{"id":1073,"date":"2026-05-23T08:37:03","date_gmt":"2026-05-23T00:37:03","guid":{"rendered":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/23\/2026-05-23-%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%9adepthpolyp-%e4%b8%8e-patch-moe-mamba\/"},"modified":"2026-05-23T08:37:03","modified_gmt":"2026-05-23T00:37:03","slug":"2026-05-23-%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%9adepthpolyp-%e4%b8%8e-patch-moe-mamba","status":"publish","type":"post","link":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/23\/2026-05-23-%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%9adepthpolyp-%e4%b8%8e-patch-moe-mamba\/","title":{"rendered":"2026-05-23 \u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\u7cbe\u8bfb\uff1aDepthPolyp \u4e0e Patch-MoE Mamba"},"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\u4ece 2026 \u5e74 5 \u6708\u4e2d\u4e0b\u65ec\u7684 arXiv \u533b\u5b66\u56fe\u50cf\u5206\u5272\u65b0\u7a3f\u4e2d\u7b5b\u5230 2 \u7bc7\u503c\u5f97\u5173\u6ce8\u7684\u8bba\u6587\uff1a\u4e00\u7bc7\u662f Mamba\/VM-UNet \u7cfb\u5217\u7684\u7ed3\u6784\u6539\u9020\uff08Patch-MoE Mamba\uff09\uff0c\u53e6\u4e00\u7bc7\u662f\u9762\u5411\u771f\u5b9e\u5185\u955c\u9000\u5316\u4e0e\u79fb\u52a8\u7aef\u90e8\u7f72\u7684\u8f7b\u91cf\u7ea7\u606f\u8089\u5206\u5272\u6846\u67b6\uff08DepthPolyp\uff09\u3002\u6574\u4f53\u8d8b\u52bf\u5f88\u660e\u786e\uff1a\u533b\u5b66\u56fe\u50cf\u5206\u5272\u65b9\u6cd5\u4e0d\u518d\u53ea\u8ffd\u6c42 clean benchmark \u4e0a\u7684 Dice\uff0c\u800c\u662f\u5728\u5f80\u201c\u72b6\u6001\u7a7a\u95f4\/\u957f\u7a0b\u4f9d\u8d56 + \u5c40\u90e8\u7ed3\u6784\u4fdd\u6301\u201d\u201c\u771f\u5b9e\u9000\u5316\u9c81\u68d2\u6027 + \u4f4e\u7b97\u529b\u90e8\u7f72\u201d\u4e24\u4e2a\u65b9\u5411\u63a8\u8fdb\u3002<\/p>\n<h2>\u68c0\u7d22\u8bf4\u660e<\/h2>\n<p>\u68c0\u7d22\u8303\u56f4\u8986\u76d6 arXiv 2025 \u5e74\u4ee5\u6765\u4e0e medical image segmentation\u3001Mamba\u3001U-Net\u3001universal medical image segmentation\u3001polyp segmentation \u76f8\u5173\u7684\u6700\u65b0\u8bba\u6587\uff0c\u5e76\u4f18\u5148\u68c0\u67e5\u8fd1\u671f 2026-05 \u9644\u8fd1\u7684\u53ef\u516c\u5f00\u4e0b\u8f7d PDF\u3002\u4eca\u5929\u5165\u9009\u7684\u4e24\u7bc7\u8bba\u6587\u5747\u4e3a 2025 \u5e74\u4ee5\u540e\u8bba\u6587\uff0c\u4e14\u76ee\u524d\u5747\u4e3a arXiv preprint\uff1b\u672a\u53d1\u73b0\u5b83\u4eec\u5df2\u7ecf\u6709 MICCAI\/CVPR\/TMI\/MedIA \u7b49\u6b63\u5f0f\u63a5\u6536\u4fe1\u606f\u3002\u5df2\u68c0\u67e5\u672c\u5730\u5386\u53f2\u63a8\u8350\u8bb0\u5f55\u4e0e\u53ef\u7528 cron\/session \u6587\u4ef6\u8def\u5f84\uff0c\u4f46\u5f53\u524d\u73af\u5883\u672a\u68c0\u51fa\u65e2\u5f80\u8f93\u51fa\u8bb0\u5f55\uff1b\u56e0\u6b64\u6309\u6807\u9898\u3001arXiv ID \u4e0e PDF \u94fe\u63a5\u5bf9\u672c\u6b21\u5019\u9009\u505a\u53bb\u91cd\uff0c\u6392\u9664\u4e86\u91cd\u590d\u8bba\u6587\u3002\u5019\u9009\u4e2d\u8fd8\u68c0\u67e5\u8fc7 USEMA\u3001Deco-Mamba\u3001CT body composition class-imbalance sampling \u7b49\u8bba\u6587\uff1b\u6700\u7ec8\u4f18\u5148\u9009\u62e9\u4e0e\u7528\u6237\u5173\u6ce8\u7684 polyp segmentation\u3001Mamba\/U-Net \u6539\u9020\u3001\u8f7b\u91cf\u90e8\u7f72\u548c\u53ef\u590d\u73b0\u5b9e\u9a8c\u66f4\u76f8\u5173\u7684\u4e24\u7bc7\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\uff1aPatch-MoE Mamba: A Patch-Ordered Mixture-of-Experts State Space Architecture for Medical Image Segmentation<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aPatch-MoE Mamba: A Patch-Ordered Mixture-of-Experts State Space Architecture for Medical Image Segmentation<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aDiego Adame et al.<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-18<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.17719<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1aMEDIA:\/root\/papers_medseg_2026-05-23\/2605.17719.pdf \uff1bhttps:\/\/arxiv.org\/pdf\/2605.17719<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1a\u672a\u83b7\u53d6 \/ \u8bba\u6587\u6b63\u6587\u672a\u7ed9\u51fa\u660e\u786e\u5b98\u65b9\u4ee3\u7801\u94fe\u63a5<\/li>\n<li>\u4efb\u52a1\uff1a2D \u533b\u5b66\u56fe\u50cf\u5206\u5272\uff1b\u4e3b\u8981\u662f polyp segmentation\uff0c\u9644\u52a0 skin lesion segmentation<\/li>\n<li>\u6570\u636e\u96c6\uff1aKvasir-SEG\u3001CVC-ClinicDB\u3001ColonDB\u3001ETIS\u3001CVC-300\uff1bISIC 2017\u3001ISIC 2018<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1aU-Net-style \/ VM-UNetV2-based \/ Mamba-based segmentation \/ patch-ordered scanning \/ MoE directional fusion<\/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\u4e0d\u662f\u201c\u53c8\u4e00\u4e2a Mamba-UNet\u201d\uff0c\u800c\u662f\u628a Vision Mamba \u4e2d\u5bb9\u6613\u7834\u574f 2D \u5c40\u90e8\u90bb\u57df\u7684\u626b\u63cf\u987a\u5e8f\u663e\u5f0f\u6539\u6210 patch-ordered scanning\uff0c\u5e76\u7528 MoE \u8ba9\u591a\u65b9\u5411\u626b\u63cf\u8f93\u51fa\u6309\u7a7a\u95f4\u4f4d\u7f6e\u81ea\u9002\u5e94\u878d\u5408\uff1b\u4f46\u4ee3\u4ef7\u662f\u53c2\u6570\u548c FLOPs \u660e\u663e\u589e\u52a0\uff0c\u65b9\u6cd5\u521b\u65b0\u503c\u5f97\u501f\u9274\uff0c\u5de5\u7a0b\u6027\u4ef7\u6bd4\u9700\u8981\u8c28\u614e\u770b\u5f85\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u8bba\u6587\u7814\u7a76\u7684\u662f Mamba\/state space model \u5728\u533b\u5b66\u56fe\u50cf\u5206\u5272\u4e2d\u7684\u7a7a\u95f4\u5efa\u6a21\u95ee\u9898\u3002\u533b\u5b66\u56fe\u50cf\u5206\u5272\u5c24\u5176\u662f\u606f\u8089\u3001\u76ae\u635f\u7b49\u4efb\u52a1\u65e2\u9700\u8981\u5c40\u90e8\u8fb9\u754c\u7ec6\u8282\uff0c\u4e5f\u9700\u8981\u8f83\u5927\u8303\u56f4\u4e0a\u4e0b\u6587\u3002CNN \u5f3a\u5728\u5c40\u90e8\u7eb9\u7406\u4f46\u957f\u7a0b\u4f9d\u8d56\u6709\u9650\uff1bTransformer \u80fd\u5efa\u6a21\u5168\u5c40\u5173\u7cfb\u4f46\u6ce8\u610f\u529b\u590d\u6742\u5ea6\u9ad8\u3001\u5bf9\u6570\u636e\u89c4\u6a21\u548c\u663e\u5b58\u66f4\u654f\u611f\uff1bMamba \u7c7b\u6a21\u578b\u7528\u7ebf\u6027\u5e8f\u5217\u590d\u6742\u5ea6\u5efa\u6a21\u957f\u7a0b\u4f9d\u8d56\uff0c\u8fd1\u4e24\u5e74\u88ab\u5927\u91cf\u7528\u4e8e VM-UNet\u3001SegMamba \u7b49\u533b\u5b66\u5206\u5272\u6846\u67b6\u3002<\/p>\n<p>\u4f5c\u8005\u6307\u51fa\u73b0\u6709 Vision Mamba \u5206\u5272\u6a21\u578b\u6709\u4e24\u4e2a\u6838\u5fc3\u95ee\u9898\uff1a\u7b2c\u4e00\uff0c\u5e38\u89c1 raster\/pixel-wise \u626b\u63cf\u628a\u4e8c\u7ef4\u7279\u5f81\u56fe\u5c55\u5e73\u6210\u4e00\u7ef4\u5e8f\u5217\uff0c\u4f1a\u8ba9\u5782\u76f4\u76f8\u90bb\u6216\u5c40\u90e8\u76f8\u90bb\u50cf\u7d20\u5728\u5e8f\u5217\u4e2d\u76f8\u8ddd\u5f88\u8fdc\uff0c\u7834\u574f dense prediction \u6240\u9700\u7684\u5c40\u90e8\u7a7a\u95f4\u76f8\u5e72\u6027\uff1b\u7b2c\u4e8c\uff0c\u591a\u65b9\u5411\u626b\u63cf\u8f93\u51fa\u901a\u5e38\u7b80\u5355\u6c42\u548c\uff0c\u9ed8\u8ba4\u6240\u6709\u65b9\u5411\u3001\u6240\u6709\u5c3a\u5ea6\u5728\u6bcf\u4e2a\u4f4d\u7f6e\u540c\u7b49\u91cd\u8981\uff0c\u4e0d\u80fd\u9002\u5e94\u606f\u8089\/\u76ae\u635f\u76ee\u6807\u5927\u5c0f\u3001\u5f62\u72b6\u3001\u8fb9\u754c\u590d\u6742\u5ea6\u7684\u53d8\u5316\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u5bf9\u73b0\u6709\u65b9\u6cd5\u7684\u6279\u8bc4\u96c6\u4e2d\u5728 Mamba \u5206\u5272\u6a21\u578b\uff0c\u800c\u4e0d\u662f\u6cdb\u6cdb\u6279\u8bc4 U-Net\uff1a<\/p>\n<ul>\n<li>U-Net \/ U-Net v2\uff1askip connection \u548c encoder-decoder \u7ed3\u6784\u6709\u6548\uff0c\u4f46\u672c\u8eab\u6ca1\u6709\u89e3\u51b3\u5168\u5c40\u4f9d\u8d56\u5efa\u6a21\u95ee\u9898\u3002<\/li>\n<li>Transformer \/ Swin-UNet \/ TransUNet \u7c7b\u65b9\u6cd5\uff1a\u5168\u5c40\u5efa\u6a21\u80fd\u529b\u5f3a\uff0c\u4f46\u81ea\u6ce8\u610f\u529b\u7684\u4e8c\u6b21\u590d\u6742\u5ea6\u548c\u8bad\u7ec3\u6570\u636e\u9700\u6c42\u5728\u533b\u5b66\u573a\u666f\u4e2d\u4e0d\u7406\u60f3\u3002<\/li>\n<li>VM-UNet \/ VM-UNetV2 \/ Vision Mamba\uff1a\u80fd\u7ebf\u6027\u590d\u6742\u5ea6\u5efa\u6a21\u957f\u7a0b\u4f9d\u8d56\uff0c\u4f46\u628a 2D feature map \u76f4\u63a5\u7ebf\u6027\u626b\u63cf\u4f1a\u635f\u5bb3\u5c40\u90e8 2D \u51e0\u4f55\uff1b\u591a\u65b9\u5411\u626b\u63cf\u7ed3\u679c\u7684\u56fa\u5b9a\u6c42\u548c\u4e0d\u591f\u81ea\u9002\u5e94\u3002<\/li>\n<li>\u666e\u901a\u591a\u5c3a\u5ea6\u878d\u5408\uff1a\u5982\u679c\u53ea\u662f\u5806\u5377\u79ef\u6216\u56fa\u5b9a\u6c42\u548c\uff0c\u96be\u4ee5\u6839\u636e\u5c40\u90e8\u75c5\u7076\u5f62\u6001\u9009\u62e9\u5408\u9002\u65b9\u5411\u548c\u5c3a\u5ea6\u3002<\/li>\n<\/ul>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>Patch-MoE Mamba \u91c7\u7528 U-Net-style \u67b6\u6784\uff0c\u7531\u4e09\u90e8\u5206\u7ec4\u6210\uff1a<\/p>\n<ol>\n<li><strong>Mamba-based encoder<\/strong>\uff1a\u4ee5 VM-UNetV2 \u4e3a\u57fa\u7840\uff0c\u628a\u539f\u6765\u7684 VSS block \u66ff\u6362\u4e3a Patch-MoE VSS block\u3002<\/li>\n<li><strong>SDI \u6a21\u5757<\/strong>\uff1a\u91c7\u7528 U-Net v2 \u7684 Semantics and Detail Infusion\uff0c\u901a\u8fc7\u9ad8\u5c42\u8bed\u4e49\u4e0e\u4f4e\u5c42\u7ec6\u8282\u7684 Hadamard product \u878d\u5408\u589e\u5f3a\u591a\u5c3a\u5ea6\u7279\u5f81\u3002<\/li>\n<li><strong>Decoder<\/strong>\uff1a\u57fa\u672c\u4fdd\u7559 VM-UNetV2 \u7684 decoder \u8bbe\u8ba1\u3002<\/li>\n<\/ol>\n<p>\u6280\u672f\u6838\u5fc3\u5728 encoder \u7684 Patch-MoE VSS block\uff1a<\/p>\n<ul>\n<li>\u5bf9\u8f93\u5165\u7279\u5f81\u56fe (X_l \\in \\mathbb{R}^{C_l \\times H_l \\times W_l})\uff0c\u4e0d\u76f4\u63a5 raster flatten\uff0c\u800c\u662f\u5148\u5212\u5206\u4e3a\u975e\u91cd\u53e0 patch\uff0c\u518d\u5728\u6bcf\u4e2a patch \u5185\u6309\u5c40\u90e8\u987a\u5e8f\u679a\u4e3e\u50cf\u7d20\uff0c\u5f62\u6210 patch-ordered permutation\u3002<\/li>\n<li>\u4f7f\u7528\u591a\u4e2a patch size \u6784\u9020\u5c42\u6b21\u5316\u626b\u63cf\u5e8f\u5217\uff0c\u8ba9\u65e9\u671f\/\u4e0d\u540c\u65b9\u5411\u53ef\u4ee5\u4f7f\u7528\u8f83\u7c97\u6216\u8f83\u7ec6\u7c92\u5ea6\u7684\u5c40\u90e8\u8fde\u7eed\u5e8f\u5217\u3002<\/li>\n<li>\u56db\u4e2a\u65b9\u5411\u626b\u63cf\uff1aforward\u3001reverse\u3001WH forward\u3001WH reverse\u3002\u6bcf\u4e2a\u65b9\u5411\u53ef\u5bf9\u5e94\u4e0d\u540c patch granularity\u3002<\/li>\n<li>\u56db\u4e2a\u65b9\u5411\u8f93\u51fa (Y_l^{(1)},\\dots,Y_l^{(4)}) \u4e0d\u518d\u7b80\u5355\u76f8\u52a0\uff0c\u800c\u662f\u4f5c\u4e3a\u4e13\u5bb6\uff0c\u5916\u52a0\u4e00\u4e2a concat expert\uff1b\u901a\u8fc7 spatial-aware router \u751f\u6210\u6bcf\u4e2a\u4f4d\u7f6e\u7684 expert \u6743\u91cd\u3002<\/li>\n<li>\u6700\u7ec8\u8f93\u51fa\u4e3a MoE \u52a0\u6743\u878d\u5408\u52a0\u4e0a\u539f\u59cb\u65b9\u5411\u8f93\u51fa residual sum\uff1a<\/li>\n<\/ul>\n<p>[<br \/>\n\\tilde{Y}<em e=\"1\">l = \\sum<\/em> \\odot E_e,<br \/>\n\\quad<br \/>\nZ_l = \\tilde{Y}}^{5} w_{l,e<em i=\"1\">l + \\sum<\/em>.<br \/>\n]}^{4} Y_l^{(i)<\/p>\n<p>\u8fd9\u4f7f\u8be5 block \u4fdd\u7559 VMamba\/VSS \u7684\u957f\u7a0b\u5efa\u6a21\uff0c\u540c\u65f6\u5f3a\u5316\u5c40\u90e8\u76f8\u5e72\u548c\u65b9\u5411\u81ea\u9002\u5e94\u3002<\/p>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<p><strong>\u6a21\u5757 A\uff1aPatch-Ordered Scanning<\/strong><br \/>\n\u8f93\u5165\u662f stage (l) \u7684 feature map (X_l)\u3002\u7ed9\u5b9a patch size (p)\uff0c\u5c06\u7a7a\u95f4\u7f51\u683c\u5212\u5206\u4e3a (\\lceil H_l\/p\\rceil \\times \\lceil W_l\/p\\rceil) \u4e2a patch\uff0c\u6bcf\u4e2a patch \u5185\u6309 row-major \u679a\u4e3e\uff0c\u7136\u540e\u518d\u8bbf\u95ee\u4e0b\u4e00\u4e2a patch\u3002\u8f93\u51fa\u662f\u957f\u5ea6\u4e3a (H_lW_l) \u7684 permutation sequence\u3002\u5b83\u4e0d\u505a pooling\u3001\u4e0d\u964d\u5206\u8fa8\u7387\uff0c\u53ea\u6539\u53d8 token visiting order\u3002\u5176\u4f5c\u7528\u662f\u8ba9 patch \u5185\u50cf\u7d20\u5728\u5e8f\u5217\u4e2d\u8fde\u7eed\uff0c\u7f13\u89e3 raster scanning \u4e0b\u5c40\u90e8\u90bb\u57df\u88ab\u6253\u6563\u7684\u95ee\u9898\u3002<\/p>\n<p><strong>\u521b\u65b0\u6027\u5224\u65ad<\/strong>\uff1a\u8fd9\u662f\u4e00\u4e2a\u6e05\u6670\u3001\u53ef\u8fc1\u79fb\u7684\u8bbe\u8ba1\u70b9\u3002\u5b83\u4e0d\u6539\u53d8 Mamba \u7b97\u5b50\u672c\u8eab\uff0c\u4f46\u6539\u53d8 2D-to-1D \u5e8f\u5217\u5316\u65b9\u5f0f\uff0c\u9002\u5408\u8fc1\u79fb\u5230 VM-UNet\u3001VM-UNetV2\u3001SegMamba \u6216 DAMamba \u7c7b\u6846\u67b6\u4e2d\u3002\u5bf9 polyp segmentation \u7684\u8fb9\u754c\u548c\u5c0f\u76ee\u6807\u8f83\u6709\u610f\u4e49\uff1b\u5bf9 3D medical image segmentation \u4e5f\u6709\u542f\u53d1\uff0c\u4f46 3D \u4e2d\u9700\u8981\u6269\u5c55\u4e3a block-ordered \/ plane-aware \/ anisotropic scan\uff0c\u5426\u5219\u76f4\u63a5\u5957\u7528 2D patch \u89c4\u5219\u4e0d\u591f\u3002<\/p>\n<p><strong>\u6a21\u5757 B\uff1aHierarchical Patch Sizes<\/strong><br \/>\n\u4f5c\u8005\u8bbe\u7f6e\u4e0d\u540c\u5c42\u6216\u65b9\u5411\u7684 patch size\uff0c\u4f8b\u5982\u5b9e\u9a8c\u4e2d\u6bd4\u8f83 8844\/8844\/8844\/8844\u30018844\/1111\/1111\/1111 \u7b49\u7ec4\u5408\u3002\u76f4\u89c9\u662f\u6d45\u5c42\/\u4e0d\u540c\u65b9\u5411\u53ef\u4ee5\u7528\u8f83\u7c97 patch \u6355\u83b7\u533a\u57df\u7ed3\u6784\uff0c\u6df1\u5c42\u6216\u8fb9\u754c\u654f\u611f\u5904\u7528\u66f4\u7ec6\u7c92\u5ea6\u626b\u63cf\u4fdd\u7559\u7ec6\u8282\u3002<\/p>\n<p><strong>\u521b\u65b0\u6027\u5224\u65ad<\/strong>\uff1a\u66f4\u50cf\u662f\u5b9e\u7528\u8bbe\u8ba1\u548c\u8d85\u53c2\u6570\u7b56\u7565\uff0c\u800c\u4e0d\u662f\u72ec\u7acb\u7406\u8bba\u8d21\u732e\u3002Table IV \u663e\u793a patch-size \u7ec4\u5408\u5f71\u54cd\u5e73\u5747 Dice\uff0c\u4f46\u641c\u7d22\u7a7a\u95f4\u6709\u9650\uff0c\u672a\u8bc1\u660e\u6700\u4f18\u6027\u3002<\/p>\n<p><strong>\u6a21\u5757 C\uff1aMoE-Based Directional Fusion<\/strong><br \/>\n\u56db\u4e2a\u65b9\u5411\u8f93\u51fa\u5148\u7ecf GroupNorm \u5f97\u5230\u56db\u4e2a directional experts\uff1a(E_1,\\dots,E_4)\u3002\u518d\u628a\u56db\u4e2a\u65b9\u5411 concat \u540e\u7528 (1\\times1) conv + BN + ReLU \u6295\u5f71\u4e3a\u7b2c\u4e94\u4e2a concat expert\uff1a(E_{concat})\u3002router \u7531 local descriptor \u548c global descriptor \u6df7\u5408\u9a71\u52a8\uff1a<\/p>\n<p>[<br \/>\nF_{local}=DWConv_{3\\times3}\\left(\\sum_iY_l^{(i)}\\right),<br \/>\n\\quad<br \/>\nF_{global}=GAP\\left(\\sum_iY_l^{(i)}\\right),<br \/>\n]<\/p>\n<p>[<br \/>\nF_l=\\alpha F_{local}+(1-\\alpha)F_{global},<br \/>\n\\quad<br \/>\nw_l=Softmax(Router(F_l)).<br \/>\n]<\/p>\n<p>\u8f93\u51fa\u662f\u4e94\u4e2a\u4e13\u5bb6\u7684\u7a7a\u95f4\u4f4d\u7f6e\u81ea\u9002\u5e94\u52a0\u6743\u548c\u3002<\/p>\n<p><strong>\u521b\u65b0\u6027\u5224\u65ad<\/strong>\uff1a\u6bd4\u56fa\u5b9a\u6c42\u548c\u66f4\u7075\u6d3b\uff0c\u9002\u5408\u75c5\u7076\u5f62\u6001\u53d8\u5316\u5927\u7684\u5206\u5272\u4efb\u52a1\u3002\u4f46 concat expert \u548c router \u5e26\u6765\u663e\u8457\u53c2\u6570\/FLOPs \u589e\u957f\uff1b\u8be5\u6a21\u5757\u662f\u5426\u6bd4\u66f4\u8f7b\u91cf\u7684 attention\/gating \u66f4\u5212\u7b97\uff0c\u8bba\u6587\u6ca1\u6709\u5145\u5206\u6bd4\u8f83\u3002<\/p>\n<p><strong>\u6a21\u5757 D\uff1aResidual Directional Aggregation<\/strong><br \/>\n\u6700\u7ec8\u8f93\u51fa\u52a0\u56de\u56db\u4e2a raw directional outputs\uff1a(Z_l=\\tilde{Y}_l+\\sum_iY_l^{(i)})\u3002\u4f5c\u7528\u662f\u9632\u6b62 router \u8bad\u7ec3\u65e9\u671f\u9000\u5316\uff0c\u5e76\u4fdd\u7559\u56fa\u5b9a\u65b9\u5411\u4fe1\u53f7\u3002<\/p>\n<p><strong>\u8fc1\u79fb\u4ef7\u503c<\/strong>\uff1a\u5bf9 DAMamba \u6216 VM-UNet \u7c7b\u7ed3\u6784\u5f88\u5b9e\u7528\u3002\u82e5\u7528\u6237\u540e\u7eed\u505a DAMamba \u6539\u9020\uff0c\u53ef\u4ee5\u5148\u53ea\u8fc1\u79fb patch-ordered scanning\uff0c\u518d\u8003\u8651\u7528\u8f7b\u91cf gating \u66ff\u4ee3\u5b8c\u6574 MoE\uff0c\u4ee5\u907f\u514d\u590d\u6742\u5ea6\u7206\u70b8\u3002<\/p>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p>\u5b9e\u9a8c\u5305\u542b\u4e24\u4e2a\u4efb\u52a1\u65cf\uff1a<\/p>\n<ul>\n<li><strong>Polyp segmentation<\/strong>\uff1aKvasir-SEG\u3001ClinicDB\u3001ColonDB\u3001ETIS\u3001CVC-300\u3002\u8bad\u7ec3\u534f\u8bae\u6cbf\u7528 U-Net v2\uff1aKvasir-SEG 900 \u5f20 + ClinicDB 550 \u5f20\u8bad\u7ec3\uff1b\u6d4b\u8bd5\u5305\u542b CVC-300 60\u3001ColonDB 380\u3001ETIS 196\u3001Kvasir 100\u3001ClinicDB 62\u3002<\/li>\n<li><strong>Skin lesion segmentation<\/strong>\uff1aISIC 2017\u3001ISIC 2018\uff0c\u91c7\u7528 U-Net v2 \u7684 split\u3002<\/li>\n<\/ul>\n<p>\u8bad\u7ec3\u8bbe\u7f6e\uff1aPyTorch\uff0cNVIDIA Tesla A100 80GB\uff0cAdamW\uff0clr=1e-3\uff0cbatch size=80\uff0c\u8f93\u5165 256\u00d7256\uff0c300 epochs\uff0ccosine annealing\uff0cVMamba-S pretrained initialization\uff0c\u6807\u51c6\u968f\u673a\u7ffb\u8f6c\/\u65cb\u8f6c\u589e\u5f3a\u3002\u6bcf\u4e2a\u5b9e\u9a8c\u8fd0\u884c 5 \u4e2a\u968f\u673a\u79cd\u5b50\u3002<\/p>\n<p>\u4e3b\u8981\u7ed3\u679c\uff1a<\/p>\n<ul>\n<li>\u5728 5 \u4e2a\u606f\u8089\u6570\u636e\u96c6\u4e0a\uff0cPatch-MoE Mamba \u7684 Dice \u5747\u4e3a\u6700\u9ad8\u6216\u63a5\u8fd1\u6700\u9ad8\u3002\u76f8\u5bf9 VM-UNetV2\uff0cKvasir-SEG \u63d0\u5347\u5f88\u5c0f\uff0890.82\u219290.90\uff09\uff0cClinicDB \u63d0\u5347\u8f83\u660e\u663e\uff0890.52\u219291.32\uff09\uff0cColonDB \u63d0\u5347\uff0876.62\u219277.94\uff09\uff0cETIS \u63d0\u5347\uff0872.56\u219274.04\uff09\uff0cCVC-300 \u63d0\u5347\uff0886.80\u219287.31\uff09\u3002<\/li>\n<li>\u5728 ISIC 2017\/2018 \u4e0a\uff0cPatch-MoE Mamba \u4e5f\u4f18\u4e8e U-Net\u3001U-Net v2\u3001VM-UNet\u3001VM-UNetV2\uff1b\u4f8b\u5982 ISIC 2018 Dice \u7531 VM-UNetV2 \u7684 88.36 \u63d0\u5347\u5230 89.34\u3002<\/li>\n<li>Ablation\uff1aVM-UNetV2 \u5e73\u5747 Dice 83.46\uff1b\u52a0\u5165 patch-ordered scanning \u540e 84.02\uff1b\u518d\u52a0\u5165 MoE fusion \u540e 84.30\u3002\u8bf4\u660e\u4e24\u4e2a\u6838\u5fc3\u7ec4\u4ef6\u90fd\u6709\u589e\u76ca\uff0c\u4f46 MoE \u7684\u989d\u5916\u5e73\u5747\u6536\u76ca\u7ea6 +0.28 Dice\u3002<\/li>\n<li>\u590d\u6742\u5ea6\uff1aU-Net v2 \u4e3a 25.15M \/ 5.58G FLOPs\uff0cVM-UNetV2 \u4e3a 22.77M \/ 5.31G\uff1bPatch-MoE Mamba \u4e3a 70.06M \/ 28.18G\u3002\u590d\u6742\u5ea6\u589e\u52a0\u975e\u5e38\u660e\u663e\u3002<\/li>\n<\/ul>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u53ef\u4fe1\u70b9\uff1a<\/p>\n<ul>\n<li>\u8bba\u6587\u4e0d\u662f\u53ea\u62a5\u5355\u6b21\u7ed3\u679c\uff0c\u800c\u662f Table I \u660e\u786e\u6bcf\u4e2a\u5b9e\u9a8c\u8fd0\u884c 5 \u4e2a\u968f\u673a\u79cd\u5b50\uff0c\u5e76\u62a5\u544a\u5747\u503c\/\u65b9\u5dee\u3002<\/li>\n<li>baseline \u8986\u76d6 U-Net\u3001U-Net v2\u3001VM-UNet\u3001VM-UNetV2\uff0c\u548c\u8bba\u6587\u65b9\u6cd5\u7684\u8840\u7f18\u5173\u7cfb\u6bd4\u8f83\u76f4\u63a5\u3002<\/li>\n<li>\u6709\u8de8\u4efb\u52a1\u9a8c\u8bc1\uff1a\u606f\u8089 + ISIC \u76ae\u635f\u3002<\/li>\n<li>\u6709\u6838\u5fc3\u7ec4\u4ef6\u6d88\u878d\uff1apatch-ordered scanning \u4e0e MoE fusion \u5206\u522b\u52a0\u5165\u3002<\/li>\n<\/ul>\n<p>\u4e3b\u8981\u98ce\u9669\uff1a<\/p>\n<ul>\n<li>\u8bba\u6587\u53ea\u6709 5 \u9875\uff0c\u5b9e\u9a8c\u4e0e\u5b9e\u73b0\u7ec6\u8282\u76f8\u5bf9\u538b\u7f29\uff1b\u6ca1\u6709\u66f4\u5927\u89c4\u6a21\u3001\u591a\u4e2d\u5fc3 3D \u533b\u5b66\u5206\u5272\u9a8c\u8bc1\u3002<\/li>\n<li>\u76f8\u6bd4 VM-UNetV2\uff0c\u6027\u80fd\u63d0\u5347\u901a\u5e38\u662f 0.x \u5230 1.x Dice\uff0c\u4f46\u53c2\u6570\u4ece 22.77M \u589e\u81f3 70.06M\u3001FLOPs \u4ece 5.31G \u589e\u81f3 28.18G\uff1b\u5982\u679c\u5f3a\u8c03\u201cefficient Mamba\u201d\uff0c\u8fd9\u4e2a\u590d\u6742\u5ea6\u4ee3\u4ef7\u4f1a\u524a\u5f31\u8bf4\u670d\u529b\u3002<\/li>\n<li>MoE fusion \u7684\u6536\u76ca\u76f8\u5bf9\u5176\u5f00\u9500\u4e0d\u7b97\u5927\uff1aTable III \u4e2d patch scanning \u5df2\u5e26\u6765\u4e3b\u8981\u63d0\u5347\uff0cMoE \u53ea\u4ece 84.02 \u5230 84.30\u3002<\/li>\n<li>\u6ca1\u6709\u4e0e\u66f4\u5f3a\u6216\u66f4\u65b0\u7684 polyp segmentation SOTA\uff08\u4f8b\u5982\u4e13\u95e8\u9762\u5411 PolypGen\u3001\u771f\u5b9e\u9000\u5316\u3001foundation model adaptation \u7684\u65b9\u6cd5\uff09\u5168\u9762\u6bd4\u8f83\u3002<\/li>\n<li>\u6ca1\u6709\u7edf\u8ba1\u663e\u8457\u6027\u68c0\u9a8c\uff0c\u867d\u6709 seed \u65b9\u5dee\uff0c\u4f46\u672a\u8bf4\u660e\u5dee\u5f02\u662f\u5426\u663e\u8457\u3002<\/li>\n<\/ul>\n<h4>8. \u4e0e\u4e3b\u6d41\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u7684\u5173\u7cfb<\/h4>\n<ul>\n<li><strong>U-Net \/ nnU-Net<\/strong>\uff1a\u6574\u4f53\u4ecd\u662f encoder-decoder + skip\/fusion \u601d\u8def\uff0c\u4e0d\u80fd\u66ff\u4ee3 nnU-Net \u7684\u5b8c\u6574 pipeline\uff1b\u66f4\u50cf\u662f\u53ef\u63d2\u5165\u67d0\u7c7b U-Net encoder \u7684\u5e8f\u5217\u5efa\u6a21\u6a21\u5757\u3002<\/li>\n<li><strong>U-Net v2 \/ MedNeXt<\/strong>\uff1a\u501f\u7528\u4e86 U-Net v2 \u7684 SDI \u6a21\u5757\uff1b\u4e0e MedNeXt \u8fd9\u7c7b\u7eaf ConvNeXt\/large-kernel CNN \u8def\u7ebf\u76f8\u6bd4\uff0c\u5b83\u66f4\u5f3a\u8c03 Mamba \u7684\u957f\u7a0b\u4f9d\u8d56\u3002<\/li>\n<li><strong>UNetR \/ Swin-UNet \/ TransUNet \/ TransFuse<\/strong>\uff1a\u89e3\u51b3 Transformer \u9ad8\u590d\u6742\u5ea6\u4e0e\u6570\u636e\u9700\u6c42\u95ee\u9898\uff0c\u4f7f\u7528 state space scan \u66ff\u4ee3 attention\uff1b\u4f46 Patch-MoE \u672c\u8eab\u7684 MoE\/concat \u53c8\u5e26\u6765\u8f83\u9ad8 FLOPs\u3002<\/li>\n<li><strong>Mamba \/ VMamba \/ VM-UNet \/ VM-UNetV2 \/ SegMamba \/ DAMamba<\/strong>\uff1a\u8fd9\u662f\u6700\u76f4\u63a5\u76f8\u5173\u8def\u7ebf\u3002\u8bba\u6587\u5bf9 VM-UNetV2 \u7684\u6539\u9020\u70b9\u5f88\u660e\u786e\uff1a\u626b\u63cf\u987a\u5e8f + \u591a\u65b9\u5411\u878d\u5408\u65b9\u5f0f\u3002\u5bf9 DAMamba \u6539\u9020\u5c24\u5176\u503c\u5f97\u501f\u9274 patch-ordered scan\u3002<\/li>\n<li><strong>Foundation model for medical segmentation<\/strong>\uff1a\u57fa\u672c\u65e0\u76f4\u63a5\u5173\u7cfb\uff1b\u6ca1\u6709\u4f7f\u7528 SAM\/MedSAM\/\u901a\u7528 prompt segmentation\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u5bf9 polyp segmentation\uff1a\u4ef7\u503c\u8f83\u9ad8\u3002\u8bba\u6587\u76f4\u63a5\u5728 5 \u4e2a\u606f\u8089\u5206\u5272\u6570\u636e\u96c6\u8bc4\u4f30\uff0c\u5e76\u5728 ETIS\u3001ColonDB \u8fd9\u79cd\u66f4\u96be\u6570\u636e\u96c6\u4e0a\u6709\u76f8\u5bf9\u66f4\u660e\u663e\u63d0\u5347\u3002patch-ordered scanning \u5bf9\u5c0f\u76ee\u6807\u3001\u4f4e\u5bf9\u6bd4\u3001\u8fb9\u754c\u4e0d\u89c4\u5219\u75c5\u7076\u6709\u5408\u7406\u5f52\u7eb3\u504f\u7f6e\u3002<\/p>\n<p>\u5bf9 DAMamba \u6539\u9020\uff1a\u4ef7\u503c\u5f88\u9ad8\u3002\u5efa\u8bae\u4f18\u5148\u590d\u73b0\/\u501f\u9274 <strong>patch-ordered scanning<\/strong>\uff0c\u56e0\u4e3a\u5b83\u7b80\u5355\u3001\u89e3\u91ca\u6e05\u695a\u3001\u5bf9\u590d\u6742\u5ea6\u5f71\u54cd\u5c0f\uff1bMoE fusion \u53ef\u4ee5\u4f5c\u4e3a\u7b2c\u4e8c\u9636\u6bb5\u63a2\u7d22\uff0c\u4f46\u5e94\u8bbe\u8ba1 lightweight router \u6216\u53bb\u6389 concat expert \u4ee5\u63a7\u5236 FLOPs\u3002<\/p>\n<p>\u5bf9 3D medical image segmentation\uff1a\u6982\u5ff5\u6709\u542f\u53d1\uff0c\u4f46\u4e0d\u80fd\u76f4\u63a5\u7167\u642c\u30023D \u4e2d\u53ef\u4ee5\u8003\u8651 patch\/block-ordered scanning\u3001slice-aware scanning\u3001\u8f74\u5411 anisotropic scanning\uff0c\u4f46\u5fc5\u987b\u91cd\u65b0\u9a8c\u8bc1\u663e\u5b58\u548c\u626b\u63cf\u8def\u5f84\u5bf9\u4f53\u7d20\u90bb\u63a5\u7684\u5f71\u54cd\u3002<\/p>\n<p>\u5bf9 related work\uff1a\u9002\u5408\u653e\u5728 Mamba-based medical segmentation \u6216 scan-order design \u76f8\u5173\u6bb5\u843d\uff0c\u8bf4\u660e\u73b0\u6709 Mamba \u5206\u5272\u5f00\u59cb\u4ece\u201c\u7528 SSM \u66ff\u4ee3 attention\u201d\u8fdb\u5165\u201c\u5982\u4f55\u628a 2D\/3D \u7a7a\u95f4\u7ed3\u6784\u5408\u7406\u5e8f\u5217\u5316\u201d\u7684\u9636\u6bb5\u3002<\/p>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5efa\u8bae\u7cbe\u8bfb\uff0c\u4f46\u91cd\u70b9\u8bfb\u65b9\u6cd5\u548c\u6d88\u878d\uff0c\u7ed3\u679c\u90e8\u5206\u8981\u5e26\u7740\u590d\u6742\u5ea6\u4ee3\u4ef7\u5ba1\u89c6\u3002<\/strong> \u5982\u679c\u7528\u6237\u505a DAMamba \u6216 Mamba-UNet \u6539\u9020\uff0c\u8fd9\u7bc7\u503c\u5f97\u5b8c\u6574\u8bfb Method \u7684 patch scanning \u4e0e MoE fusion\uff1b\u5982\u679c\u76ee\u6807\u662f\u8f7b\u91cf\u5b9e\u65f6\u90e8\u7f72\uff0c\u5219\u4e0d\u5efa\u8bae\u628a\u5b8c\u6574 Patch-MoE \u4f5c\u4e3a\u9996\u9009 baseline\uff0c\u56e0\u4e3a\u5176 70.06M \u53c2\u6570\u548c 28.18G FLOPs \u5df2\u7ecf\u504f\u91cd\u3002<\/p>\n<hr \/>\n<h2>\u8bba\u6587 2\uff1aDepthPolyp: Pseudo-Depth Guided Lightweight Segmentation for Real-Time Colonoscopy<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aDepthPolyp: Pseudo-Depth Guided Lightweight Segmentation for Real-Time Colonoscopy<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aZhuoyu Wu et al.<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-15<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.16519<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1aMEDIA:\/root\/papers_medseg_2026-05-23\/2605.16519.pdf \uff1bhttps:\/\/arxiv.org\/pdf\/2605.16519<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/ReaganWu\/DepthPolyp\/ \uff08\u8bba\u6587\u6b63\u6587\u7ed9\u51fa\uff1b\u672c\u6b21\u672a\u80fd\u5728\u53d7\u9650\u73af\u5883\u4e2d\u5b8c\u6210\u4ed3\u5e93\u8fde\u901a\u6027\u9a8c\u8bc1\uff09<\/li>\n<li>\u4efb\u52a1\uff1areal-time polyp segmentation \/ lightweight endoscopy segmentation \/ robustness under degradation<\/li>\n<li>\u6570\u636e\u96c6\uff1aKvasir-SEG\u3001CVC-ClinicDB\u3001CVC-ColonDB\u3001PolypGen sequences 18\u201322<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1alightweight segmentation \/ pseudo-depth-guided multi-task learning \/ MiT-B0 encoder \/ Ghost Factorization \/ Interleaved Shuffle Fusion \/ Dynamic Group Gating \/ uncertainty-weighted multi-task loss<\/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\u7684\u4ef7\u503c\u5728\u4e8e\u628a\u606f\u8089\u5206\u5272\u4ece clean benchmark \u63a8\u5411\u771f\u5b9e\u5185\u955c\u9000\u5316\u548c\u79fb\u52a8\u7aef\u5b9e\u65f6\u90e8\u7f72\uff1a\u7528 pseudo-depth \u4f5c\u4e3a\u8bad\u7ec3\u671f\u7ed3\u6784\u76d1\u7763\uff0c\u5e76\u914d\u5408\u8f7b\u91cf decoder\uff0c\u5728 PolypGen \u548c noisy cross-dataset \u8bbe\u7f6e\u4e2d\u5c55\u793a\u4e86\u6bd4\u8bb8\u591a\u5927\u6a21\u578b\u66f4\u7a33\u7684\u9c81\u68d2\u6027\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u606f\u8089\u5206\u5272\u5bf9\u65e9\u671f\u7ed3\u76f4\u80a0\u764c\u7b5b\u67e5\u5f88\u91cd\u8981\uff0c\u4f46\u771f\u5b9e\u7ed3\u80a0\u955c\u89c6\u9891\u4e2d\u5e38\u89c1 motion blur\u3001specular reflection\u3001illumination instability\u3001defocus\u3001JPEG\/\u4f4e\u8d28\u4f20\u8f93\u7b49\u9000\u5316\u3002\u5f88\u591a polyp segmentation \u65b9\u6cd5\u5728 Kvasir\u3001ClinicDB \u7b49 clean image benchmark \u4e0a Dice \u5f88\u9ad8\uff0c\u4f46\u771f\u5b9e\u624b\u672f\u89c6\u9891\u91cc\u9884\u6d4b\u4f1a\u788e\u88c2\u3001\u6f0f\u68c0\u6216\u4ea7\u751f\u53cd\u5149\u533a\u57df\u5047\u9633\u6027\u3002<\/p>\n<p>\u8bba\u6587\u7814\u7a76\u7684\u95ee\u9898\u662f\uff1a\u5982\u4f55\u5728\u4f4e\u53c2\u6570\u3001\u4f4e GMAC\u3001\u79fb\u52a8\u7aef\u5b9e\u65f6\u6761\u4ef6\u4e0b\uff0c\u4f7f\u606f\u8089\u5206\u5272\u6a21\u578b\u5bf9\u771f\u5b9e\u5185\u955c\u9000\u5316\u66f4\u9c81\u68d2\uff1f\u4f5c\u8005\u4e0d\u53ea\u63d0\u51fa\u4e00\u4e2a\u6a21\u578b\uff0c\u8fd8\u63d0\u51fa\u56db\u8c61\u9650\u9c81\u68d2\u8bc4\u4f30\u534f\u8bae\uff1aClean\u2192Clean\u3001Clean\u2192Noisy\u3001Noisy\u2192Clean\u3001Noisy\u2192Noisy\uff0c\u7528\u6765\u5206\u79bb clean \u6027\u80fd\u3001\u5206\u5e03\u504f\u79fb\u3001noisy training \u7684\u6062\u590d\u6548\u679c\u548c clean-domain penalty\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u8ba4\u4e3a\u73b0\u6709\u65b9\u6cd5\u4e3b\u8981\u6709\u4e09\u7c7b\u4e0d\u8db3\uff1a<\/p>\n<ul>\n<li><strong>Transformer \/ hybrid \u5927\u6a21\u578b<\/strong>\uff1a\u5982 TransUNet\u3001SegFormer\u3001PraNet\u3001CFFormer \u7b49\uff0c\u5728 clean benchmark \u4e0a\u5f3a\uff0c\u4f46\u53c2\u6570\u91cf\u901a\u5e38\u8f83\u5927\uff0c\u4e14\u5728 blur\/noise \u4e0b Dice \u4e0b\u964d\u660e\u663e\uff0c\u4e0d\u9002\u5408\u4f4e\u8d44\u6e90\u5b9e\u65f6\u90e8\u7f72\u3002<\/li>\n<li><strong>\u8f7b\u91cf\u6a21\u578b<\/strong>\uff1aMobilePolypNet\u3001ULite\u3001UNeXt\u3001CMUNeXt \u7b49\u91cd\u89c6\u6548\u7387\uff0c\u4f46\u8868\u793a\u80fd\u529b\u6709\u9650\uff0c\u5728\u9000\u5316\u8f93\u5165\u4e0a\u7a33\u5b9a\u6027\u4e0d\u8db3\u3002<\/li>\n<li><strong>\u591a\u4efb\u52a1\/\u8fb9\u754c\/\u663e\u8457\u6027\u8f85\u52a9\u65b9\u6cd5<\/strong>\uff1a\u8fb9\u7f18\u6216\u663e\u8457\u6027\u76d1\u7763\u672c\u8eab\u4e5f\u4f9d\u8d56\u5916\u89c2\u8d28\u91cf\uff0c\u9047\u5230\u53cd\u5149\u3001\u6a21\u7cca\u3001\u7167\u660e\u53d8\u5316\u65f6\u4e0d\u4e00\u5b9a\u53ef\u9760\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cmonocular pseudo-depth \u63d0\u4f9b\u76f8\u5bf9\u7ed3\u6784\u7ebf\u7d22\uff0c\u53ef\u80fd\u6bd4\u7eaf appearance cues \u66f4\u8010\u9000\u5316\u3002<\/li>\n<\/ul>\n<p>\u4f5c\u8005\u8fd8\u5f3a\u8c03\u4e00\u4e2a\u8bc4\u4ef7\u5c42\u9762\u7684\u4e0d\u8db3\uff1a\u591a\u6570\u65e2\u6709 polyp segmentation \u8bba\u6587\u53ea\u5728\u9ad8\u8d28\u91cf\u9759\u6001\u56fe\u50cf\u4e0a\u6d4b\u8bd5\uff0c\u4ece\u800c\u9ad8\u4f30\u4e86\u4e34\u5e8a\u90e8\u7f72\u53ef\u9760\u6027\u3002<\/p>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>DepthPolyp \u7684 pipeline \u5982\u4e0b\uff1a<\/p>\n<ol>\n<li>\u8f93\u5165\u56fe\u50cf (I \\in \\mathbb{R}^{B\\times3\\times H\\times W})\u3002<\/li>\n<li>\u4f7f\u7528 MiT-B0 encoder \u63d0\u53d6\u56db\u5c3a\u5ea6\u7279\u5f81 ({c_1,c_2,c_3,c_4})\u3002\u6bcf\u4e2a\u5c3a\u5ea6\u7ecf MLP \u6295\u5f71\u3001reshape\u3001\u4e0a\u91c7\u6837\u5230 (H\/4\\times W\/4)\uff1a<\/li>\n<\/ol>\n<p>[<br \/>\n\\tilde{c}_i = Upsample(reshape(MLP_i(c_i)), size=(H\/4,W\/4)).<br \/>\n]<\/p>\n<ol start=\"3\">\n<li>Decoder \u7528 GFM\u3001ISF\u3001DGG \u8fdb\u884c\u8f7b\u91cf\u591a\u5c3a\u5ea6\u878d\u5408\uff0c\u5f97\u5230 (F_{out})\u3002<\/li>\n<li>\u4e24\u4e2a head\uff1asegmentation head \u8f93\u51fa mask logits\uff1bdepth head \u8f93\u51fa\u5f52\u4e00\u5316\u6df1\u5ea6\u56fe\u3002<\/li>\n<li>\u8bad\u7ec3\u65f6\u7528 frozen Depth-Anything v2-small \u751f\u6210 pseudo-depth target\uff0c\u53ea\u4f5c\u4e3a\u8f85\u52a9\u76d1\u7763\uff1b\u63a8\u7406\u65f6\u4e0d\u9700\u8981 Depth-Anything\uff0c\u56e0\u6b64\u4e0d\u589e\u52a0 inference cost\u3002<\/li>\n<li>\u635f\u5931\u7531 Dice segmentation loss\u3001Smooth-L1 depth loss \u548c uncertainty-based multi-task weighting \u7ec4\u6210\uff1a<\/li>\n<\/ol>\n<p>[<br \/>\nL = \\frac{1}{2\\sigma_s^2}L_{seg}+\\frac{1}{2\\sigma_d^2}L_{depth}+\\log\\sigma_s+\\log\\sigma_d.<br \/>\n]<\/p>\n<p>\u6574\u4f53\u601d\u60f3\u662f\uff1a\u7528 pseudo-depth \u5728\u8bad\u7ec3\u671f\u7ed9\u6a21\u578b\u6ce8\u5165\u7ed3\u6784\u5f52\u7eb3\u504f\u7f6e\uff1b\u7528\u8f7b\u91cf decoder \u4fdd\u6301\u5b9e\u65f6\u6027\uff1b\u7528 noisy training \u548c\u56db\u8c61\u9650\u8bc4\u4f30\u903c\u8fd1\u771f\u5b9e\u5185\u955c\u90e8\u7f72\u573a\u666f\u3002<\/p>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<p><strong>\u6a21\u5757 A\uff1aGhost Factorization Module (GFM)<\/strong><br \/>\n\u8f93\u5165 (X\\in\\mathbb{R}^{B\\times C_{in}\\times H\\times W})\u3002\u5148\u7528 pointwise (1\\times1) convolution \u4ea7\u751f primary component\uff1a<\/p>\n<p>[<br \/>\nX_p=PWConv(X),<br \/>\n]<\/p>\n<p>\u518d\u7528 depthwise convolution \u4ea7\u751f cheap auxiliary component\uff1a<\/p>\n<p>[<br \/>\nX_a=DWConv(X_p),<br \/>\n]<\/p>\n<p>\u8f93\u51fa ((X_p,X_a))\uff0c\u6ee1\u8db3 (C_p+C_a=C_{out})\u3002\u5b83\u501f\u9274 GhostNet \u7684 cheap feature generation\uff0c\u7528\u8f83\u5c11\u53c2\u6570\u8fd1\u4f3c\u66f4\u5bc6\u96c6\u7684\u5377\u79ef\u8868\u8fbe\u3002<\/p>\n<p><strong>\u521b\u65b0\u6027\u5224\u65ad<\/strong>\uff1a\u6a21\u5757\u672c\u8eab\u4e0d\u662f\u5168\u65b0\u601d\u60f3\uff0c\u4e3b\u8981\u4ef7\u503c\u5728\u4e8e\u7528\u4e8e\u9ad8\u5206\u8fa8\u7387\u5206\u5272 decoder \u7684\u5c42\u6b21\u5316\u8f7b\u91cf\u805a\u5408\u3002\u9002\u5408\u8fc1\u79fb\u5230 U-Net decoder\u3001polyp segmentation \u8f7b\u91cf\u6a21\u578b\u3001\u79fb\u52a8\u7aef baseline\u3002<\/p>\n<p><strong>\u6a21\u5757 B\uff1aHierarchical Factorized Decoder<\/strong><br \/>\n\u56db\u4e2a\u5c3a\u5ea6 (\\tilde{c}<em out=\"out\">i) \u5206\u522b\u7ecf GFM\uff0c\u4ea7\u751f primary \u548c auxiliary stream\uff1b\u5206\u522b concat \u540e\u8fdb\u5165 ISF\uff0c\u518d\u7ecf GFM \u5f97\u5230 (SS,AS,SA,AA)\uff0c\u6700\u540e concat \u540e\u7528 DGG \u8f93\u51fa (F<\/em>)\u3002\u8fd9\u76f8\u5f53\u4e8e\u628a\u591a\u5c3a\u5ea6\u878d\u5408\u62c6\u6210\u201c\u6bcf\u5c3a\u5ea6\u538b\u7f29\u2014\u8de8\u5c3a\u5ea6\u8f7b\u91cf\u4ea4\u4e92\u2014\u52a8\u6001\u5206\u7ec4\u805a\u5408\u201d\u3002<\/p>\n<p><strong>\u6a21\u5757 C\uff1aInterleaved Shuffle Fusion (ISF)<\/strong><br \/>\n\u5c06\u8f93\u5165 feature \u6309 channel \u5206\u6210 (G=4) \u7ec4\uff0c\u505a deterministic channel shuffle\uff1a<\/p>\n<p>[<br \/>\n\\hat{F}=Shuffle_G(F),<br \/>\n]<\/p>\n<p>\u518d\u7528 depthwise convolution \u5f97\u5230 (U)\uff0c\u5e76\u901a\u8fc7 group-wise learnable scale (\\gamma) \u6b8b\u5dee\u76f8\u52a0\uff1a<\/p>\n<p>[<br \/>\nF'=F+expand(\\gamma)\\odot U.<br \/>\n]<\/p>\n<p>\u4f5c\u7528\u662f\u4f4e\u6210\u672c\u8de8 group \/ \u8de8\u5c3a\u5ea6\u4ea4\u4e92\uff0c\u51e0\u4e4e\u4e0d\u589e\u52a0\u53c2\u6570\u3002<\/p>\n<p><strong>\u6a21\u5757 D\uff1aDynamic Group Gating (DGG)<\/strong><br \/>\n\u5c06\u8f93\u5165\u6309 channel \u5206\u7ec4\u4e3a (\\tilde{X}\\in\\mathbb{R}^{B\\times G\\times C_g\\times H\\times W})\uff0c\u5bf9 channel \u548c spatial \u5e73\u5747\u6c60\u5316\u5f97\u5230 group descriptor (z\\in\\mathbb{R}^{B\\times G})\uff0c\u7ebf\u6027\u5c42 + sigmoid \u5f97\u5230 gate\uff1a<\/p>\n<p>[<br \/>\nw=\\sigma(\\phi(z)),<br \/>\n\\quad<br \/>\n\\tilde{X}'=\\tilde{X}\\odot w^\\uparrow,<br \/>\n\\quad<br \/>\nX_{out}=X+Reshape(\\tilde{X}').<br \/>\n]<\/p>\n<p>\u4f5c\u7528\u662f\u8ba9\u4e0d\u540c feature group \u6309\u8f93\u5165\u5185\u5bb9\u52a8\u6001\u52a0\u6743\uff0c\u5bf9\u53cd\u5149\u3001\u6a21\u7cca\u3001\u5c0f\u606f\u8089\u7b49\u573a\u666f\u66f4\u7075\u6d3b\u3002<\/p>\n<p><strong>\u6a21\u5757 E\uff1aPseudo-depth-guided multi-task learning<\/strong><br \/>\nDepth-Anything v2-small \u751f\u6210\u76f8\u5bf9\u6df1\u5ea6 (D^*)\uff0c\u6a21\u578b depth head \u9884\u6d4b (D)\uff0c\u7528 Smooth-L1 \u7ea6\u675f\u3002\u6df1\u5ea6\u76d1\u7763\u4ec5\u8bad\u7ec3\u671f\u5b58\u5728\uff0c\u63a8\u7406\u65f6\u65e0\u989d\u5916\u5f00\u9500\u3002\u4f5c\u8005\u7684\u5173\u952e\u5047\u8bbe\u662f\uff1a\u76f8\u5bf9 depth\/geometry cue \u6bd4 RGB appearance \u66f4\u4e0d\u53d7\u53cd\u5149\u548c\u6a21\u7cca\u5f71\u54cd\uff0c\u56e0\u6b64\u80fd\u4fc3\u4f7f\u5171\u4eab\u7279\u5f81\u5b66\u5230\u66f4\u7a33\u5b9a\u7684\u7ed3\u6784\u8868\u5f81\u3002<\/p>\n<p><strong>\u662f\u5426\u9002\u5408\u8fc1\u79fb<\/strong>\uff1a<br \/>\n- \u5bf9 polyp segmentation\uff1a\u975e\u5e38\u9002\u5408\uff0c\u4efb\u52a1\u4e0e\u6570\u636e\u76f4\u63a5\u5339\u914d\u3002<br \/>\n- \u5bf9 3D medical image segmentation\uff1apseudo-depth \u673a\u5236\u4e0d\u76f4\u63a5\u9002\u7528\uff1b\u4f46\u201c\u8bad\u7ec3\u671f\u8f85\u52a9\u7ed3\u6784\u76d1\u7763\u3001\u63a8\u7406\u671f\u96f6\u5f00\u9500\u201d\u7684\u601d\u60f3\u53ef\u8fc1\u79fb\u4e3a signed distance map\u3001boundary distance transform\u3001surface\/centerline \u8f85\u52a9\u76d1\u7763\u3002<br \/>\n- \u5bf9 DAMamba \/ Transformer-based segmentation\uff1a\u53ef\u628a depth\/\u7ed3\u6784\u8f85\u52a9 loss \u4f5c\u4e3a\u8bad\u7ec3\u6b63\u5219\uff0c\u4e0e Mamba\/Transformer backbone \u89e3\u8026\u3002<\/p>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p>\u6570\u636e\u96c6\uff1a<\/p>\n<ul>\n<li>Kvasir-SEG\uff1a1000 \u5f20\uff0c\u9ad8\u8d28\u91cf\u606f\u8089\u56fe\u50cf\uff0c80\/20 train\/val\u3002<\/li>\n<li>CVC-ClinicDB\uff1a612 \u5f20\uff0c\u7528\u4e8e OOD validation\u3002<\/li>\n<li>CVC-ColonDB\uff1a380 \u5f20\uff0c\u7528\u4e8e OOD validation\u3002<\/li>\n<li>PolypGen sequences 18\u201322\uff1a273 \u5f20\uff0c\u771f\u5b9e\u624b\u672f\u573a\u666f\uff0c\u5305\u542b blur \u548c reflection artifacts\u3002<\/li>\n<\/ul>\n<p>\u9000\u5316\u534f\u8bae\uff1a<\/p>\n<ul>\n<li>synthetic degradation \u5305\u62ec motion blur\u3001Gaussian blur\u3001brightness\/contrast\u3001JPEG compression\u3001light spots\/reflection\u3001fog\u3001optical distortion\u3002<\/li>\n<li>\u56db\u8c61\u9650\u8bbe\u7f6e\uff1aClean\u2192Clean\u3001Clean\u2192Noisy\u3001Noisy\u2192Clean\u3001Noisy\u2192Noisy\u3002<\/li>\n<li>PolypGen \u4f7f\u7528\u539f\u59cb\u771f\u5b9e\u9000\u5316\uff0c\u4e0d\u505a synthetic augmentation\u3002<\/li>\n<\/ul>\n<p>\u8bad\u7ec3\u548c\u5b9e\u73b0\uff1a<\/p>\n<ul>\n<li>encoder\uff1aMiT-B0\u3002<\/li>\n<li>\u8f93\u5165\uff1a224\u00d7224\u3002<\/li>\n<li>optimizer\uff1aAdamW\uff0clr=1e-4\uff0cweight decay=1e-4\u3002<\/li>\n<li>\u8bad\u7ec3 200 epochs\uff0c\u524d 10% warm-up\uff0ccosine annealing\u3002<\/li>\n<li>batch size 16\uff0cNVIDIA A100\u3002<\/li>\n<li>pseudo-depth\uff1aDepth-Anything v2-small\u3002<\/li>\n<li>\u63a8\u7406\u5e73\u53f0\uff1aRTX 3090\u3001Apple iPhone 15 CoreML FP16\u3001Raspberry Pi 4\u3002<\/li>\n<\/ul>\n<p>\u5173\u952e\u7ed3\u679c\uff1a<\/p>\n<ul>\n<li>\u56db\u8c61\u9650\u9c81\u68d2\u5206\u6790\uff1aDepthPolyp Clean\u2192Noisy Dice \u4e3a 0.8126\uff0c\u9ad8\u4e8e UNet 0.6478\u3001SegFormer-B0 0.6962\u3001PraNet 0.7143\u3001CFFormer 0.7556\uff1bNoisy\u2192Noisy Dice \u4e3a 0.8525\u3002<\/li>\n<li>noisy training \u7684 clean-domain penalty \u8f83\u5c0f\uff1aDepthPolyp Noisy\u2192Clean \u76f8\u5bf9 Clean\u2192Clean \u4e3a -0.0197 Dice\uff0c\u8bf4\u660e\u52a0\u5165\u9000\u5316\u8bad\u7ec3\u6ca1\u6709\u4e25\u91cd\u727a\u7272 clean \u6027\u80fd\u3002<\/li>\n<li>cross-dataset noisy training\uff1aDepthPolyp \u5728 Kvasir\u3001ClinicDB\u3001ColonDB \u7684 N\u2192N Dice \u5206\u522b\u4e3a 0.853\u30010.751\u30010.734\uff0c\u4f18\u4e8e SegFormer-B0 \u7684 0.823\u30010.698\u30010.621\uff1b\u53c2\u6570 3.57M\u3001GMACs 0.86\uff0c\u4e5f\u4f4e\u4e8e SegFormer-B0 \u7684 3.71M\u30011.30 GMACs\u3002<\/li>\n<li>PolypGen\uff1aDepthPolyp PolypGen Dice 0.679\uff0c\u9ad8\u4e8e SegFormer-B0 0.634\u3001CFFormer 0.643\u3001SegFormer-B5 0.671\uff1biPhone 15 \u8fbe 181.54 FPS\uff0c\u63a5\u8fd1 SegFormer-B0 186.72 FPS\uff0c\u4f46 PolypGen Dice \u66f4\u9ad8\u3002<\/li>\n<li>Ablation\uff1afull model Avg Dice 0.784\uff1bw\/o depth guidance \u4e3a 0.759\uff1bw\/o uncertainty loss \u4e3a 0.605\uff1bw\/o GFM \u4e3a 0.776 \u4f46 iPhone FPS \u4ece 181.54 \u964d\u5230 131.39\uff1bw\/o ISF \u4e3a 0.760\uff1bw\/o DGG \u4e3a 0.736\u3002<\/li>\n<\/ul>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u53ef\u4fe1\u70b9\uff1a<\/p>\n<ul>\n<li>\u8bba\u6587\u6ca1\u6709\u53ea\u62a5 clean Kvasir\/ClinicDB\uff0c\u800c\u662f\u663e\u5f0f\u6784\u9020\u56db\u8c61\u9650\u9c81\u68d2\u8bc4\u4f30\uff0c\u80fd\u56de\u7b54\u201cclean \u8bad\u7ec3\u5728 noisy \u6d4b\u8bd5\u4e0b\u4f1a\u600e\u6837\u201d\u201cnoisy \u8bad\u7ec3\u662f\u5426\u4f24\u5bb3 clean \u6027\u80fd\u201d\u3002<\/li>\n<li>\u5305\u542b\u771f\u5b9e PolypGen sequences 18\u201322\uff0c\u800c\u4e0d\u53ea\u662f synthetic degradation\u3002<\/li>\n<li>baseline \u8986\u76d6 heavy\u3001mid-size\u3001lightweight \u4e09\u7c7b\uff0c\u5171 19 \u4e2a\u4ee3\u8868\u6a21\u578b\uff1b\u540c\u65f6\u62a5\u544a\u53c2\u6570\u3001GMACs \u548c\u591a\u5e73\u53f0 FPS\u3002<\/li>\n<li>ablation \u6bd4\u8f83\u5b8c\u6574\uff0c\u80fd\u533a\u5206 pseudo-depth\u3001uncertainty loss\u3001GFM\u3001ISF\u3001DGG \u5bf9\u51c6\u786e\u7387\u548c\u901f\u5ea6\u7684\u8d21\u732e\u3002<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\u5728\u8bba\u6587\u4e2d\u7ed9\u51fa\uff0c\u7406\u8bba\u4e0a\u6709\u5229\u4e8e\u590d\u73b0\u3002<\/li>\n<\/ul>\n<p>\u4e3b\u8981\u98ce\u9669\uff1a<\/p>\n<ul>\n<li>pseudo-depth target \u7531 Depth-Anything v2-small \u751f\u6210\uff0c\u5176\u5728\u5185\u955c\u56fe\u50cf\u4e0a\u7684\u6df1\u5ea6\u53ef\u9760\u6027\u6ca1\u6709\u5355\u72ec\u5b9a\u91cf\u9a8c\u8bc1\uff1b\u5b83\u53ef\u80fd\u53ea\u662f\u63d0\u4f9b\u4e00\u79cd regularization signal\uff0c\u800c\u4e0d\u662f\u771f\u6b63\u51e0\u4f55\u6df1\u5ea6\u3002<\/li>\n<li>\u201csynthetic degradation accurately replicates real conditions\u201d\u7684\u8bf4\u6cd5\u9700\u8981\u8c28\u614e\u3002\u867d\u7136 PolypGen \u9a8c\u8bc1\u589e\u5f3a\u4e86\u8bf4\u670d\u529b\uff0c\u4f46 synthetic blur\/reflection\/fog \u4e0e\u771f\u5b9e\u5185\u955c\u9000\u5316\u5206\u5e03\u4ecd\u53ef\u80fd\u6709\u5dee\u8ddd\u3002<\/li>\n<li>\u8bad\u7ec3\u548c\u6d4b\u8bd5\u90fd\u56f4\u7ed5\u606f\u8089\u5206\u5272\uff0c\u7ed3\u8bba\u4e0d\u5e94\u6cdb\u5316\u5230\u6240\u6709\u533b\u5b66\u56fe\u50cf\u5206\u5272\u3002<\/li>\n<li>w\/o uncertainty loss \u6027\u80fd\u5d29\u5f97\u5f88\u5389\u5bb3\uff080.784\u21920.605\uff09\uff0c\u63d0\u793a\u591a\u4efb\u52a1\u6743\u91cd\u5bf9\u8bad\u7ec3\u975e\u5e38\u654f\u611f\uff1b\u5982\u679c\u590d\u73b0\u65f6 loss \u6743\u91cd\u3001pseudo-depth normalization \u6216\u8bad\u7ec3 schedule \u4e0d\u4e00\u81f4\uff0c\u7ed3\u679c\u53ef\u80fd\u4e0d\u7a33\u5b9a\u3002<\/li>\n<li>\u6ca1\u6709\u62a5\u544a\u7edf\u8ba1\u663e\u8457\u6027\u6216\u591a seed \u65b9\u5dee\uff0c\u9c81\u68d2\u63d0\u5347\u867d\u7136\u5e45\u5ea6\u8f83\u5927\uff0c\u4f46\u590d\u73b0\u65f6\u4ecd\u9700\u68c0\u67e5\u65b9\u5dee\u3002<\/li>\n<\/ul>\n<h4>8. \u4e0e\u4e3b\u6d41\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u7684\u5173\u7cfb<\/h4>\n<ul>\n<li><strong>U-Net \/ nnU-Net<\/strong>\uff1aDepthPolyp \u4e0d\u662f nnU-Net pipeline\uff0c\u4e5f\u4e0d\u662f\u901a\u7528\u81ea\u52a8\u914d\u7f6e\u6846\u67b6\uff1b\u5b83\u66f4\u50cf\u4e00\u4e2a\u8f7b\u91cf encoder-decoder + \u591a\u4efb\u52a1\u8bad\u7ec3\u7b56\u7565\u3002GFM\/ISF\/DGG \u53ef\u8fc1\u79fb\u5230 U-Net decoder\u3002<\/li>\n<li><strong>MedNeXt \/ CNN-based segmentation<\/strong>\uff1aGFM\/ISF\/DGG \u5ef6\u7eed\u8f7b\u91cf CNN\/\u5377\u79ef\u8c03\u5236\u8def\u7ebf\uff0c\u5f3a\u8c03\u9ad8\u6548\u5c40\u90e8\u4e0e\u8de8\u5c3a\u5ea6\u878d\u5408\u3002<\/li>\n<li><strong>UNetR \/ Swin-UNet \/ TransUNet \/ TransFuse \/ SegFormer<\/strong>\uff1a\u4f7f\u7528 MiT-B0\/SegFormer-style encoder\uff0c\u4f46\u8d21\u732e\u4e3b\u8981\u5728\u8f7b\u91cf decoder \u548c pseudo-depth multi-task learning\uff1b\u5b9e\u9a8c\u4e2d\u4e5f\u628a SegFormer-B0\/B5 \u4f5c\u4e3a\u91cd\u8981\u5bf9\u6bd4\u3002<\/li>\n<li><strong>Mamba \/ VMamba \/ SegMamba \/ DAMamba<\/strong>\uff1a\u6ca1\u6709\u4f7f\u7528 Mamba\uff0c\u4f46\u5176\u8bad\u7ec3\u671f\u7ed3\u6784\u76d1\u7763\u548c\u9000\u5316\u9c81\u68d2\u8bc4\u4f30\u53ef\u4e0e Mamba backbone \u7ed3\u5408\u3002\u5bf9 DAMamba \u7528\u6237\u800c\u8a00\uff0c\u5b83\u63d0\u4f9b\u7684\u662f\u201crobustness protocol + auxiliary geometry supervision\u201d\uff0c\u4e0d\u662f SSM \u6a21\u5757\u3002<\/li>\n<li><strong>Foundation model for medical segmentation \/ MedSAM<\/strong>\uff1a\u6ca1\u6709\u4f7f\u7528 SAM\/MedSAM\u3002\u5b83\u4f7f\u7528 Depth-Anything v2 \u4f5c\u4e3a frozen pseudo-label generator\uff0c\u66f4\u63a5\u8fd1\u201cfoundation depth model \u63d0\u4f9b\u8f85\u52a9\u76d1\u7763\u201d\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u5bf9 polyp segmentation\uff1a\u975e\u5e38\u9ad8\u3002\u5b83\u76f4\u63a5\u9488\u5bf9\u771f\u5b9e\u5185\u955c\u9000\u5316\u3001PolypGen\u3001\u79fb\u52a8\u7aef FPS \u548c\u8f7b\u91cf\u6a21\u578b\uff0c\u9002\u5408\u4f5c\u4e3a\u7528\u6237\u540e\u7eed polyp segmentation \u7814\u7a76\u7684 baseline \u6216\u5b9e\u9a8c\u534f\u8bae\u53c2\u8003\u3002<\/p>\n<p>\u5bf9 DAMamba \u6539\u9020\uff1a\u4e2d\u9ad8\u3002\u867d\u7136\u6ca1\u6709 Mamba \u6a21\u5757\uff0c\u4f46\u53ef\u628a pseudo-depth auxiliary head \u6216 uncertainty-weighted multi-task loss \u63a5\u5165 DAMamba\uff0c\u7528\u4e8e\u6d4b\u8bd5 Mamba backbone \u5728 noisy colonoscopy \u4e0b\u662f\u5426\u66f4\u7a33\u3002\u66f4\u91cd\u8981\u7684\u662f\u56db\u8c61\u9650\u8bc4\u4f30\u534f\u8bae\u503c\u5f97\u76f4\u63a5\u590d\u7528\u3002<\/p>\n<p>\u5bf9\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u9009\u62e9\uff1a\u5b83\u63d0\u9192\u4e0d\u8981\u53ea\u770b clean Dice\u3002\u82e5\u7528\u6237\u8981\u5199 introduction\/related work\uff0c\u53ef\u7528\u5b83\u652f\u6301\u201cclinical deployment requires robustness to realistic degradations and edge efficiency\u201d\u7684\u8bba\u70b9\u3002<\/p>\n<p>\u5bf9 3D medical segmentation\uff1a\u6a21\u5757\u672c\u8eab\u4e0d\u662f\u76f4\u63a5\u9488\u5bf9 3D\uff1b\u4f46\u8bad\u7ec3\u671f\u8f85\u52a9\u7ed3\u6784\u76d1\u7763\u601d\u60f3\u53ef\u8fc1\u79fb\u5230 3D distance transform\u3001surface-aware loss\u3001boundary-aware auxiliary head\u3002<\/p>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5f3a\u70c8\u5efa\u8bae\u7cbe\u8bfb\u3002<\/strong> \u5982\u679c\u7528\u6237\u5173\u6ce8 polyp segmentation \u6216\u771f\u5b9e\u4e34\u5e8a\u90e8\u7f72\uff0c\u8fd9\u7bc7\u6bd4\u5355\u7eaf\u5806 Mamba\/Transformer \u6a21\u5757\u7684\u8bba\u6587\u66f4\u503c\u5f97\u8bfb\u3002\u5efa\u8bae\u4f18\u5148\u590d\u73b0\u5176\u56db\u8c61\u9650 noisy evaluation \u548c pseudo-depth\/uncertainty loss ablation\uff0c\u800c\u4e0d\u662f\u4e00\u5f00\u59cb\u5b8c\u6574\u590d\u73b0\u6240\u6709\u8f7b\u91cf decoder \u7ec6\u8282\u3002<\/p>\n<hr \/>\n<h2>\u4eca\u65e5\u63a8\u8350\u4f18\u5148\u7ea7<\/h2>\n<ol>\n<li><strong>DepthPolyp<\/strong>\uff1a\u6700\u503c\u5f97\u4f18\u5148\u6df1\u5165\u8bfb\u3002\u539f\u56e0\u662f\u95ee\u9898\u5b9a\u4e49\u66f4\u8d34\u8fd1\u771f\u5b9e\u4e34\u5e8a\u90e8\u7f72\uff0c\u5b9e\u9a8c\u534f\u8bae\u66f4\u5b8c\u6574\uff0c\u5305\u542b PolypGen \u771f\u5b9e\u9000\u5316\u3001\u8de8\u6570\u636e\u96c6\u3001\u8f7b\u91cf\u901f\u5ea6\u548c\u7cfb\u7edf\u6d88\u878d\uff1b\u5bf9 polyp segmentation \u8bfe\u9898\u76f4\u63a5\u4ef7\u503c\u6700\u9ad8\u3002<\/li>\n<li><strong>Patch-MoE Mamba<\/strong>\uff1a\u9002\u5408\u505a Mamba\/VM-UNet\/DAMamba \u7ed3\u6784\u6539\u9020\u53c2\u8003\u3002\u4f18\u5148\u501f\u9274 patch-ordered scanning\uff1b\u5b8c\u6574 MoE \u7248\u672c\u56e0\u590d\u6742\u5ea6\u8f83\u9ad8\uff0c\u66f4\u9002\u5408\u4f5c\u4e3a idea source \u800c\u975e\u76f4\u63a5\u90e8\u7f72\u6846\u67b6\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\u8def\u5f84 MEDIA:\/root\/papers_medseg_2026-05-23\/2605.17719.pdf\uff1bPDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.17719<\/li>\n<li>\u8bba\u6587 2\uff1a\u5df2\u9644 PDF\uff1b\u672c\u5730\u8def\u5f84 MEDIA:\/root\/papers_medseg_2026-05-23\/2605.16519.pdf\uff1bPDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.16519<\/li>\n<\/ul>\n<h2>\u4eca\u65e5\u53ef\u6267\u884c\u5efa\u8bae<\/h2>\n<ol>\n<li>\u5148\u7cbe\u8bfb\u5e76\u590d\u73b0 <strong>DepthPolyp \u7684\u56db\u8c61\u9650\u9c81\u68d2\u8bc4\u4f30\u534f\u8bae<\/strong>\uff1aClean\u2192Clean\u3001Clean\u2192Noisy\u3001Noisy\u2192Clean\u3001Noisy\u2192Noisy\u3002\u8fd9\u4e2a\u534f\u8bae\u53ef\u4ee5\u76f4\u63a5\u7528\u6765\u8bc4\u4f30\u4f60\u7684 polyp segmentation \u6216 DAMamba \u6539\u9020\u662f\u5426\u771f\u7684\u63d0\u5347\u771f\u5b9e\u9000\u5316\u9c81\u68d2\u6027\u3002<\/li>\n<li>\u5bf9 DAMamba\/Mamba-UNet \u6539\u9020\uff0c\u4f18\u5148\u5c1d\u8bd5 <strong>Patch-MoE Mamba \u7684 patch-ordered scanning<\/strong>\uff0c\u5148\u4e0d\u8981\u5b8c\u6574\u5f15\u5165 MoE concat expert\uff1b\u56e0\u4e3a\u8bba\u6587\u663e\u793a\u4e3b\u8981\u6536\u76ca\u5df2\u6765\u81ea patch scanning\uff0c\u800c\u5b8c\u6574 MoE \u5e26\u6765 70.06M \u53c2\u6570\u548c 28.18G FLOPs\u3002<\/li>\n<li>related work \u53ef\u4ee5\u52a0\u5165\u8fd9\u4e24\u7bc7\uff1aDepthPolyp \u653e\u5728 robust\/lightweight polyp segmentation \u4e0e deployment-focused evaluation\uff1bPatch-MoE Mamba \u653e\u5728 Mamba-based medical segmentation \u4e2d\u5173\u4e8e scan order \u548c directional fusion \u7684\u6700\u65b0\u6539\u9020\u3002<\/li>\n<\/ol>\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\u4ece 2026 \u5e74 5 \u6708\u4e2d\u4e0b\u65ec\u7684 arXiv \u533b\u5b66\u56fe\u50cf\u5206\u5272\u65b0\u7a3f\u4e2d\u7b5b\u5230 2 \u7bc7\u503c\u5f97\u5173 &#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-1073","post","type-post","status-publish","format-standard","hentry","category-85"],"views":16,"_links":{"self":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1073","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=1073"}],"version-history":[{"count":0,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1073\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/media?parent=1073"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/categories?post=1073"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/tags?post=1073"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}