fundus optic cup\/rim\u30012D skin lesion\u30013D radiotherapy target segmentation<\/li>\n<li>\u6570\u636e\u96c6\uff1aHarvard-FairSeg\uff1bHAM10000\uff1bin-house 3D pelvic CT prostate radiotherapy CTV cohort<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1afairness-aware segmentation\uff1bdistribution-aware mixture-of-experts\uff1bsubgroup-conditioned KL-DRO \/ robust loss aggregation<\/li>\n<\/ul>\n<h3>paper-deep-reader \u7cbe\u8bfb\u7ed3\u679c<\/h3>\n<h4>1. \u4e00\u53e5\u8bdd\u7ed3\u8bba<\/h4>\n<p>\u8fd9\u7bc7\u8bba\u6587\u6700\u91cd\u8981\u7684\u4ef7\u503c\uff0c\u662f\u628a\u533b\u5b66\u5206\u5272\u516c\u5e73\u6027\u4ece\u7ec4\u95f4\u5e73\u5747\u6307\u6807\u5dee\u8ddd\u63a8\u8fdb\u5230\u7ec4\u95f4\u9002\u914d + \u7ec4\u5185\u56f0\u96be\u60a3\u8005\u4e0d\u88ab\u5e73\u5747\u503c\u63a9\u76d6\u7684\u53cc\u8f74\u95ee\u9898\uff0c\u5e76\u7528 dMoE + subgroup-conditioned DRO \u7ed9\u51fa\u4e00\u4e2a\u53ef\u5b9e\u73b0\u6846\u67b6\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6b63\u5728\u8fdb\u5165\u7b5b\u67e5\u3001\u8bca\u65ad\u3001\u653e\u7597\u8ba1\u5212\u7b49\u4e34\u5e8a\u73af\u8282\uff0c\u6a21\u578b\u5728\u4e0d\u540c\u4eba\u53e3\u5b66\u3001\u4e34\u5e8a\u6216\u673a\u6784\u5b50\u7fa4\u4e0a\u7684\u8868\u73b0\u5dee\u5f02\u4f1a\u5bfc\u81f4\u5b9e\u9645\u98ce\u9669\u3002\u4f8b\u5982 glaucoma screening \u4e2d optic cup\/rim \u8fb9\u754c\u9519\u8bef\u53ef\u80fd\u5f71\u54cd\u5224\u65ad\uff0cradiotherapy target segmentation \u9519\u8bef\u53ef\u80fd\u6f0f\u7167\u75c5\u7076\u6216\u7167\u5c04\u5065\u5eb7\u7ec4\u7ec7\u3002\u5df2\u6709\u516c\u5e73\u6027\u5de5\u4f5c\u5f80\u5f80\u6309 race\u3001age\u3001sex\u3001tumor stage\u3001institution \u7b49\u9884\u5b9a\u4e49 subgroup \u8ba1\u7b97\u5e73\u5747\u6027\u80fd\u5e76\u7f29\u5c0f\u7ec4\u95f4\u5dee\u8ddd\uff1b\u4f46\u540c\u4e00 subgroup \u5185\u4e5f\u6709\u56fe\u50cf\u8d28\u91cf\u3001\u89e3\u5256\u5dee\u5f02\u3001\u75c5\u7076\u8303\u56f4\u3001\u8fb9\u754c\u6a21\u7cca\u7b49\u5dee\u5f02\uff0c\u5e73\u5747\u503c\u4f1a\u63a9\u76d6\u5c11\u6570\u9ad8\u635f\u5931\u60a3\u8005\u3002<\/p>\n<p>paper map\uff1a\u8bba\u6587\u7814\u7a76\u516c\u5e73\u533b\u5b66\u5206\u5272\u4e2d subgroup \u5e73\u5747\u6027\u80fd\u63a9\u76d6\u7ec4\u5185\u56f0\u96be\u6837\u672c\u7684\u95ee\u9898\uff1b\u4e3b\u7ebf\u63d0\u51fa DuetFair \u53cc\u8f74\u673a\u5236\uff0c\u5e76\u5b9e\u4f8b\u5316\u4e3a FairDRO\uff1a\u7528 dMoE \u5904\u7406 inter-subgroup representation adaptation\uff0c\u7528 subgroup-conditioned KL-DRO \u5904\u7406 intra-subgroup hard sample robustness\uff1b\u58f0\u79f0\u5728 Harvard-FairSeg\u3001HAM10000 \u548c 3D \u653e\u7597 CTV \u5206\u5272\u4e2d\u63d0\u5347 equity-scaled metrics \u548c worst subgroup performance\uff1b\u5173\u952e\u6280\u672f\u5bf9\u8c61\u662f subgroup attribute\u3001ERM subgroup risk\u3001dMoE routing\u3001robust risk\u3001KL ambiguity set\u3001ES-Dice \/ ES-IoU\uff1b\u4e3b\u8981\u98ce\u9669\u662f\u65b9\u6cd5\u4f9d\u8d56\u8bad\u7ec3\u548c\u63a8\u7406\u65f6\u53ef\u7528\u4e14\u53ef\u9760\u7684 subgroup \u5c5e\u6027\uff0c\u4e14\u5728 HAM10000 \u8fd9\u7c7b\u4e3b\u8981\u7531\u7ec4\u95f4\u5dee\u5f02\u89e3\u91ca\u7684\u4efb\u52a1\u4e0a\u4f18\u52bf\u4e0d\u7a33\u5b9a\u3002<\/p>\n<p>\u8def\u7ebf\u8bb0\u5f55\uff1aPrimary adapter = method-algorithm\uff1bSecondary adapter = benchmark-evaluation\uff1bEvidence packs = general\u3001experimental-eval\u3001robustness-and-ood\u3001benchmark-fairness-and-contamination\uff1bRoute confidence = \u9ad8\u3002\u539f\u56e0\uff1a\u8bba\u6587\u8d21\u732e\u662f\u4e00\u4e2a\u516c\u5e73\u6027\u8bad\u7ec3\u76ee\u6807\u548c\u67b6\u6784\u7ec4\u5408\uff0c\u5e76\u7528\u591a benchmark \u516c\u5e73\u6027\u6307\u6807\u652f\u6491\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u8ba4\u4e3a\u5df2\u6709 FairSeg\u3001FairDiff\u3001FairDomain\u3001dMoE \u7b49\u533b\u5b66\u5206\u5272\u516c\u5e73\u6027\u65b9\u6cd5\u4e3b\u8981\u5173\u6ce8 inter-subgroup fairness\uff1a\u6309 race\u3001age\u3001domain \u6216 clinical subgroup \u8861\u91cf\u548c\u6539\u5584\u7fa4\u4f53\u5e73\u5747\u6027\u80fd\u3002\u4f46 subgroup \u5185\u90e8\u5e76\u4e0d\u5747\u8d28\uff0c\u56f0\u96be\u75c5\u4f8b\u53ef\u80fd\u88ab subgroup mean \u6df9\u6ca1\u3002\u53e6\u4e00\u65b9\u9762\uff0cCVaR\u3001tilted ERM\u3001focal loss\u3001OHEM \u7b49 sample-wise robust learning \u4f1a\u5173\u6ce8\u56f0\u96be\u6837\u672c\uff0c\u4f46\u901a\u5e38\u5728\u5168\u5c40\u6837\u672c\u6c60\u4e2d\u5b9a\u4e49 hard cases\uff0c\u5c0f subgroup \u5185\u90e8\u7684\u56f0\u96be\u6837\u672c\u4ecd\u53ef\u80fd\u56e0\u4e3a\u6570\u91cf\u5c11\u800c\u5f71\u54cd\u4e0d\u8db3\u3002\u8bba\u6587\u7684\u6838\u5fc3\u6279\u8bc4\u662f\uff1a\u5355\u72ec\u6309 group \u6216\u5355\u72ec\u6309 sample \u90fd\u4e0d\u591f\uff0c\u516c\u5e73\u533b\u5b66\u5206\u5272\u9700\u8981\u540c\u65f6\u5904\u7406\u7ec4\u95f4\u5f02\u8d28\u6027\u548c\u7ec4\u5185\u56f0\u96be\u6837\u672c\u3002<\/p>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>FairDRO \u4ece\u6807\u51c6 ERM \u5199\u8d77\uff1a\u6807\u51c6 ERM \u53ef\u5199\u6210\u5404 subgroup \u98ce\u9669\u7684\u52a0\u6743\u5e73\u5747\u3002\u8fd9\u4e2a\u5f0f\u5b50\u66b4\u9732\u4e24\u5c42\u5e73\u5747\uff1a\u7ec4\u95f4\u7531 subgroup frequency \u51b3\u5b9a\uff0c\u5c0f\u7ec4\u5f71\u54cd\u53ef\u80fd\u5c0f\uff1b\u7ec4\u5185 subgroup risk \u662f\u5e73\u5747\u98ce\u9669\uff0c\u9ad8\u635f\u5931\u6837\u672c\u53ef\u80fd\u88ab\u4f4e\u635f\u5931\u6837\u672c\u7a00\u91ca\u3002\u4f5c\u8005\u6307\u51fa\uff0c\u7b80\u5355\u4e24\u5c42 reweighting \u5728\u4e00\u9636\u68af\u5ea6\u91cc\u53ea\u4f53\u73b0\u4e3a subgroup weight \u4e0e sample weight \u7684\u4e58\u79ef\uff0c\u65e0\u6cd5\u6e05\u695a\u533a\u5206\u201c\u5f3a\u8c03 subgroup\u201d\u8fd8\u662f\u201c\u5f3a\u8c03 subgroup \u5185\u56f0\u96be\u6837\u672c\u201d\u3002\u56e0\u6b64 FairDRO \u628a\u4e24\u4ef6\u4e8b\u62c6\u5f00\uff1a\u7528 dMoE \u5728 representation \u5c42\u5904\u7406\u7ec4\u95f4\u5dee\u5f02\uff1b\u7528 subgroup-conditioned DRO \u5728 loss aggregation \u5c42\u5904\u7406\u7ec4\u5185\u56f0\u96be\u6837\u672c\u3002\u6700\u7ec8\u76ee\u6807\u662f\u5747\u5300\u805a\u5408\u5404 subgroup \u7684 robust risk\uff0c\u4e0d\u624b\u52a8\u504f\u7f6e\u67d0\u4e2a subgroup\u3002<\/p>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<ol>\n<li>\n<p><strong>DuetFair \u673a\u5236<\/strong>\uff1a\u8fd9\u4e0d\u662f\u4e00\u4e2a\u5177\u4f53 layer\uff0c\u800c\u662f\u95ee\u9898\u5206\u89e3\uff1ainter-subgroup axis \u5173\u6ce8\u4e0d\u540c subgroup \u7684\u5206\u5e03\u548c\u5f62\u6001\u5dee\u5f02\uff1bintra-subgroup axis \u5173\u6ce8\u540c\u4e00 subgroup \u5185 hard patients\u3002\u8fd9\u4e2a\u89c6\u89d2\u5f88\u6709\u5199\u4f5c\u4ef7\u503c\uff0c\u56e0\u4e3a\u5b83\u80fd\u89e3\u91ca\u4e3a\u4ec0\u4e48\u5355\u770b subgroup mean \u4e0d\u8db3\u4ee5\u4fdd\u8bc1\u60a3\u8005\u7ea7\u53ef\u9760\u6027\u3002<\/p>\n<\/li>\n<li>\n<p><strong>dMoE \u8868\u5f81\u9002\u914d<\/strong>\uff1a\u7ed9\u5b9a\u6837\u672c\u548c subgroup attribute\uff0c\u5728 backbone \u4e2d\u95f4\u5c42\u53d6\u7279\u5f81\uff0c\u901a\u8fc7 subgroup-conditioned top-K expert routing \u6dfb\u52a0\u6b8b\u5dee\u4e13\u5bb6\u53d8\u6362\uff0c\u7136\u540e decoder \u8f93\u51fa mask\u3002dMoE \u7684\u4f5c\u7528\u662f\u628a\u4e0d\u540c race \/ age \/ tumor stage \/ institution \u7684\u5206\u5e03\u5dee\u5f02\u653e\u5728 feature path \u4e2d\u5904\u7406\uff0c\u800c\u4e0d\u662f\u53ea\u9760 loss \u6743\u91cd\u3002<\/p>\n<\/li>\n<li>\n<p><strong>Subgroup-conditioned DRO<\/strong>\uff1a\u5bf9\u6bcf\u4e2a subgroup \u5355\u72ec\u5b9a\u4e49 KL ambiguity set\u3002robust risk \u662f\u8be5 subgroup \u5185\u6240\u6709\u8fd1\u90bb\u5206\u5e03\u4e0a\u7684 worst-case expected loss\u3002\u6709\u9650\u6837\u672c\u4e0b\uff0c\u9644\u5f55\u7ed9\u51fa log-sum-exp \u5f62\u5f0f\u548c\u6307\u6570\u503e\u659c\u6743\u91cd\uff1b\u8fd9\u610f\u5473\u7740\u9ad8\u635f\u5931\u6837\u672c\u83b7\u5f97\u66f4\u5927\u68af\u5ea6\u6743\u91cd\uff0c\u4f46\u8be5\u6743\u91cd\u662f\u5728 subgroup \u5185\u90e8\u5f52\u4e00\u5316\u7684\uff0c\u6240\u4ee5\u5c0f subgroup \u7684\u56f0\u96be\u75c5\u4f8b\u4e0d\u4f1a\u88ab\u5168\u5c40\u5927 subgroup \u6837\u672c\u538b\u6389\u3002<\/p>\n<\/li>\n<li>\n<p><strong>Equity-scaled metrics<\/strong>\uff1a\u8bba\u6587\u4f7f\u7528 ES-Dice \/ ES-IoU\uff0c\u628a\u603b\u4f53\u6307\u6807\u6309 subgroup \u6307\u6807\u504f\u79bb\u603b\u4f53\u503c\u7684\u7edd\u5bf9\u5dee\u8fdb\u884c down-scale\u3002\u8fd9\u6bd4\u53ea\u62a5\u5e73\u5747 Dice \u66f4\u80fd\u53cd\u6620 subgroup disparity\u3002<\/p>\n<\/li>\n<\/ol>\n<p>\u6a21\u5757\u8fc1\u79fb\u6027\uff1adMoE routing \u53ef\u5d4c\u5165 TransUNet\u30013D U-Net\u3001\u751a\u81f3 Mamba\/VMamba encoder \u7684\u4e2d\u95f4\u5c42\uff1bDRO loss \u662f architecture-agnostic\u3002\u5bf9 polyp segmentation\uff0c\u82e5\u6ca1\u6709\u660e\u786e demographic \/ institution subgroup\uff0c\u65b9\u6cd5\u4e0d\u80fd\u76f4\u63a5\u7528\uff1b\u4f46\u53ef\u628a\u6570\u636e\u96c6\u6765\u6e90\uff08Kvasir\u3001ClinicDB\u3001ETIS\u3001CVC-ColonDB\uff09\u6216\u6210\u50cf\u4e2d\u5fc3\u89c6\u4e3a subgroup\uff0c\u7814\u7a76\u8de8\u6570\u636e\u96c6\u516c\u5e73 \/ \u6cdb\u5316\u3002<\/p>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p>\u5b9e\u9a8c\u8986\u76d6\u4e09\u4e2a benchmark\uff1aHarvard-FairSeg\uff0c10,000 \u5f20 SLO fundus \u56fe\u50cf\uff0coptic cup \/ neuroretinal rim mask\uff0crace \u5206\u4e3a Black \/ Asian \/ White\uff0c\u6d4b\u8bd5\u96c6 2,000\uff0c\u4f7f\u7528 TransUNet ViT-B\uff1bHAM10000\uff0c10,015 \u5f20 dermoscopy \u56fe\u50cf\uff0cbinary lesion mask\uff0cage \u5206 5 \u7ec4\uff0c\u6d4b\u8bd5 1,061\uff0c\u540c\u6837\u4f7f\u7528 TransUNet ViT-B\uff1bRadiotherapy Target Dataset\uff0cin-house pelvic CT prostate cancer CTV \u5206\u5272\uff0cT-stage subgroup \u548c institution subgroup \u4e24\u79cd\u8bbe\u7f6e\uff0c\u4f7f\u7528 3D Residual U-Net\uff0cpatch 384x384x128\uff0cbatch size 4\uff0csliding-window inference\u3002<\/p>\n<p>\u4e3b\u8981\u7ed3\u679c\uff1aHarvard-FairSeg race \/ rim \u4e0a FairDRO ES-Dice 0.755\u3001Dice 0.816\u3001ES-IoU 0.639\u3001IoU 0.701\uff1bAsian subgroup Dice \u4ece dMoE \u7684 0.769 \u63d0\u5347\u5230 0.780\u3002cup \u5206\u5272\u4e2d Asian Dice \u4ece 0.844 \u63d0\u5347\u5230 0.852\u3002HAM10000 age \u4e0a FairDRO \u5bf9 80 \u5c81\u4ee5\u4e0a minority group \u76f8\u6bd4 dMoE \u6709\u63d0\u5347\uff08Dice 0.864 \u5230 0.876\uff0cIoU 0.791 \u5230 0.840\uff09\uff0c\u4f46\u6574\u4f53\u6307\u6807\u4e0d\u4f18\u4e8e dMoE\uff0c\u4f5c\u8005\u627f\u8ba4\u5f53\u5dee\u5f02\u4e3b\u8981\u662f\u7ec4\u95f4\u65f6 GDRO \u5df2\u7ecf\u5f88\u5f3a\uff0cFairDRO \u989d\u5916\u6536\u76ca\u6709\u9650\u30023D radiotherapy \/ T-stage \u4e0a FairDRO ES-Dice 0.530\u3001Dice 0.665\u3001ES-IoU 0.411\u3001IoU 0.521\uff1b\u76f8\u5bf9 dMoE\uff0cworst-group Dice \u4ece T2 \u7684 0.585 \u63d0\u5347\u5230 0.620\u30023D radiotherapy \/ institution \u4e0a FairDRO ES-Dice 0.589\u3001Dice 0.688\u3001ES-IoU 0.471\u3001IoU 0.554\uff1b\u6700\u96be GN Dice \u4ece 3D ResUNet \u7684 0.551 \/ dMoE \u7684 0.547 \u63d0\u5347\u5230 0.592\uff0cSC-GN gap \u4ece\u7ea6 0.18 \u7f29\u5c0f\u5230\u7ea6 0.14\u3002\u6d88\u878d Table 6 \u663e\u793a dMoE-ERM\u3001SG-DRO\u3001FairDRO \u4e09\u8005\u5bf9\u6bd4\u4e2d\u7ec4\u5408\u6700\u597d\uff0c\u652f\u6301\u53cc\u8f74\u8bbe\u8ba1\u3002<\/p>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u53ef\u4fe1\u70b9\uff1a\u8bba\u6587\u5305\u542b\u516c\u5f00 2D fairness benchmark\u3001\u516c\u5f00 skin lesion benchmark \u548c\u4e34\u5e8a 3D CT \u653e\u7597\u4efb\u52a1\uff1b\u62a5\u544a\u4e86 subgroup-wise metrics\u3001worst subgroup\u3001ES metrics \u548c 1000 bootstrap 95% CI\uff1b\u6709 dMoE\u3001GDRO\u3001Prompt-GDRO\u3001MoE\u3001FEBS\u3001FairDiff \u7b49\u6bd4\u8f83\uff1b\u9644\u5f55\u7ed9\u51fa\u4e24\u8f74\u6d88\u878d\u3001KL-DRO \u6743\u91cd\u89e3\u91ca\u548c\u6570\u636e\u5206\u5e03\u3002<\/p>\n<p>\u9650\u5236\uff1a\u7b2c\u4e00\uff0c\u5173\u952e 3D radiotherapy \u6570\u636e\u662f in-house\uff0c\u65e0\u6cd5\u72ec\u7acb\u590d\u73b0\u5168\u90e8\u7ed3\u8bba\u3002\u7b2c\u4e8c\uff0cFairDRO \u4f9d\u8d56\u63a8\u7406\u65f6 subgroup attribute \u53ef\u7528\u4e14\u53ef\u9760\uff1b\u5728\u5b9e\u9645 polyp \u6570\u636e\u4e2d\uff0c\u4e2d\u5fc3\u3001\u8bbe\u5907\u3001\u60a3\u8005\u5c5e\u6027\u672a\u5fc5\u53ef\u7528\u6216\u6807\u6ce8\u4e00\u81f4\u3002\u7b2c\u4e09\uff0cHAM10000 \u4e0a FairDRO \u4e0d\u603b\u662f\u4f18\u4e8e GDRO \/ dMoE\uff0c\u8bf4\u660e\u65b9\u6cd5\u4f18\u52bf\u4f9d\u8d56\u201c\u7ec4\u5185\u5f02\u8d28\u6027\u5f3a\u201d\u8fd9\u4e00\u6761\u4ef6\u3002\u7b2c\u56db\uff0c\u4ee3\u7801\u94fe\u63a5\u672a\u63d0\u4f9b\u3002\u7b2c\u4e94\uff0c\u8bba\u6587\u5173\u6ce8 fairness \/ robustness\uff0c\u4e0d\u662f\u63d0\u9ad8\u7edd\u5bf9 Dice \u7684\u901a\u7528\u65b0\u5206\u5272\u4e3b\u5e72\uff1b\u5982\u679c\u53ea\u8ffd\u6c42 leaderboard \u5e73\u5747 Dice\uff0c\u4ef7\u503c\u53ef\u80fd\u4f4e\u4e8e\u4e13\u95e8\u7684 U-Net \/ Transformer \/ Mamba \u7ed3\u6784\u8bba\u6587\u3002<\/p>\n<h4>8. \u4e0e\u4e3b\u6d41\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u7684\u5173\u7cfb<\/h4>\n<p>\u4e0e U-Net \/ 3D U-Net\uff1aradiotherapy \u5b9e\u9a8c\u76f4\u63a5\u4f7f\u7528 3D Residual U-Net\uff0c\u8bf4\u660e FairDRO \u53ef\u4f5c\u4e3a U-Net \u7c7b\u6a21\u578b\u7684\u8bad\u7ec3\u76ee\u6807\u548c\u4e2d\u95f4\u5c42 routing \u6269\u5c55\u3002\u4e0e nnU-Net\uff1a\u8bba\u6587\u672a\u9a8c\u8bc1 nnU-Net\uff0c\u4f46\u4ece\u65b9\u6cd5\u4e0a\u53ef\u628a subgroup-conditioned experts \u653e\u5230 encoder \/ bottleneck\uff0c\u5e76\u7528 DRO loss \u66ff\u4ee3\u666e\u901a Dice\/CE \u7684\u6837\u672c\u805a\u5408\u3002\u4e0e TransUNet \/ Swin-UNet \/ UNetR\uff1a2D \u5b9e\u9a8c\u4f7f\u7528 TransUNet ViT-B\uff0c\u8bf4\u660e\u8be5\u601d\u8def\u517c\u5bb9 Transformer encoder\u3002\u4e0e Mamba \/ VMamba \/ SegMamba \/ DAMamba\uff1a\u8bba\u6587\u6ca1\u6709 Mamba \u6a21\u5757\uff1b\u4f46 dMoE + subgroup-conditioned DRO \u53ef\u4f5c\u4e3a Mamba-based segmentation \u7684\u516c\u5e73\u6027\u8bad\u7ec3\u6269\u5c55\uff0c\u5c24\u5176\u9002\u5408\u591a\u4e2d\u5fc3 3D \u533b\u5b66\u5206\u5272\u3002\u4e0e foundation model\uff1a\u5f15\u7528 SAM \/ FairSeg \u80cc\u666f\uff0c\u4f46 FairDRO \u672c\u8eab\u4e0d\u4f9d\u8d56 SAM\uff1b\u53ef\u7528\u4e8e MedSAM \/ SAM-Med3D \u9002\u914d\u540e\u7684 subgroup robustness fine-tuning\u3002<\/p>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u5bf9 polyp segmentation\uff1a\u4e2d\u7b49\u3002\u666e\u901a\u606f\u8089\u5206\u5272\u6570\u636e\u96c6\u6ca1\u6709\u660e\u786e race \/ age \/ institution \u4fe1\u606f\uff0c\u4f46\u53ef\u4ee5\u628a\u6570\u636e\u96c6\u6765\u6e90\u3001\u5185\u955c\u4e2d\u5fc3\u3001\u8bbe\u5907\u7c7b\u578b\u6216\u56f0\u96be\u6837\u672c\u7c7b\u522b\u4f5c\u4e3a subgroup\uff0c\u501f\u9274 intra-subgroup DRO \u505a\u8de8\u6570\u636e\u96c6\u6cdb\u5316\u548c\u516c\u5e73\u6027\u5206\u6790\u3002\u5bf9 DAMamba \u6539\u9020\uff1a\u65b9\u6cd5\u4e0d\u662f\u4e3b\u5e72\u6a21\u5757\uff0c\u4f46\u53ef\u4f5c\u4e3a DAMamba \u5728\u591a\u4e2d\u5fc3 \/ \u591a\u6570\u636e\u96c6 polyp segmentation \u4e0a\u7684\u8bad\u7ec3\u76ee\u6807\uff0c\u5f3a\u8c03\u6bcf\u4e2a\u6570\u636e\u57df\u5185\u90e8 hard cases\uff0c\u800c\u4e0d\u662f\u53ea\u4f18\u5316\u5e73\u5747 Dice\u3002\u5bf9 related work\uff1a\u5f88\u9002\u5408\u7528\u4e8e\u53ef\u4fe1\u533b\u5b66\u5206\u5272\u3001fairness-aware segmentation\u3001clinical deployment \u6bb5\u843d\u3002\u5bf9\u590d\u73b0\u5b9e\u9a8c\uff1a\u5982\u679c\u5df2\u6709 subgroup metadata\uff0cFairDRO \u6bd4\u8f83\u503c\u5f97\u5b9e\u73b0\uff1b\u6ca1\u6709 metadata \u65f6\u9700\u5148\u505a latent subgroup discovery \u6216\u6309 dataset\/domain \u8fd1\u4f3c\u3002<\/p>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5efa\u8bae\u7cbe\u8bfb<\/strong>\uff0c\u4f46\u9605\u8bfb\u76ee\u7684\u5e94\u653e\u5728\u516c\u5e73\u6027\u4e0e\u9c81\u68d2\u8bad\u7ec3\uff0c\u4e0d\u8981\u628a\u5b83\u5f53\u4f5c\u65b0 backbone \u8bba\u6587\u3002\u5efa\u8bae\u4f18\u5148\u8bfb Introduction\u3001Method\u3001Table 3\/4\u3001Appendix A.2 \u548c A.5\uff1b\u5982\u679c\u8fd1\u671f\u8bba\u6587\u53ea\u805a\u7126 polyp segmentation \u5e73\u5747\u7cbe\u5ea6\uff0c\u53ef\u7565\u8bfb\u5b9e\u9a8c\u90e8\u5206\u5373\u53ef\u3002<\/p>\n<hr \/>\n<h2>\u4eca\u65e5\u63a8\u8350\u4f18\u5148\u7ea7<\/h2>\n<ol>\n<li><strong>DuetFair \/ FairDRO<\/strong>\uff1a\u66f4\u503c\u5f97\u4ece\u7814\u7a76\u89c6\u89d2\u6df1\u5165\u8bfb\uff0c\u56e0\u4e3a\u5b83\u63d0\u51fa\u7684 inter-subgroup + intra-subgroup \u53cc\u8f74\u6846\u67b6\u6709\u8f83\u5f3a\u6982\u5ff5\u4ef7\u503c\uff0c\u9002\u5408\u5199\u5165\u53ef\u4fe1\u533b\u5b66\u5206\u5272\u3001\u8de8\u4e2d\u5fc3\u6cdb\u5316\u3001\u516c\u5e73\u6027\u548c clinical deployment \u76f8\u5173\u7ae0\u8282\uff0c\u4e5f\u80fd\u542f\u53d1 DAMamba \u5728\u591a\u57df\u6570\u636e\u4e0a\u7684\u8bad\u7ec3\u76ee\u6807\u8bbe\u8ba1\u3002<\/li>\n<li><strong>SplitFed-CL<\/strong>\uff1a\u66f4\u9002\u5408\u5de5\u7a0b\u548c\u5b9e\u9a8c\u8bbe\u8ba1\u53c2\u8003\uff0c\u7279\u522b\u662f noisy label\u3001\u5f31\u6807\u6ce8\u3001\u591a\u4e2d\u5fc3\u9690\u79c1\u8bad\u7ec3\u548c boundary-centric annotation error simulation\uff1b\u5982\u679c\u8981\u505a polyp noisy-label robustness\uff0c\u5b83\u7684\u53ef\u8fc1\u79fb\u6027\u5f88\u5f3a\u3002<\/li>\n<\/ol>\n<h2>\u4eca\u65e5 PDF \u83b7\u53d6\u60c5\u51b5<\/h2>\n<ul>\n<li>\u8bba\u6587 1\uff1a\u5df2\u9644 PDF\uff1b\u672c\u5730\u6587\u4ef6 <code>MEDIA:\/tmp\/medseg_daily_2026-05-17\/2605.11060.pdf<\/code>\uff0c\u53ef\u8bbf\u95ee PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.11060<\/li>\n<li>\u8bba\u6587 2\uff1a\u5df2\u9644 PDF\uff1b\u672c\u5730\u6587\u4ef6 <code>MEDIA:\/tmp\/medseg_daily_2026-05-17\/2605.10521.pdf<\/code>\uff0c\u53ef\u8bbf\u95ee PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.10521<\/li>\n<\/ul>\n<h2>\u4eca\u65e5\u53ef\u6267\u884c\u5efa\u8bae<\/h2>\n<ol>\n<li>\u5982\u679c\u4eca\u5929\u53ea\u80fd\u8bfb\u4e00\u7bc7\uff0c\u5148\u8bfb <strong>DuetFair<\/strong> \u7684\u65b9\u6cd5\u548c 3D radiotherapy \u5b9e\u9a8c\uff1a\u5b83\u5bf9\u5982\u4f55\u8bc1\u660e\u6a21\u578b\u4e0d\u4ec5\u5e73\u5747 Dice \u9ad8\uff0c\u800c\u4e14\u5bf9\u4e0d\u540c subgroup \/ hard patients \u66f4\u53ef\u9760\u5f88\u6709\u542f\u53d1\u3002<\/li>\n<li>\u5982\u679c\u51c6\u5907\u505a polyp segmentation \u7684\u5f31\u6807\u6ce8\u6216 noisy-label \u5b9e\u9a8c\uff0c\u4f18\u5148\u590d\u73b0 <strong>SplitFed-CL<\/strong> \u7684 difficulty-guided boundary deformation\uff1b\u8fd9\u4e2a\u566a\u58f0\u751f\u6210\u6bd4\u968f\u673a erosion\/dilation \u66f4\u50cf\u606f\u8089\u8fb9\u754c\u6a21\u7cca\u548c\u6807\u6ce8\u5206\u6b67\u3002<\/li>\n<li>\u5bf9 DAMamba \/ Mamba-based segmentation \u540e\u7eed\u5de5\u4f5c\uff0c\u53ef\u4ee5\u628a\u4e24\u7bc7\u5408\u5e76\u6210\u4e00\u4e2a\u5b9e\u9a8c\u65b9\u5411\uff1a\u7528 DAMamba \u4f5c\u4e3a backbone\uff0c\u518d\u52a0\u5165 SplitFed-CL \u7684 label correction \u6216 FairDRO \u7684 subgroup-conditioned robust loss\uff0c\u9a8c\u8bc1\u5728\u8de8\u6570\u636e\u96c6 polyp segmentation \u4e0a\u662f\u5426\u964d\u4f4e worst-domain failure\u3002<\/li>\n<\/ol>\n<h2>\u53c2\u8003\u94fe\u63a5<\/h2>\n<ul>\n<li>SplitFed-CL arXiv\uff1ahttps:\/\/arxiv.org\/abs\/2605.11060<\/li>\n<li>SplitFed-CL PDF\uff1ahttps:\/\/arxiv.org\/pdf\/2605.11060<\/li>\n<li>DuetFair arXiv\uff1ahttps:\/\/arxiv.org\/abs\/2605.10521<\/li>\n<li>DuetFair PDF\uff1ahttps:\/\/arxiv.org\/pdf\/2605.10521<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u4eca\u65e5\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6700\u65b0\u8bba\u6587\u7cbe\u8bfb\u8ffd\u8e2a \u4eca\u65e5\u7ed3\u8bba \u4eca\u5929\u68c0\u7d22\u5230\u7684\u6700\u65b0\u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\u4e2d\uff0c\u65b9\u6cd5\u521b\u65b0\u6700\u503c\u5f97\u5173\u6ce8\u7684\u4e0d\u662f\u53c8\u4e00\u4e2a U-Net \u6a21\u5757\u5806 &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"emotion":"","emotion_color":"","title_style":"","license":"","footnotes":""},"categories":[85],"tags":[],"class_list":["post-1064","post","type-post","status-publish","format-standard","hentry","category-85"],"views":6,"_links":{"self":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1064","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/comments?post=1064"}],"version-history":[{"count":0,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1064\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/media?parent=1064"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/categories?post=1064"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/tags?post=1064"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}