{"id":1077,"date":"2026-05-26T08:34:58","date_gmt":"2026-05-26T00:34:58","guid":{"rendered":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/26\/2026-05-26-%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%9amamba-based-segmentation-%e6%9c%80%e6%96%b0%e8%bf%9b%e5%b1%95\/"},"modified":"2026-05-26T08:34:58","modified_gmt":"2026-05-26T00:34:58","slug":"2026-05-26-%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%9amamba-based-segmentation-%e6%9c%80%e6%96%b0%e8%bf%9b%e5%b1%95","status":"publish","type":"post","link":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/26\/2026-05-26-%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%9amamba-based-segmentation-%e6%9c%80%e6%96%b0%e8%bf%9b%e5%b1%95\/","title":{"rendered":"2026-05-26 \u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\u7cbe\u8bfb\uff1aMamba-based segmentation \u6700\u65b0\u8fdb\u5c55"},"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\u6ca1\u6709\u68c0\u7d22\u5230\u5df2\u7ecf\u660e\u786e\u8fdb\u5165 MICCAI\/CVPR\/ICCV\/ECCV\/NeurIPS\/ICLR\/AAAI\/IJCAI\/ISBI \u6216\u9876\u520a\u6b63\u5f0f\u63a5\u6536\u7684\u5168\u65b0\u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\uff1b\u56e0\u6b64\u6309\u4efb\u52a1\u8981\u6c42\u5411 2026 \u5e74 5 \u6708\u4e2d\u65ec arXiv preprint \u56de\u6eaf\u7b5b\u9009\u3002\u4eca\u65e5\u6700\u503c\u5f97\u5173\u6ce8\u7684\u8d8b\u52bf\u662f\uff1aMamba\/State Space Model \u4ecd\u5728\u533b\u5b66\u5206\u5272\u4e2d\u6301\u7eed\u88ab\u6539\u9020\uff0c\u8fd1\u671f\u8bba\u6587\u4e3b\u8981\u56f4\u7ed5\u201c\u5982\u4f55\u628a 2D\/3D \u7a7a\u95f4\u7ed3\u6784\u66f4\u5408\u7406\u5730\u5e8f\u5217\u5316\u201d\u548c\u201c\u5982\u4f55\u8ba9\u591a\u65b9\u5411\/\u591a\u5934 SSM \u8f93\u51fa\u66f4\u7a33\u5b9a\u5730\u878d\u5408\u201d\u3002<\/p>\n<h2>\u68c0\u7d22\u8bf4\u660e<\/h2>\n<p>\u68c0\u7d22\u8303\u56f4\u8986\u76d6 arXiv \u6700\u65b0\u63d0\u4ea4\uff0c\u5e76\u4ee5 medical image segmentation\u3001Mamba medical image segmentation\u30013D brain tumor segmentation\u3001polyp segmentation\u3001universal medical segmentation \u7b49\u5173\u952e\u8bcd\u5411 2025 \u5e74\u4ee5\u540e\u56de\u6eaf\uff1b\u540c\u65f6\u68c0\u67e5\u4e86\u672c\u5730\u53ef\u7528\u7684 cron\/session \u5386\u53f2\u6587\u4ef6\u8def\u5f84\uff0c\u4f46\u5f53\u524d\u73af\u5883\u672a\u53d1\u73b0 <code>session_cron_def34ee3de23_*<\/code> \u6216\u53ef\u68c0\u7d22\u7684\u5386\u53f2\u8f93\u51fa\u6587\u4ef6\uff0c\u56e0\u6b64\u672c\u6b21\u4ee5\u53ef\u7528\u5386\u53f2\u8bb0\u5f55\u4e3a\u7a7a\u8fdb\u884c\u53bb\u91cd\u3002\u6240\u6709\u5165\u9009\u8bba\u6587\u5747\u4e3a 2025 \u5e74\u53ca\u4ee5\u540e\uff0c\u4e14\u5747\u4e3a arXiv preprint\uff0c\u5c1a\u672a\u786e\u8ba4\u9876\u4f1a\/\u9876\u520a\u6b63\u5f0f\u63a5\u6536\u3002\u5df2\u68c0\u67e5\u5386\u53f2\u63a8\u8350\u8bb0\u5f55\u5e76\u6392\u9664\u4e86\u91cd\u590d\u8bba\u6587\uff1b\u672c\u6b21\u6ca1\u6709\u53d1\u73b0\u53ef\u786e\u8ba4\u7684\u5386\u53f2\u91cd\u590d\u5019\u9009\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, Fabian Vazquez, Jose A. Nu\u00f1ez, Huimin Li, Jinghao Yang, Erik Enriquez, DongChul Kim, Haoteng Tang, Bin Fu, Pengfei Gu \/ \u7b2c\u4e00\u4f5c\u8005 Diego Adame<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-18\uff08arXiv v1\uff09<\/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\uff1ahttps:\/\/arxiv.org\/pdf\/2605.17719<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1a\u672a\u83b7\u53d6 \/ arXiv \u9875\u9762\u4e0e\u8bba\u6587\u6b63\u6587\u672a\u7ed9\u51fa\u660e\u786e\u4ee3\u7801\u4ed3\u5e93<\/li>\n<li>\u4efb\u52a1\uff1a2D medical image segmentation\uff1b\u4e3b\u8981\u662f\u606f\u8089\u5206\u5272\uff0c\u53e6\u542b\u76ae\u80a4\u75c5\u7076\u5206\u5272<\/li>\n<li>\u6570\u636e\u96c6\uff1aKvasir-SEG\u3001ClinicDB\u3001ColonDB\u3001ETIS\u3001CVC-300\uff1bISIC 2017\u3001ISIC 2018<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1aU-Net-style encoder-decoder + VMamba\/VM-UNetV2 \u6539\u9020\uff1bpatch-ordered scanning + MoE directional fusion \u7684 Mamba-based segmentation framework<\/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\u70b9\u4e0d\u662f\u201c\u53c8\u505a\u4e86\u4e00\u4e2a Mamba U-Net\u201d\uff0c\u800c\u662f\u628a Vision Mamba \u4e2d\u5bb9\u6613\u7834\u574f 2D \u90bb\u57df\u5173\u7cfb\u7684\u50cf\u7d20\u7ea7\u626b\u63cf\uff0c\u6539\u6210 patch-ordered scanning\uff0c\u5e76\u7528\u7a7a\u95f4\u81ea\u9002\u5e94 MoE \u66ff\u4ee3\u7b80\u5355\u65b9\u5411\u6c42\u548c\uff1b\u5bf9 polyp segmentation \u7684 DAMamba\/VM-UNet \u7cfb\u5217\u6539\u9020\u6709\u76f4\u63a5\u53c2\u8003\u4ef7\u503c\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u533b\u5b66\u56fe\u50cf\u5206\u5272\u540c\u65f6\u9700\u8981\u5c40\u90e8\u8fb9\u754c\u7ec6\u8282\u548c\u8f83\u5927\u8303\u56f4\u4e0a\u4e0b\u6587\uff1a\u606f\u8089\u8fb9\u754c\u5e38\u4f4e\u5bf9\u6bd4\u3001\u5f62\u72b6\u4e0d\u89c4\u5219\uff0c\u76ae\u80a4\u75c5\u7076\u4e5f\u5b58\u5728\u5c3a\u5ea6\u548c\u7eb9\u7406\u53d8\u5316\u3002CNN\/U-Net \u7684\u5c40\u90e8\u5f52\u7eb3\u504f\u7f6e\u6709\u5229\u4e8e\u8fb9\u754c\uff0c\u4f46\u5168\u5c40\u4f9d\u8d56\u5efa\u6a21\u5f31\uff1bTransformer \u53ef\u4ee5\u5efa\u6a21\u5168\u5c40\u5173\u7cfb\uff0c\u4f46\u6807\u51c6 self-attention \u5728\u9ad8\u5206\u8fa8\u7387 dense prediction \u4e2d\u6210\u672c\u9ad8\u3002Mamba\/SSM \u63d0\u4f9b\u7ebf\u6027\u590d\u6742\u5ea6\u7684\u957f\u7a0b\u5efa\u6a21\uff0c\u56e0\u6b64\u8fd1\u5e74\u88ab\u7528\u4e8e VM-UNet\u3001VM-UNetV2 \u7b49\u533b\u5b66\u5206\u5272\u7f51\u7edc\u3002<\/p>\n<p>\u4f5c\u8005\u6293\u4f4f\u7684\u6838\u5fc3\u95ee\u9898\u662f\uff1a\u73b0\u6709 Vision Mamba \u901a\u5e38\u628a 2D feature map \u76f4\u63a5\u5c55\u5e73\u6210 1D \u5e8f\u5217\uff0c\u5e76\u6cbf\u56fa\u5b9a\u65b9\u5411\u626b\u63cf\u3002\u8fd9\u6837\u4f1a\u8ba9\u7a7a\u95f4\u76f8\u90bb\u50cf\u7d20\u5728\u5e8f\u5217\u4e2d\u76f8\u8ddd\u5f88\u8fdc\uff0c\u6216\u8ba9\u5e8f\u5217\u76f8\u90bb token \u5728\u56fe\u50cf\u5e73\u9762\u4e0a\u5e76\u4e0d\u76f8\u90bb\u3002\u5bf9 segmentation \u6765\u8bf4\uff0c\u8fd9\u53ef\u80fd\u635f\u4f24\u8fb9\u754c\u548c\u5c0f\u76ee\u6807\u7684\u5c40\u90e8\u4e00\u81f4\u6027\u3002\u53e6\u4e00\u4e2a\u95ee\u9898\u662f\u591a\u65b9\u5411\u626b\u63cf\u8f93\u51fa\u901a\u5e38\u76f4\u63a5\u76f8\u52a0\uff0c\u65e0\u6cd5\u6839\u636e\u5c40\u90e8\u533a\u57df\u7684\u5c3a\u5ea6\u3001\u65b9\u5411\u3001\u8fb9\u754c\u590d\u6742\u5ea6\u81ea\u9002\u5e94\u9009\u62e9\u66f4\u6709\u7528\u7684\u65b9\u5411\u54cd\u5e94\u3002<\/p>\n<p>\u5185\u90e8 paper map\uff1a\u8bba\u6587\u7814\u7a76\u533b\u5b66\u56fe\u50cf\u5206\u5272\u4e2d Vision Mamba \u7684\u7a7a\u95f4\u5e8f\u5217\u5316\u4e0e\u65b9\u5411\u878d\u5408\u95ee\u9898\uff0c\u8bbe\u7f6e\u4e3a 2D \u606f\u8089\u548c\u76ae\u80a4\u75c5\u7076\u5206\u5272\uff1b\u4e3b\u62db\u662f patch-ordered scanning + MoE-based directional fusion\uff1b\u58f0\u79f0\u5728 5 \u4e2a\u606f\u8089\u6570\u636e\u96c6\u548c 2 \u4e2a ISIC \u6570\u636e\u96c6\u4e0a\u4f18\u4e8e U-Net\u3001U-Net v2\u3001VM-UNet\u3001VM-UNetV2\uff1b\u5173\u952e\u5bf9\u8c61\u662f Patch-MoE VSS block\u3001\u56db\u65b9\u5411\u626b\u63cf\u3001\u4e94\u4e13\u5bb6\u878d\u5408\u3001spatial-aware router\u3001SDI skip enhancement\uff1b\u771f\u6b63\u8d1f\u91cd\u5728\u201c\u626b\u63cf\u987a\u5e8f\u662f\u5426\u786e\u5b9e\u4fdd\u7559\u5c40\u90e8\u7ed3\u6784\u201d\u548c\u201cMoE \u589e\u76ca\u662f\u5426\u62b5\u6d88\u989d\u5916\u8ba1\u7b97\u201d\uff1b\u4e3b\u8981\u5931\u8d25\u98ce\u9669\u662f\u5b9e\u9a8c\u53ea\u6bd4\u8f83\u5c11\u6570 baselines\uff0c\u4e14\u53c2\u6570\/FLOPs \u589e\u5e45\u5f88\u5927\u3002<\/p>\n<p>\u8def\u7531\u8bb0\u5f55\uff1aPrimary adapter = method-algorithm\uff1bSecondary adapter = benchmark-evaluation\uff1bEvidence packs = general\u3001experimental-eval\u3001ablation-and-mechanism-isolation\u3001reproducibility-and-compute\uff1bRoute confidence = \u9ad8\u3002\u539f\u56e0\uff1a\u8bba\u6587\u4e3b\u8981\u8d21\u732e\u662f\u53ef\u79fb\u690d\u7684\u6a21\u578b\u6a21\u5757\uff0c\u7ed3\u8bba\u4f9d\u8d56\u591a\u6570\u636e\u96c6\u5b9e\u9a8c\u3001\u6d88\u878d\u548c\u590d\u6742\u5ea6\u6bd4\u8f83\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u8ba4\u4e3a\u73b0\u6709\u65b9\u6cd5\u7684\u95ee\u9898\u5206\u4e24\u5c42\uff1a<\/p>\n<ol>\n<li>CNN\/Transformer \u5c42\u9762\uff1aCNN \u53d7\u5c40\u90e8\u611f\u53d7\u91ce\u9650\u5236\uff0c\u957f\u7a0b\u4f9d\u8d56\u4e0d\u8db3\uff1bTransformer \u867d\u7136\u5168\u5c40\uff0c\u4f46 quadratic complexity \u548c\u6807\u6ce8\u6570\u636e\u9700\u6c42\u8ba9\u533b\u5b66\u5206\u5272\u4e0d\u591f\u7ecf\u6d4e\u3002<\/li>\n<li>Mamba \u533b\u5b66\u5206\u5272\u5c42\u9762\uff1a<br \/>\n   - pixel-wise raster scan \u4f1a\u7834\u574f 2D \u5c40\u90e8\u7ed3\u6784\u3002\u8bba\u6587\u4e3e\u4f8b\u8bf4\u660e\uff0c\u5728 16\u00d716 grid \u4e2d\u5782\u76f4\u76f8\u90bb\u50cf\u7d20\u53ef\u80fd\u5728\u5e8f\u5217\u4e2d\u76f8\u9694 16 \u4e2a step\u3002<br \/>\n   - \u591a\u65b9\u5411\u8f93\u51fa\u7b80\u5355 summation \u5047\u8bbe\u6240\u6709\u65b9\u5411\u3001\u6240\u6709\u5c3a\u5ea6\u5728\u6240\u6709\u4f4d\u7f6e\u540c\u7b49\u91cd\u8981\uff0c\u800c\u533b\u5b66\u76ee\u6807\u7684\u5c3a\u5ea6\u3001\u5f62\u72b6\u548c\u8fb9\u754c\u590d\u6742\u5ea6\u660e\u663e\u4e0d\u540c\u3002<\/li>\n<\/ol>\n<p>\u8fd9\u4e2a\u95ee\u9898\u5b9a\u4e49\u5bf9 polyp segmentation \u5f88\u8d34\u5207\uff0c\u56e0\u4e3a\u606f\u8089\u8fb9\u754c\u3001\u4f4e\u5bf9\u6bd4\u533a\u57df\u3001\u5c0f\u76ee\u6807\u786e\u5b9e\u5bb9\u6613\u88ab\u4e0d\u5408\u7406\u7684\u5e8f\u5217\u5316\u548c\u7c97\u7cd9\u878d\u5408\u635f\u4f24\u3002<\/p>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>\u6574\u4f53\u6846\u67b6\u662f U-Net-style \u7ed3\u6784\uff1aMamba-based encoder + SDI module + decoder\u3002\u5b83\u4ee5 VM-UNetV2 \u4e3a\u57fa\u7840\uff0c\u66ff\u6362\u5176\u4e2d\u7684 VSS block \u4e3a Patch-MoE VSS block\uff1b\u968f\u540e\u91c7\u7528 U-Net v2 \u7684 Semantics and Detail Infusion\uff08SDI\uff09\u6a21\u5757\uff0c\u7528\u9ad8\u5c42\u8bed\u4e49\u4e0e\u4f4e\u5c42\u7ec6\u8282\u901a\u8fc7 Hadamard product \u5f3a\u5316\u591a\u5c3a\u5ea6\u7279\u5f81\uff1bdecoder \u57fa\u672c\u6cbf\u7528 VM-UNetV2\u3002<\/p>\n<p>\u6838\u5fc3\u6d41\u7a0b\uff1a<\/p>\n<ol>\n<li>\u8f93\u5165\u56fe\u50cf\u7ecf encoder \u63d0\u53d6\u591a\u5c3a\u5ea6 feature map <code>X_l \u2208 R^{C_l \u00d7 H_l \u00d7 W_l}<\/code>\u3002<\/li>\n<li>\u5728\u6bcf\u4e2a Patch-MoE VSS block \u4e2d\uff0c\u4e0d\u6309\u666e\u901a raster order \u76f4\u63a5\u9010\u50cf\u7d20\u626b\u63cf\uff0c\u800c\u5148\u628a feature map \u6309 patch size <code>p<\/code> \u5206\u6210\u5c40\u90e8 patch\u3002<\/li>\n<li>\u6bcf\u4e2a patch \u5185\u90e8\u7528 row-major \u987a\u5e8f\u679a\u4e3e\u50cf\u7d20\uff0c\u518d\u8fdb\u5165\u4e0b\u4e00\u4e2a patch\uff0c\u7531\u6b64\u5f97\u5230\u4e00\u4e2a permutation vector\uff1b\u5b83\u4e0d\u51cf\u5c11 token \u6570\uff0c\u53ea\u6539\u53d8\u8bbf\u95ee\u987a\u5e8f\u3002<\/li>\n<li>\u4f7f\u7528\u56db\u4e2a\u65b9\u5411\u626b\u63cf\uff1aforward\u3001reverse\u3001width-height forward\u3001width-height reverse\uff1b\u4e0d\u540c\u65b9\u5411\u53ef\u5bf9\u5e94\u4e0d\u540c patch size\uff0c\u4ee5\u5f15\u5165\u591a\u5c3a\u5ea6\/\u5404\u5411\u5f02\u6027\u626b\u63cf\u3002<\/li>\n<li>\u56db\u4e2a directional Mamba outputs \u7ecf GroupNorm \u540e\u4f5c\u4e3a 4 \u4e2a expert\uff1b\u518d\u628a\u56db\u4e2a\u65b9\u5411 concat \u540e\u7ecf 1\u00d71 conv + BN + ReLU \u5f62\u6210\u7b2c 5 \u4e2a concat expert\u3002<\/li>\n<li>spatial-aware router \u540c\u65f6\u5229\u7528 local descriptor\uff08\u65b9\u5411\u8f93\u51fa\u6c42\u548c\u540e depthwise 3\u00d73 conv\uff09\u548c global descriptor\uff08\u65b9\u5411\u8f93\u51fa\u6c42\u548c\u540e GAP\uff0c\u518d broadcast\uff09\uff0c\u7528\u53ef\u5b66\u4e60 <code>\u03b1<\/code> \u6df7\u5408\u540e\u901a\u8fc7\u4e24\u5c42 1\u00d71 conv \u751f\u6210\u6bcf\u4e2a\u7a7a\u95f4\u4f4d\u7f6e\u7684 5 \u4e2a expert \u6743\u91cd\u3002<\/li>\n<li>\u4e94\u4e2a expert \u6309 softmax \u6743\u91cd\u52a0\u6743\u6c42\u548c\uff0c\u6700\u540e\u518d\u52a0\u56de\u56db\u4e2a raw directional outputs \u7684 residual sum\uff0c\u4ee5\u907f\u514d\u8def\u7531\u9000\u5316\u5e76\u4fdd\u7559\u5f3a\u65b9\u5411\u54cd\u5e94\u3002<\/li>\n<\/ol>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<p><strong>\u6a21\u5757 A\uff1aPatch-Ordered Scanning<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\uff1astage <code>l<\/code> \u7684 feature map <code>X_l \u2208 R^{C_l\u00d7H_l\u00d7W_l}<\/code>\u3002<\/li>\n<li>\u8f93\u51fa\uff1a\u4fdd\u6301 token \u6570\u4e0d\u53d8\u7684\u5e8f\u5217\u5316 feature\uff0c\u53ea\u6539\u53d8 token \u987a\u5e8f\u3002<\/li>\n<li>\u89e3\u51b3\u95ee\u9898\uff1a\u8ba9 patch \u5185\u7a7a\u95f4\u90bb\u8fd1\u50cf\u7d20\u5728 1D sequence \u4e2d\u8fde\u7eed\uff0c\u4ece\u800c\u51cf\u8f7b\u666e\u901a raster scan \u5bf9\u5c40\u90e8\u7ed3\u6784\u7684\u7834\u574f\u3002<\/li>\n<li>\u521b\u65b0\u6027\u5224\u65ad\uff1a\u601d\u60f3\u6734\u7d20\u4f46\u5408\u7406\uff0c\u50cf\u662f\u7ed9 Vision Mamba \u52a0\u4e0a\u66f4\u9002\u5408 dense prediction \u7684\u5c40\u90e8\u6027\u7ea6\u675f\uff1b\u6bd4\u201c\u6362\u4e00\u4e2a Mamba block \u540d\u5b57\u201d\u66f4\u6709\u673a\u5236\u610f\u4e49\u3002<\/li>\n<li>\u53ef\u8fc1\u79fb\u6027\uff1a\u9002\u5408\u8fc1\u79fb\u5230 VM-UNet\u3001VM-UNetV2\u3001DAMamba\u3001\u8f7b\u91cf polyp segmentation encoder\uff1b\u5b9e\u73b0\u4e0a\u53ea\u9700\u66ff\u6362 scan index\/permutation\uff0c\u4fb5\u5165\u6027\u8f83\u4f4e\u3002<\/li>\n<li>\u5bf9 3D segmentation\uff1a\u53ef\u6269\u5c55\u4e3a block-ordered \/ patch-cube-ordered scanning\uff0c\u4f46 3D \u4e2d scan path\u3001\u663e\u5b58\u548c anisotropic voxel spacing \u9700\u8981\u91cd\u65b0\u8bbe\u8ba1\u3002<\/li>\n<\/ul>\n<p><strong>\u6a21\u5757 B\uff1aHierarchical patch sizes \/ direction-specific scanning<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\uff1a\u56db\u4e2a\u65b9\u5411\u7684\u626b\u63cf\u8def\u5f84\u548c\u4e00\u7ec4 patch size\u3002<\/li>\n<li>\u8f93\u51fa\uff1a\u4e0d\u540c\u65b9\u5411\u3001\u4e0d\u540c\u7a7a\u95f4\u7c92\u5ea6\u7684 Mamba outputs\u3002<\/li>\n<li>\u89e3\u51b3\u95ee\u9898\uff1a\u5c0f patch \u504f\u5411\u7ec6\u8fb9\u754c\uff0c\u5927 patch \u504f\u5411\u8f83\u7c97\u7ed3\u6784\u548c\u66f4\u5927\u90bb\u57df\u3002<\/li>\n<li>\u8bc1\u636e\uff1aTable IV \u6bd4\u8f83\u4e86\u591a\u79cd patch size \u914d\u7f6e\uff0c<code>8844\/1111\/1111\/1111<\/code> \u5e73\u5747 Dice \u6700\u597d\uff0884.01\uff09\uff0c\u4f46\u8fd9\u91cc\u8868\u683c\u7b26\u53f7\u6bd4\u8f83\u7b80\u7565\uff0c\u8bba\u6587\u6ca1\u6709\u5145\u5206\u89e3\u91ca\u6bcf\u4e2a stage\/direction \u7684\u7cbe\u786e\u542b\u4e49\uff0c\u590d\u73b0\u65f6\u9700\u8981\u8bfb\u4ee3\u7801\u6216\u8054\u7cfb\u4f5c\u8005\u786e\u8ba4\u3002<\/li>\n<\/ul>\n<p><strong>\u6a21\u5757 C\uff1aMoE-Based Directional Fusion<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\uff1a\u56db\u4e2a\u65b9\u5411 feature maps <code>{Y_l^(1),...,Y_l^(4)}<\/code>\u3002<\/li>\n<li>\u8f93\u51fa\uff1a\u878d\u5408\u540e\u7684 <code>Z_l<\/code>\u3002<\/li>\n<li>\u673a\u5236\uff1a\u56db\u4e2a\u65b9\u5411 map \u7ecf GN \u4f5c\u4e3a\u56db\u4e2a expert\uff1bconcat \u540e 1\u00d71 conv \u5f62\u6210\u7b2c\u4e94 expert\uff1brouter \u751f\u6210\u7a7a\u95f4\u4f4d\u7f6e\u76f8\u5173\u6743\u91cd <code>w_l \u2208 R^{5\u00d7H_l\u00d7W_l}<\/code>\uff1b\u52a0\u6743\u878d\u5408\u540e\u518d residual add raw directional outputs\u3002<\/li>\n<li>\u89e3\u51b3\u95ee\u9898\uff1a\u4e0d\u540c\u533a\u57df\u53ef\u80fd\u9700\u8981\u4e0d\u540c\u65b9\u5411\/\u5c3a\u5ea6\u54cd\u5e94\uff0c\u56fa\u5b9a\u6c42\u548c\u8fc7\u4e8e\u7c97\u7cd9\u3002<\/li>\n<li>\u521b\u65b0\u6027\u5224\u65ad\uff1aMoE router \u5728 dense prediction \u4e2d\u5e38\u89c1\uff0c\u4f46\u7528\u4e8e\u66ff\u4ee3 Vision Mamba \u591a\u65b9\u5411\u626b\u63cf\u6c42\u548c\u662f\u6709\u9488\u5bf9\u6027\u7684\u6539\u52a8\uff1b\u771f\u6b63\u521b\u65b0\u5f3a\u5ea6\u4e2d\u7b49\uff0c\u5de5\u7a0b\u5b9e\u7528\u6027\u8f83\u5f3a\u3002<\/li>\n<li>\u6210\u672c\u95ee\u9898\uff1a\u8fd9\u662f\u6700\u5927\u5f31\u70b9\u3002Table V \u663e\u793a Patch-MoE Mamba \u8fbe\u5230 70.06M \u53c2\u6570\u300128.18G FLOPs\uff0c\u800c VM-UNetV2 \u662f 22.77M\u30015.31G FLOPs\u3002\u4e5f\u5c31\u662f\u8bf4\u51c6\u786e\u7387\u589e\u76ca\u4f34\u968f\u7ea6 3 \u500d\u53c2\u6570\u548c 5 \u500d FLOPs\uff0c\u4e0d\u9002\u5408\u4f5c\u4e3a\u201c\u8f7b\u91cf\u5316\u201d\u65b9\u6cd5\u3002<\/li>\n<\/ul>\n<p><strong>\u6a21\u5757 D\uff1aSDI module<\/strong><\/p>\n<ul>\n<li>\u6765\u6e90\uff1a\u6cbf\u7528 U-Net v2\uff0c\u4e0d\u662f\u672c\u6587\u6838\u5fc3\u539f\u521b\u3002<\/li>\n<li>\u4f5c\u7528\uff1a\u7528\u9ad8\u5c42\u8bed\u4e49\u548c\u4f4e\u5c42\u7ec6\u8282\u589e\u5f3a\u5404\u5c3a\u5ea6 feature\u3002<\/li>\n<li>\u5bf9\u7528\u6237\u4ef7\u503c\uff1a\u5982\u679c\u5df2\u6709 DAMamba\/VM-UNet \u7cfb\u5217\u6846\u67b6\uff0cSDI \u53ef\u4ee5\u4f5c\u4e3a skip refinement \u7ec4\u4ef6\u52a0\u5165\uff0c\u4f46\u9700\u8981\u5355\u72ec\u6d88\u878d\u786e\u8ba4\u5176\u8d21\u732e\u3002<\/li>\n<\/ul>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p>\u5b9e\u9a8c\u5305\u62ec 5 \u4e2a\u606f\u8089\u6570\u636e\u96c6\u548c 2 \u4e2a\u76ae\u80a4\u75c5\u7076\u6570\u636e\u96c6\uff1a<\/p>\n<ul>\n<li>Polyp\uff1aKvasir-SEG\u3001ClinicDB\u3001ColonDB\u3001ETIS\u3001CVC-300\u3002\u8bad\u7ec3\u534f\u8bae\u8ddf\u968f U-Net v2\uff1aKvasir-SEG 900 \u5f20 + ClinicDB 550 \u5f20\u8bad\u7ec3\uff1b\u6d4b\u8bd5\u5305\u62ec CVC-300 60\u3001ColonDB 380\u3001ETIS 196\u3001Kvasir-SEG 100\u3001ClinicDB 62\u3002<\/li>\n<li>Skin lesion\uff1aISIC 2017\uff082,150 images\uff09\u548c ISIC 2018\uff082,694 images\uff09\uff0c\u4f7f\u7528 U-Net v2 \u7684\u5212\u5206\u3002<\/li>\n<li>\u8f93\u5165\u5c3a\u5bf8\uff1a256\u00d7256\u3002<\/li>\n<li>\u8bad\u7ec3\uff1aPyTorch\uff0cNVIDIA Tesla A100 80GB\uff0cAdamW\uff0clr 1e-3\uff0cbatch size 80\uff0c300 epochs\uff0ccosine annealing\uff0cVMamba-S pretrained initialization\uff0c\u968f\u673a\u7ffb\u8f6c\/\u65cb\u8f6c\u589e\u5f3a\u3002<\/li>\n<li>\u6307\u6807\uff1aDice\/DSC\u3001IoU\u3001MAE\u3002<\/li>\n<li>\u91cd\u590d\uff1aTable I \u660e\u786e\u8bf4\u6bcf\u4e2a\u5b9e\u9a8c\u7528 5 \u4e2a\u968f\u673a\u79cd\u5b50\u8fd0\u884c\u3002<\/li>\n<\/ul>\n<p>\u4e3b\u8981\u7ed3\u679c\uff1a<\/p>\n<ul>\n<li>Polyp Table I\uff1aPatch-MoE Mamba \u5728\u4e94\u4e2a\u606f\u8089\u96c6\u4e0a\u7684 Dice \u5206\u522b\u4e3a Kvasir-SEG 90.90\u3001ClinicDB 91.32\u3001ColonDB 77.94\u3001ETIS 74.04\u3001CVC-300 87.31\uff0c\u6574\u4f53\u8d85\u8fc7 VM-UNetV2\u3002\u6700\u503c\u5f97\u6ce8\u610f\u7684\u662f ColonDB \u548c ETIS \u8fd9\u7c7b\u8de8\u6570\u636e\u96c6\u3001\u4f4e\u5bf9\u6bd4\/\u590d\u6742\u8fb9\u754c\u6d4b\u8bd5\u96c6\u4e0a\u6709\u63d0\u5347\u3002<\/li>\n<li>ISIC Table II\uff1aISIC 2017 Dice 90.85\u3001IoU 84.45\u3001MAE 0.0293\uff1bISIC 2018 Dice 89.34\u3001IoU 82.28\u3001MAE 0.0496\uff0c\u5747\u4f18\u4e8e\u5217\u51fa\u7684 U-Net\/U-Net v2\/VM-UNet\/VM-UNetV2\u3002<\/li>\n<li>Ablation Table III\uff1aVM-UNetV2 \u5e73\u5747 Dice 83.46\uff1b\u52a0 patch-ordered scanning \u5230 84.02\uff1b\u518d\u52a0 MoE fusion \u5230 84.30\u3002\u8bf4\u660e\u4e3b\u8981\u589e\u76ca\u6765\u81ea patch-ordered scanning\uff0cMoE \u8fdb\u4e00\u6b65\u5e26\u6765\u8f83\u5c0f\u589e\u76ca\u3002<\/li>\n<li>Complexity Table V\uff1a\u5b8c\u6574\u6a21\u578b 70.06M \u53c2\u6570\u300128.18G FLOPs\uff0c\u76f8\u6bd4 VM-UNetV2 \u7684 22.77M\/5.31G \u6210\u672c\u663e\u8457\u4e0a\u5347\u3002<\/li>\n<\/ul>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u53ef\u4fe1\u4e4b\u5904\uff1a<\/p>\n<ul>\n<li>\u8986\u76d6\u4e86\u606f\u8089\u548c\u76ae\u80a4\u75c5\u7076\u4e24\u7c7b\u6570\u636e\uff0c\u4e14\u606f\u8089\u4efb\u52a1\u5305\u542b\u591a\u4e2a\u5e38\u7528\u8de8\u6570\u636e\u96c6\u6d4b\u8bd5\u96c6\u3002<\/li>\n<li>\u4e0e U-Net\u3001U-Net v2\u3001VM-UNet\u3001VM-UNetV2 \u76f4\u63a5\u6bd4\u8f83\uff0cbaseline \u9009\u62e9\u4e0e\u8bba\u6587\u95ee\u9898\u76f8\u5173\u3002<\/li>\n<li>\u6709\u6a21\u5757\u7ea7\u6d88\u878d\u3001patch size \u6d88\u878d\u3001\u590d\u6742\u5ea6\u8868\uff1b\u5e76\u62a5\u544a\u591a\u968f\u673a\u79cd\u5b50\u5747\u503c\/\u65b9\u5dee\u3002<\/li>\n<\/ul>\n<p>\u9700\u8981\u8c28\u614e\u7684\u5730\u65b9\uff1a<\/p>\n<ul>\n<li>\u6bd4\u8f83\u5bf9\u8c61\u504f\u5c11\uff0c\u6ca1\u6709\u5305\u542b PraNet\u3001Polyp-PVT\u3001SANet\u3001TransFuse\u3001HarDNet-MSEG\u3001CaraNet \u7b49\u7ecf\u5178\/\u5f3a polyp segmentation baseline\uff1b\u56e0\u6b64\u201cpolyp SOTA\u201d\u4e0d\u80fd\u8f7b\u6613\u63a5\u53d7\u3002<\/li>\n<li>\u6210\u672c\u5f88\u9ad8\uff1a70M \u53c2\u6570\u548c 28G FLOPs \u5bf9 256\u00d7256 2D \u5206\u5272\u5e76\u4e0d\u8f7b\u3002\u82e5\u7528\u6237\u76ee\u6807\u662f\u8f7b\u91cf\u90e8\u7f72\uff0c\u8fd9\u7bc7\u4e0d\u662f\u76f4\u63a5\u7b54\u6848\u3002<\/li>\n<li>\u4ee3\u7801\u672a\u83b7\u53d6\uff0cpatch size \u914d\u7f6e\u8868\u8ff0\u4e0d\u591f\u6e05\u695a\uff0c\u590d\u73b0\u98ce\u9669\u4e2d\u7b49\u3002<\/li>\n<li>\u6ca1\u6709\u7edf\u8ba1\u663e\u8457\u6027\u68c0\u9a8c\uff0c\u867d\u7136\u7ed9\u4e86 5 seeds \u65b9\u5dee\uff0c\u4f46\u6ca1\u6709\u62a5\u544a p-value \u6216\u7f6e\u4fe1\u533a\u95f4\u68c0\u9a8c\u3002<\/li>\n<li>\u6ca1\u6709\u5916\u90e8\u4e34\u5e8a\u6570\u636e\u3001domain shift \u6216\u4e0d\u540c\u4e2d\u5fc3\u6570\u636e\u9a8c\u8bc1\uff1b\u6cdb\u5316\u7ed3\u8bba\u53ea\u80fd\u9650\u4e8e\u516c\u5f00 benchmark\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>\u4e0e U-Net\uff1a\u603b\u4f53\u4ecd\u662f encoder-decoder + skip \u7684 U-Net \u8303\u5f0f\uff0c\u6539\u52a8\u4e3b\u8981\u5728 encoder \u7684 VSS block \u548c\u7279\u5f81\u878d\u5408\u3002<\/li>\n<li>\u4e0e nnU-Net\uff1a\u4e0d\u662f\u81ea\u914d\u7f6e pipeline\uff0c\u4e5f\u6ca1\u6709\u8986\u76d6 nnU-Net \u7684\u6570\u636e\u9884\u5904\u7406\/\u8bad\u7ec3\u7b56\u7565\u4f18\u52bf\uff1b\u4e0d\u80fd\u66ff\u4ee3 nnU-Net\uff0c\u53ea\u80fd\u4f5c\u4e3a\u7f51\u7edc\u7ed3\u6784\u5019\u9009\u3002<\/li>\n<li>\u4e0e MedNeXt\uff1a\u6ca1\u6709\u76f4\u63a5\u6bd4\u8f83\uff1bMedNeXt \u662f ConvNeXt-style 3D\/2D medical segmentation \u5f3a baseline\uff0c\u8fd9\u7bc7\u66f4\u504f Mamba scan mechanism\u3002<\/li>\n<li>\u4e0e UNetR\/Swin-UNet\/TransUNet\/TransFuse\uff1a\u5b83\u7684\u5b9a\u4f4d\u662f\u7528\u7ebf\u6027\u590d\u6742\u5ea6 SSM \u66ff\u4ee3 Transformer attention \u7684\u5168\u5c40\u5efa\u6a21\uff0c\u4f46\u5b9e\u9a8c\u6ca1\u6709\u6b63\u9762\u5bf9\u8fd9\u4e9b Transformer \u533b\u5b66\u5206\u5272\u6a21\u578b\u505a\u5145\u5206\u6bd4\u8f83\u3002<\/li>\n<li>\u4e0e VMamba\/VM-UNet\/VM-UNetV2\uff1a\u8fd9\u662f\u6700\u76f4\u63a5\u7684\u7ee7\u627f\u5173\u7cfb\u3002\u8bba\u6587\u660e\u786e\u66ff\u6362 VM-UNetV2 \u7684 VSS block\uff0c\u5e76\u4fdd\u7559 VM-UNetV2 decoder\u3002<\/li>\n<li>\u4e0e DAMamba\uff1a\u5982\u679c DAMamba \u4e2d\u4e5f\u6709\u591a\u65b9\u5411 scanning \u6216 direction fusion\uff0c\u8fd9\u7bc7\u7684 patch-ordered index \u4e0e spatial-aware MoE router \u90fd\u662f\u53ef\u501f\u9274\u6539\u9020\u70b9\u3002<\/li>\n<li>\u4e0e foundation model\uff1a\u6ca1\u6709\u4f7f\u7528 SAM\/MedSAM \u6216\u901a\u7528\u533b\u5b66\u5206\u5272 foundation model\uff0c\u4e0d\u5c5e\u4e8e foundation model \u8def\u7ebf\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\u5b83\u76f4\u63a5\u5728\u4e94\u4e2a\u606f\u8089\u6570\u636e\u96c6\u9a8c\u8bc1\uff0c\u4e14\u63d0\u51fa\u7684\u95ee\u9898\uff08\u8fb9\u754c\u3001\u5c0f\u76ee\u6807\u3001\u4f4e\u5bf9\u6bd4\u3001\u65b9\u5411\u878d\u5408\uff09\u4e0e\u606f\u8089\u5206\u5272\u9ad8\u5ea6\u76f8\u5173\u3002\u5efa\u8bae\u4f18\u5148\u590d\u73b0 patch-ordered scanning\uff0c\u800c\u4e0d\u662f\u4e00\u5f00\u59cb\u5c31\u590d\u73b0\u5b8c\u6574 MoE\uff0c\u56e0\u4e3a ablation \u663e\u793a\u524d\u8005\u8d21\u732e\u66f4\u5927\u4e14\u6210\u672c\u66f4\u4f4e\u3002<\/p>\n<p>\u5bf9 DAMamba \u6539\u9020\uff1a\u4ef7\u503c\u9ad8\u3002\u53ef\u4ee5\u5c1d\u8bd5\u4e09\u79cd\u4f4e\u98ce\u9669\u6539\u6cd5\uff1a<\/p>\n<ol>\n<li>\u628a DAMamba\/VMamba \u7684 raster scan \u66ff\u6362\u4e3a patch-ordered scan\uff1b<\/li>\n<li>\u4fdd\u7559\u539f\u65b9\u5411\u6c42\u548c\uff0c\u4f46\u52a0\u5165\u8f7b\u91cf gating\uff0c\u4f8b\u5982\u53ea\u7528 1\u00d71 conv \u751f\u6210\u56db\u65b9\u5411\u6743\u91cd\uff0c\u4e0d\u52a0 concat expert\uff1b<\/li>\n<li>\u5728 decoder skip \u5904\u7ed3\u5408 SDI \u6216\u8f7b\u91cf\u8bed\u4e49-\u7ec6\u8282\u4ea4\u4e92\u6a21\u5757\u3002<\/li>\n<\/ol>\n<p>\u5bf9 3D medical segmentation\uff1a\u6982\u5ff5\u53ef\u8fc1\u79fb\uff0c\u4f46\u4e0d\u80fd\u76f4\u63a5\u7167\u642c\u30023D \u4e2d\u9700\u8981\u8bbe\u8ba1 cube-ordered scanning\u3001\u8f74\u5411\u65b9\u5411\u7ec4\u5408\u3001anisotropic patch size\uff0c\u5e76\u91cd\u65b0\u8bc4\u4f30\u663e\u5b58\u3002<\/p>\n<p>\u5bf9 related work\uff1a\u503c\u5f97\u52a0\u5165 Mamba-based medical segmentation \/ VM-UNet \u6539\u8fdb\u90e8\u5206\uff0c\u5c24\u5176\u4f5c\u4e3a\u201cscan order matters for dense prediction\u201d\u7684\u4f8b\u5b50\u3002<\/p>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p>\u5efa\u8bae\u7cbe\u8bfb\u3002\u7406\u7531\u662f\u8bba\u6587\u673a\u5236\u6e05\u695a\uff0c\u548c\u7528\u6237\u5173\u5fc3\u7684 polyp segmentation\u3001Mamba-based segmentation\u3001DAMamba \u6539\u9020\u9ad8\u5ea6\u76f8\u5173\uff1b\u4f46\u9605\u8bfb\u65f6\u8981\u91cd\u70b9\u76ef\u4f4f Table III \u548c Table V\uff1a\u5b83\u7684\u589e\u76ca\u5e76\u4e0d\u5b8c\u5168\u514d\u8d39\uff0c\u5b8c\u6574 MoE \u6a21\u578b\u8ba1\u7b97\u6210\u672c\u660e\u663e\u5347\u9ad8\u3002\u6700\u63a8\u8350\u5148\u8bfb Methods II-B\/II-C \u548c Tables I\/III\/V\u3002<\/p>\n<hr \/>\n<h2>\u8bba\u6587 2\uff1aMHMamba: Multi-Head Mamba for 3D Brain Tumor Segmentation<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aMHMamba: Multi-Head Mamba for 3D Brain Tumor Segmentation<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aHanjun Tao, Hua Wang, Fan Zhang \/ \u7b2c\u4e00\u4f5c\u8005 Hanjun Tao<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-15\uff08arXiv v1\uff09<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.16464<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.16464<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1a\u672a\u83b7\u53d6 \/ arXiv \u9875\u9762\u4e0e\u8bba\u6587\u6b63\u6587\u672a\u7ed9\u51fa\u660e\u786e\u4ee3\u7801\u4ed3\u5e93<\/li>\n<li>\u4efb\u52a1\uff1a3D brain tumor segmentation<\/li>\n<li>\u6570\u636e\u96c6\uff1aBraTS2021\u3001BraTS2023<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1a3D U-shaped Mamba segmentation network\uff1bmulti-head SSM + channel-spatial calibration + adaptive gated skip 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\u7684\u4ef7\u503c\u5728\u4e8e\u628a Mamba \u7528\u201c\u591a\u5934\u901a\u9053\u62c6\u5206 + \u6821\u51c6 + \u95e8\u63a7 skip fusion\u201d\u7684\u65b9\u5f0f\u7cfb\u7edf\u5316\u5230 3D BraTS \u5206\u5272\uff0c\u4f46\u5176\u5b9e\u9a8c\u548c\u5199\u4f5c\u5b58\u5728\u82e5\u5e72\u9700\u8981\u6838\u67e5\u7684\u5730\u65b9\uff0c\u66f4\u9002\u5408\u4f5c\u4e3a 3D Mamba \u7ed3\u6784\u8bbe\u8ba1\u53c2\u8003\uff0c\u800c\u4e0d\u662f\u76f4\u63a5\u5f53\u4f5c\u5f3a SOTA \u4f9d\u636e\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u8111\u80bf\u7624 MRI \u5206\u5272\u9700\u8981\u540c\u65f6\u8bc6\u522b WT\u3001TC\u3001ET \u7b49\u4e0d\u540c\u80bf\u7624\u5b50\u533a\u57df\u3002ET \u7b49\u5c0f\u4f53\u79ef\u589e\u5f3a\u533a\u57df\u5f62\u6001\u53d8\u5316\u5927\u3001\u8fb9\u754c\u590d\u6742\uff0c\u5bf9\u5168\u5c40\u4e0a\u4e0b\u6587\u548c\u5c40\u90e8\u8fb9\u754c\u90fd\u654f\u611f\u30023D U-Net\/CNN \u64c5\u957f\u5c40\u90e8\u7eb9\u7406\uff0c\u4f46\u957f\u7a0b\u4f9d\u8d56\u5efa\u6a21\u5f31\uff1bTransformer \u80fd\u5efa\u7acb\u5168\u5c40\u5173\u7cfb\uff0c\u4f46 3D MRI token \u6570\u5de8\u5927\uff0c\u6807\u51c6 attention \u7684\u4e8c\u6b21\u590d\u6742\u5ea6\u4f1a\u5e26\u6765\u663e\u5b58\u548c\u8ba1\u7b97\u538b\u529b\uff1b\u7a97\u53e3 attention \u53c8\u53ef\u80fd\u5bfc\u81f4\u8de8\u7a97\u53e3\u4e0a\u4e0b\u6587\u4e0d\u8fde\u8d2f\u548c sliding-window inference \u7684\u8fb9\u754c\u4e0d\u5e73\u6ed1\u3002<\/p>\n<p>\u4f5c\u8005\u7684\u95ee\u9898\u8bbe\u5b9a\u662f\uff1a\u5982\u4f55\u5728 3D multimodal MRI \u4e2d\u4fdd\u6301 Mamba \u7684\u7ebf\u6027\u590d\u6742\u5ea6\u4f18\u52bf\uff0c\u540c\u65f6\u63d0\u9ad8\u591a\u6a21\u6001\u3001\u591a\u5c3a\u5ea6\u3001\u590d\u6742\u80bf\u7624\u533a\u57df\u7684\u5168\u5c40-\u5c40\u90e8\u8868\u8fbe\u7a33\u5b9a\u6027\u3002<\/p>\n<p>\u5185\u90e8 paper map\uff1a\u8bba\u6587\u7814\u7a76 3D brain tumor segmentation \u4e2d\u9ad8\u6548\u957f\u7a0b\u5efa\u6a21\u95ee\u9898\uff0c\u8bbe\u7f6e\u4e3a BraTS2021\/2023 \u591a\u6a21\u6001 MRI\uff1b\u4e3b\u62db\u662f U-shaped architecture \u4e2d\u5f15\u5165 Multi-Head Mamba\u3001CSCA \u6821\u51c6\u548c AGF skip fusion\uff1b\u58f0\u79f0\u5728 Dice \u548c HD95 \u4e0a\u4f18\u4e8e nnFormer\u3001VcaNet\u3001nnU-Net\u3001LightUNet\u3001SegMamba\u3001SegMamba-V2\uff1b\u5173\u952e\u5bf9\u8c61\u662f GLA\u3001MHM\u3001CSCA\u3001AGF\uff1b\u771f\u6b63\u8d1f\u91cd\u5728\u7edf\u4e00\u8bad\u7ec3\u534f\u8bae\u4e0b\u7684 BraTS \u5bf9\u6bd4\u548c\u6a21\u5757\u6d88\u878d\uff1b\u4e3b\u8981\u5931\u8d25\u98ce\u9669\u662f arXiv preprint \u7684\u5b9e\u73b0\/\u590d\u73b0\u4fe1\u606f\u4e0d\u8db3\u3001\u4ee3\u7801\u7f3a\u5931\u3001\u90e8\u5206\u5f15\u7528\u548c\u8bba\u8ff0\u53ef\u7591\uff0c\u4e14\u6ca1\u6709\u5b98\u65b9 BraTS test server \u7ed3\u679c\u3002<\/p>\n<p>\u8def\u7531\u8bb0\u5f55\uff1aPrimary adapter = method-algorithm\uff1bSecondary adapter = benchmark-evaluation\uff1bEvidence packs = general\u3001experimental-eval\u3001ablation-and-mechanism-isolation\u3001reproducibility-and-compute\uff1bRoute confidence = \u4e2d\u9ad8\u3002\u539f\u56e0\uff1a\u8bba\u6587\u8d21\u732e\u662f\u7b97\u6cd5\u6a21\u5757\uff0c\u8bc1\u636e\u6765\u81ea BraTS benchmark \u4e0e\u6d88\u878d\uff0c\u4f46\u590d\u73b0\u548c\u516c\u5e73\u6027\u4ecd\u9700\u8c28\u614e\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u5f52\u7eb3\u4e86\u4e09\u7c7b\u4e0d\u8db3\uff1a<\/p>\n<ol>\n<li>CNN\/3D U-Net\uff1a\u6709\u6548\u611f\u53d7\u91ce\u53d7\u5c40\u90e8\u5377\u79ef\u9650\u5236\uff0c\u8fde\u7eed\u4e0b\u91c7\u6837\u4f1a\u524a\u5f31\u5c0f\u75c5\u7076\u548c\u8fb9\u754c\u7ec6\u8282\u3002<\/li>\n<li>Transformer\/TransBTS\/nnFormer\/Swin \u7c7b\u65b9\u6cd5\uff1a3D token \u6570\u5bfc\u81f4\u8ba1\u7b97\u548c\u663e\u5b58\u6210\u672c\u9ad8\uff1bwindow\/block attention \u964d\u4f4e\u6210\u672c\u4f46\u8de8\u7a97\u53e3\u5173\u7cfb\u5f31\uff0csliding-window \u63a8\u7406\u53ef\u80fd\u5f15\u5165\u6982\u7387\u573a\u4e0d\u8fde\u7eed\u3002<\/li>\n<li>\u73b0\u6709 Mamba\/SSM \u533b\u5b66\u5206\u5272\uff1a\u7b80\u5355\u987a\u5e8f SSM \u5728 3D MRI \u4e2d\u53ef\u80fd\u8bad\u7ec3\u4e0d\u7a33\u5b9a\uff0c\u5168\u5c40-\u5c40\u90e8\u8868\u8fbe\u4e0d\u8db3\uff0cskip connection \u4fe1\u606f\u878d\u5408\u4e0d\u591f\uff0c\u5c24\u5176\u5f71\u54cd ring enhancement\u3001necrosis\/living tissue \u8fb9\u754c\u548c\u5c0f\u4f53\u79ef ET \u533a\u57df\u3002<\/li>\n<\/ol>\n<p>\u8fd9\u4e9b\u95ee\u9898\u57fa\u672c\u7b26\u5408 3D BraTS \u7684\u4efb\u52a1\u75db\u70b9\uff1b\u4e0d\u8fc7\u4f5c\u8005\u5bf9 Transformer \u201cinter-block contextual incoherence\u201d\u7684\u63cf\u8ff0\u8f83\u5f3a\uff0c\u8bba\u6587\u6ca1\u6709\u5355\u72ec\u5b9e\u9a8c\u8bc1\u660e\u8fd9\u4e00\u70b9\uff0c\u53ea\u662f\u4f5c\u4e3a\u65b9\u6cd5\u52a8\u673a\u3002<\/p>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>MHMamba \u662f\u4e00\u4e2a U-shaped 3D encoder-decoder\uff1a<\/p>\n<ol>\n<li>\u8f93\u5165\u4e3a\u591a\u6a21\u6001 MRI volume\uff0c\u56db\u4e2a\u6a21\u6001 T1\u3001T1ce\u3001T2\u3001FLAIR \u88ab concat \u4e3a 4-channel input\u3002<\/li>\n<li>stem \u4f7f\u7528 7\u00d77\u00d77 depthwise separable convolution\uff0cstride=2\uff0c\u628a\u8f93\u5165\u6295\u5f71\u5230 48 channels\uff0c\u7a7a\u95f4\u5c3a\u5bf8\u51cf\u534a\u3002<\/li>\n<li>encoder \u6709 4 \u4e2a stage\uff0c\u6bcf\u4e2a stage \u5806\u53e0 MHMamba block\uff0c\u5e76\u5728 stage \u672b\u5c3e\u7528 3\u00d73\u00d73 stride=2 convolution \u4e0b\u91c7\u6837\uff0c\u9010\u6b65\u6269\u5927\u901a\u9053\u6570\u3001\u7f29\u5c0f\u7a7a\u95f4\u5c3a\u5bf8\u3002<\/li>\n<li>MHMamba block \u5185\u5148\u7528 Gated Local Aggregation\uff08GLA\uff09\u589e\u5f3a\u8fb9\u754c\/\u7ec6\u8282\uff0c\u518d\u7528 Multi-Head Mamba\uff08MHM\uff09\u505a\u5168\u5c40\u5e8f\u5217\u5efa\u6a21\uff0c\u518d\u7528 Channel-Spatial Calibration Attention\uff08CSCA\uff09\u8fdb\u884c\u901a\u9053\u4e0e\u7a7a\u95f4\u6821\u51c6\uff0c\u6700\u540e\u901a\u8fc7 LayerNorm\/MLP \u548c residual \u5b8c\u6210\u8f93\u51fa\u3002<\/li>\n<li>decoder \u4f7f\u7528 3D convolution + upsampling \u6062\u590d\u5206\u8fa8\u7387\u3002<\/li>\n<li>skip connection \u4e0d\u505a\u7b80\u5355 concat\/add\uff0c\u800c\u4f7f\u7528 Adaptive Gated Fusion\uff08AGF\uff09\u5728\u901a\u9053\u5b50\u7ec4\u4e0a\u5b66\u4e60 encoder feature \u4e0e decoder feature \u7684\u878d\u5408\u6743\u91cd\u3002<\/li>\n<\/ol>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<p><strong>\u6a21\u5757 A\uff1aGated Local Aggregation\uff08GLA\uff09<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\uff1aencoder feature <code>F<\/code>\u3002<\/li>\n<li>\u8f93\u51fa\uff1a\u5c40\u90e8\u589e\u5f3a feature <code>F_GLA<\/code>\u3002<\/li>\n<li>\u516c\u5f0f\uff1a<code>F_edge = Sobel3D(F)<\/code>\uff1b<code>F_detail = Conv(ReLU(IN(F)))<\/code>\uff1b<code>F_GLA = \u03b1\u00b7F_edge + \u03b2\u00b7F_detail<\/code>\u3002<\/li>\n<li>\u89e3\u51b3\u95ee\u9898\uff1a\u5728\u8fdb\u5165\u5168\u5c40 Mamba \u524d\u5148\u7a81\u51fa 3D \u80bf\u7624\u8fb9\u754c\u4e0e\u5c40\u90e8\u51e0\u4f55\u7ec6\u8282\u3002<\/li>\n<li>\u521b\u65b0\u6027\u5224\u65ad\uff1aSobel\/\u8fb9\u7f18\u5206\u652f + \u5377\u79ef\u5206\u652f\u662f\u5e38\u89c1\u8fb9\u754c\u589e\u5f3a\u601d\u8def\uff0c\u521b\u65b0\u6027\u4e00\u822c\uff0c\u4f46\u5bf9 BraTS \u7684\u8fb9\u754c HD95 \u76ee\u6807\u5408\u7406\u3002<\/li>\n<li>\u53ef\u8fc1\u79fb\u6027\uff1a\u9002\u5408\u8fc1\u79fb\u5230 3D U-Net\/SegMamba\/DAMamba \u7684\u6d45\u5c42\u6216\u6bcf\u4e2a stage \u524d\uff1b\u5bf9 2D polyp segmentation \u4e5f\u53ef\u6539\u6210 Sobel2D + detail branch\uff0c\u4f46\u8981\u6ce8\u610f\u566a\u58f0\u8fb9\u7f18\u3002<\/li>\n<\/ul>\n<p><strong>\u6a21\u5757 B\uff1aMulti-Head Mamba\uff08MHM\uff09<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\uff1a<code>F_GLA<\/code> \u7ecf LayerNorm \u540e\u7684 feature\u3002<\/li>\n<li>\u5904\u7406\uff1a\u6cbf channel \u7ef4\u62c6\u6210 <code>N_h<\/code> \u4e2a head\uff1b\u6bcf\u4e2a head flatten \u4e3a <code>B\u00d7N\u00d7(C\/N_h)<\/code>\uff0c\u5176\u4e2d <code>N=D\u00d7H\u00d7W<\/code>\uff1b\u6bcf\u4e2a head \u901a\u8fc7 selective state space recurrence\uff1b\u6700\u540e concat \u5404 head \u8f93\u51fa\uff0c\u7528 1\u00d71\u00d71 projection \u878d\u5408\uff0c\u5e76\u52a0\u4e0a learnable residual <code>\u03b4\u00b7F_GLA<\/code>\u3002<\/li>\n<li>\u89e3\u51b3\u95ee\u9898\uff1a\u5355\u4e00\u8def\u5f84 SSM \u53ef\u80fd\u628a\u4e0d\u540c\u6a21\u6001\/\u4e0d\u540c\u65b9\u5411\u4fe1\u606f\u5e73\u5747\u5316\uff1b\u591a\u5934\u62c6\u5206\u8ba9\u4e0d\u540c head \u5b66\u5230\u4e92\u8865\u957f\u7a0b\u6a21\u5f0f\u3002<\/li>\n<li>\u521b\u65b0\u6027\u5224\u65ad\uff1a\u8fd9\u662f\u628a Transformer multi-head \u7684\u601d\u60f3\u79fb\u690d\u5230 Mamba channel split \u4e0a\uff0c\u5408\u7406\u4f46\u4e0d\u7b97\u975e\u5e38\u65b0\uff1b\u5173\u952e\u770b head \u6570\u6d88\u878d\u662f\u5426\u652f\u6301\u3002<\/li>\n<li>\u8bc1\u636e\uff1aTable 4 \u663e\u793a BraTS2021 \u4e0a <code>N=4<\/code> \u6700\u597d\uff0c\u5e73\u5747 Dice\/HD \u4e3a 91.02\/3.38\uff1b<code>N=2<\/code> \u4e3a 90.44\/4.04\uff0c<code>N=8<\/code> \u4e3a 90.48\/4.05\uff0c\u8bf4\u660e\u8fc7\u7ec6\u62c6\u5206\u4f1a\u635f\u5bb3\u6bcf\u4e2a head \u5bb9\u91cf\u5e76\u4f7f batch size 1 \u7684 3D \u8bad\u7ec3\u66f4\u4e0d\u7a33\u5b9a\u3002<\/li>\n<\/ul>\n<p><strong>\u6a21\u5757 C\uff1aChannel-Spatial Calibration Attention\uff08CSCA\uff09<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\uff1aMHM \u8f93\u51fa <code>F_MHM<\/code>\u3002<\/li>\n<li>\u901a\u9053\u8def\u5f84\uff1aGAP \u4e0e GMP \u7ecf MLP \u4ea7\u751f channel weights\u3002<\/li>\n<li>\u7a7a\u95f4\u8def\u5f84\uff1a\u4f7f\u7528 Mean\u3001Std\u3001Max\u3001Min \u56db\u7c7b\u7edf\u8ba1\u62fc\u63a5\u540e\u751f\u6210 spatial weight map\u3002<\/li>\n<li>\u878d\u5408\uff1a\u7528 gate <code>\u03bb<\/code> \u81ea\u9002\u5e94\u878d\u5408 channel-calibrated \u548c spatial-calibrated feature\uff0c\u518d\u52a0 residual\u3002<\/li>\n<li>\u89e3\u51b3\u95ee\u9898\uff1a\u591a\u5934 Mamba \u8f93\u51fa\u53ef\u80fd\u5c3a\u5ea6\/\u7edf\u8ba1\u4e0d\u4e00\u81f4\uff0c\u9700\u8981\u6821\u51c6\u5e76\u589e\u5f3a\u80bf\u7624\u76f8\u5173\u54cd\u5e94\u3002<\/li>\n<li>\u521b\u65b0\u6027\u5224\u65ad\uff1a\u7c7b\u4f3c CBAM\/SE \u7684\u6269\u5c55\u7248\uff0c\u52a0\u4e0a Std\/Min \u7edf\u8ba1\u5bf9\u5f02\u8d28\u80bf\u7624\u6709\u4e00\u5b9a\u76f4\u89c9\uff0c\u4f46\u4e0d\u662f\u6839\u672c\u6027\u65b0\u7ed3\u6784\u3002<\/li>\n<li>\u53ef\u8fc1\u79fb\u6027\uff1a\u53ef\u7528\u4e8e 3D segmentation encoder block \u540e\uff1b\u5bf9 polyp segmentation \u4e5f\u53ef\u4f5c\u4e3a\u8f7b\u91cf attention\uff0c\u4f46\u8981\u63a7\u5236\u53c2\u6570\u3002<\/li>\n<\/ul>\n<p><strong>\u6a21\u5757 D\uff1aAdaptive Gated Fusion\uff08AGF\uff09<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\uff1aencoder skip feature <code>F_enc<\/code> \u548c decoder feature <code>F_dec<\/code>\u3002<\/li>\n<li>\u5904\u7406\uff1a\u4e8c\u8005\u5404\u81ea\u6cbf channel \u5206\u6210 4 \u4e2a subgroup\uff1b\u6bcf\u4e2a subgroup \u5b66\u4e60\u4e00\u4e2a gate <code>\u03b4_k = \u03c3(w_k \u00b7 [F_enc^k, F_dec^k])<\/code>\uff1b\u878d\u5408\u4e3a <code>F_fused^k = \u03b4_k\u00b7F_enc^k + (1-\u03b4_k)\u00b7F_dec^k<\/code>\uff1b\u6700\u540e concat \u540e\u5377\u79ef\u8f93\u51fa\u3002<\/li>\n<li>\u89e3\u51b3\u95ee\u9898\uff1a\u4f20\u7edf skip concat\/add \u5bf9\u8bed\u4e49\u4e0e\u7ec6\u8282\u51b2\u7a81\u5904\u7406\u7c97\u7cd9\uff0c\u5c24\u5176\u80bf\u7624\u8fb9\u754c\u5904\u53ef\u80fd\u5f15\u5165\u4e0d\u4e00\u81f4\u3002<\/li>\n<li>\u521b\u65b0\u6027\u5224\u65ad\uff1a\u95e8\u63a7 skip fusion \u662f\u6210\u719f\u601d\u8def\uff0c\u4f46\u4ee5 subgroup gating \u7528\u4e8e 3D Mamba U-Net \u662f\u5408\u7406\u5de5\u7a0b\u8bbe\u8ba1\u3002<\/li>\n<li>\u5bf9 DAMamba\/3D segmentation\uff1a\u53ef\u4f5c\u4e3a skip fusion \u6a21\u5757\u66ff\u6362 naive concat\uff0c\u4f18\u5148\u7ea7\u9ad8\u4e8e\u590d\u6742 attention\uff0c\u56e0\u4e3a\u5b9e\u73b0\u7b80\u5355\u3001\u89e3\u91ca\u660e\u786e\u3002<\/li>\n<\/ul>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p>\u6570\u636e\u96c6\u4e0e\u9884\u5904\u7406\uff1a<\/p>\n<ul>\n<li>BraTS2021\uff1a1,251 training samples \u548c 219 validation cases\u3002<\/li>\n<li>BraTS2023\uff1a1,534 training samples\u3002<\/li>\n<li>\u6bcf\u4e2a case \u6709 T1\u3001T1ce\u3001T2\u3001FLAIR \u56db\u6a21\u6001\uff0cconcat \u4e3a 4-channel input\u3002<\/li>\n<li>\u7edf\u4e00\u9884\u5904\u7406\u5305\u62ec\u591a\u6a21\u6001\u914d\u51c6\u3001skull stripping\u3001isotropic resampling\uff0c\u7edf\u4e00\u5c3a\u5bf8 240\u00d7240\u00d7155\u3002<\/li>\n<li>\u6807\u6ce8\uff1aET\u3001TC\u3001WT\u3002<\/li>\n<li>\u4f5c\u8005\u6ca1\u6709\u7528\u5b98\u65b9\u6d4b\u8bd5\u670d\u52a1\u5668\uff0c\u800c\u662f\u5728 official training set \u4e0a\u505a\u5185\u90e8 70%\/10%\/20% train\/val\/test split\uff0c\u56fa\u5b9a\u968f\u673a\u79cd\u5b50\u3002<\/li>\n<\/ul>\n<p>\u8bad\u7ec3\u8bbe\u7f6e\uff1a<\/p>\n<ul>\n<li>PyTorch 2.1.2\uff0c\u5355\u5f20 NVIDIA RTX 3090\u3002<\/li>\n<li>\u4ece\u968f\u673a\u521d\u59cb\u5316\u8bad\u7ec3 300 epochs\uff0cbatch size 1\uff0c\u521d\u59cb lr=0.001\uff0cweight decay=1e-5\uff0cpoly LR decay\u3002<\/li>\n<li>patch crop\uff1a128\u00d7128\u00d7128\u3002<\/li>\n<li>augmentation\uff1abrightness\u3001gamma\u3001rotation\u3001scaling\u3001mirror flipping\u3001elastic deformation\u3002<\/li>\n<li>loss\uff1aDice loss + cross entropy loss \u7b49\u6743\u91cd\u3002<\/li>\n<li>inference\uff1a\u7edf\u4e00 sliding-window\uff0c\u65e0 TTA\uff0c\u65e0 post-processing\u3002<\/li>\n<li>\u6307\u6807\uff1aDice \u548c HD95\u3002<\/li>\n<\/ul>\n<p>\u4e3b\u7ed3\u679c Table 1\uff1a<\/p>\n<ul>\n<li>BraTS2021\uff1aMHMamba WT Dice\/HD 93.54\/3.74\uff0cTC 92.23\/2.89\uff0cET 87.28\/3.50\uff0c\u5e73\u5747 Dice\/HD 91.02\/3.38\u3002\u76f8\u5bf9 SegMamba 90.32\/4.18 \u548c SegMamba-V2 89.14\/3.80 \u6709\u63d0\u5347\uff1b\u4f46 ET Dice \u4f4e\u4e8e VcaNet \u7684 87.55\u3002<\/li>\n<li>BraTS2023\uff1aMHMamba WT 93.87\/3.30\uff0cTC 91.10\/3.49\uff0cET 85.72\/4.23\uff0c\u5e73\u5747 90.23\/3.67\u3002\u5e73\u5747 Dice \u9ad8\u4e8e SegMamba 89.38 \u548c SegMamba-V2 89.98\uff0c\u4f46 ET Dice \u4f4e\u4e8e VcaNet 86.57 \u548c SegMamba-V2 86.64\u3002<\/li>\n<\/ul>\n<p>\u6d88\u878d\uff1a<\/p>\n<ul>\n<li>BraTS2021 Table 2\uff1abase\uff08SegMamba\uff09\u5e73\u5747 Dice\/HD 90.32\/4.18\uff1b+CSCA \u4e3a 90.83\/3.47\uff1b+AGF \u4e3a 90.66\/3.67\uff1b+MHM \u4e3a 90.62\/3.94\uff1b\u5b8c\u6574 ours \u4e3a 91.02\/3.38\u3002<\/li>\n<li>BraTS2023 Table 3\uff1abase 89.38\/4.25\uff1b+CSCA 89.83\/4.14\uff1b+AGF 89.72\/4.13\uff1b+MHM 89.73\/4.05\uff1b\u5b8c\u6574 90.23\/3.67\u3002<\/li>\n<li>Head number Table 4\uff1a4 heads \u6700\u597d\uff0c2 heads \u5bb9\u91cf\u4e0d\u8db3\uff0c8 heads \u8fc7\u5ea6\u62c6\u5206\u3002<\/li>\n<\/ul>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u53ef\u4fe1\u4e4b\u5904\uff1a<\/p>\n<ul>\n<li>\u4f7f\u7528 BraTS2021 \u548c BraTS2023 \u4e24\u4e2a\u4e3b\u6d41 3D \u8111\u80bf\u7624\u5206\u5272\u6570\u636e\u96c6\u3002<\/li>\n<li>\u6307\u6807\u5305\u542b Dice \u548c HD95\uff0c\u80fd\u540c\u65f6\u53cd\u6620\u533a\u57df\u91cd\u53e0\u548c\u8fb9\u754c\u8bef\u5dee\u3002<\/li>\n<li>baseline \u5305\u542b nnU-Net\u3001nnFormer\u3001SegMamba\u3001SegMamba-V2\uff0c\u548c\u8bba\u6587\u5b9a\u4f4d\u76f8\u5173\u3002<\/li>\n<li>\u6d88\u878d\u8986\u76d6 MHM\u3001CSCA\u3001AGF \u548c head number\uff0c\u80fd\u521d\u6b65\u652f\u6301\u6a21\u5757\u4f5c\u7528\u3002<\/li>\n<li>\u4f5c\u8005\u5f3a\u8c03\u7edf\u4e00\u9884\u5904\u7406\u3001\u8bad\u7ec3\u3001loss\u3001sliding-window inference\uff0c\u5bf9\u516c\u5e73\u6bd4\u8f83\u662f\u52a0\u5206\u9879\u3002<\/li>\n<\/ul>\n<p>\u9700\u8981\u8c28\u614e\u7684\u5730\u65b9\uff1a<\/p>\n<ul>\n<li>\u6ca1\u6709\u5b98\u65b9 BraTS test server \u7ed3\u679c\uff0c\u53ea\u662f\u5185\u90e8 split\uff1b\u4e0e\u6b63\u5f0f challenge leaderboard \u4e0d\u53ef\u76f4\u63a5\u6bd4\u8f83\u3002<\/li>\n<li>\u6587\u4e2d\u79f0 Table 1 \u7ed3\u679c averaged over multiple runs\uff0c\u4f46\u6ca1\u6709\u6e05\u695a\u62a5\u544a seed \u6570\u3001\u5747\u503c\u00b1\u65b9\u5dee\uff0c\u7edf\u8ba1\u7a33\u5b9a\u6027\u4e0d\u8db3\u3002<\/li>\n<li>\u4ee3\u7801\u672a\u83b7\u53d6\uff0c\u590d\u73b0\u96be\u5ea6\u548c\u5b9e\u73b0\u7ec6\u8282\u4e0d\u900f\u660e\u3002<\/li>\n<li>\u53c2\u6570\u91cf\u3001FLOPs\u3001\u663e\u5b58\u3001\u63a8\u7406\u901f\u5ea6\u6ca1\u6709\u7ed9\u51fa\uff1b\u867d\u7136\u5f3a\u8c03 Mamba \u7ebf\u6027\u590d\u6742\u5ea6\uff0c\u4f46\u6ca1\u6709\u5b9e\u6d4b\u6548\u7387\u5bf9\u6bd4\uff0c\u8fd9\u662f reproducibility\/compute \u65b9\u9762\u7684\u660e\u663e\u7f3a\u53e3\u3002<\/li>\n<li>\u90e8\u5206\u53c2\u8003\u6587\u732e\u548c\u76f8\u5173\u5de5\u4f5c\u5f15\u7528\u663e\u5f97\u6df7\u6742\uff0c\u4f8b\u5982\u628a\u56fe\u6587\u68c0\u7d22\/\u89c6\u9891\u68c0\u7d22\u8bba\u6587\u7528\u4e8e\u89e3\u91ca CSCA\/MHM \u7684\u5408\u7406\u6027\uff0c\u964d\u4f4e\u4e86\u8bba\u8bc1\u53ef\u4fe1\u5ea6\u3002<\/li>\n<li>ET \u6307\u6807\u4e0d\u662f\u6240\u6709\u60c5\u51b5\u4e0b\u6700\u5f3a\uff1b\u82e5\u7814\u7a76\u91cd\u70b9\u662f enhancing tumor\uff0c\u5c0f\u75c5\u7076\u654f\u611f\u6027\u7ed3\u8bba\u9700\u8981\u5f31\u5316\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>\u4e0e U-Net\/3D U-Net\uff1a\u6574\u4f53\u4ecd\u662f U-shaped encoder-decoder\uff1b\u521b\u65b0\u4e3b\u8981\u5728 encoder block \u548c skip fusion\u3002<\/li>\n<li>\u4e0e nnU-Net\uff1annU-Net \u662f\u5f3a pipeline baseline\uff1b\u672c\u6587\u53ea\u6bd4\u8f83 nnU-Net \u6307\u6807\uff0c\u6ca1\u6709\u5438\u6536 nnU-Net \u7684\u81ea\u52a8\u914d\u7f6e\u3001\u9884\u5904\u7406\u3001\u8bad\u7ec3\u7b56\u7565\u3002\u5b9e\u9645\u590d\u73b0\u65f6\u5e94\u8003\u8651\u201c\u7528 nnU-Net pipeline + MHMamba backbone\u201d\u662f\u5426\u66f4\u516c\u5e73\u3002<\/li>\n<li>\u4e0e MedNeXt\uff1a\u6ca1\u6709\u6bd4\u8f83\u3002MedNeXt \u662f\u5f3a CNN baseline\uff0c\u7f3a\u5931\u5b83\u4f1a\u5f71\u54cd\u5bf9\u201cCNN \u4e0d\u8db3\u201d\u7684\u5224\u65ad\u3002<\/li>\n<li>\u4e0e UNetR\/Swin-UNet\/TransUNet\/TransBTS\/nnFormer\uff1a\u672c\u6587\u5c5e\u4e8e\u7528 Mamba \u66ff\u4ee3 Transformer \u7684 3D \u5168\u5c40\u5efa\u6a21\u8def\u7ebf\uff1b\u4e0e nnFormer\u3001TransBTS \u7cfb\u5217\u5173\u7cfb\u5bc6\u5207\uff0c\u4f46\u5b9e\u9a8c\u4e3b\u8981\u5217 nnFormer\u3002<\/li>\n<li>\u4e0e SegMamba\/SegMamba-V2\uff1a\u8fd9\u662f\u6700\u76f4\u63a5 baseline\u3002MHMamba \u53ef\u89c6\u4e3a\u5728 SegMamba \u5f0f 3D Mamba segmentation \u4e0a\u589e\u52a0\u591a\u5934\u62c6\u5206\u3001\u8fb9\u754c\u9884\u589e\u5f3a\u3001\u6821\u51c6\u548c\u95e8\u63a7 skip\u3002<\/li>\n<li>\u4e0e DAMamba\uff1a\u5982\u679c DAMamba \u5173\u6ce8 direction-aware \u6216 dual-branch Mamba\uff0c\u8fd9\u7bc7\u7684 multi-head channel split \u548c AGF skip fusion \u53ef\u4f5c\u4e3a 3D \u7248\u672c\u8bbe\u8ba1\u53c2\u8003\u3002<\/li>\n<li>\u4e0e foundation model\uff1a\u4e0d\u5c5e\u4e8e MedSAM\/SAM-Med3D\/foundation model \u8def\u7ebf\uff0c\u672a\u8ba8\u8bba promptable \u6216 universal segmentation\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u5bf9 3D medical image segmentation\uff1a\u6709\u53c2\u8003\u4ef7\u503c\uff0c\u5c24\u5176\u662f\u591a\u5934 Mamba \u4e0e AGF skip fusion \u7684\u7ed3\u6784\u7ec4\u5408\u3002\u82e5\u7528\u6237\u540e\u7eed\u505a BraTS\u3001multi-organ CT\/MRI \u6216 3D Mamba backbone\uff0c\u53ef\u4ee5\u628a\u5b83\u4f5c\u4e3a\u7ed3\u6784\u8bbe\u8ba1\u5019\u9009\u3002<\/p>\n<p>\u5bf9 polyp segmentation\uff1a\u95f4\u63a5\u4ef7\u503c\u3002MHMamba \u662f 3D BraTS \u8bba\u6587\uff0c\u4e0d\u76f4\u63a5\u670d\u52a1 2D polyp\uff1b\u4f46\u4ee5\u4e0b\u6a21\u5757\u53ef\u8fc1\u79fb\uff1a<\/p>\n<ol>\n<li>Multi-head channel split Mamba\uff1a\u53ef\u7528\u4e8e 2D DAMamba encoder\uff0c\u4f7f\u4e0d\u540c head \u5b66\u4e0d\u540c\u65b9\u5411\/\u5c3a\u5ea6\u54cd\u5e94\u3002<\/li>\n<li>AGF skip fusion\uff1a\u9002\u5408\u66ff\u6362 U-Net\/DAMamba decoder \u7684\u7b80\u5355 concat\uff0c\u5c24\u5176\u7528\u4e8e\u8fb9\u754c\u7ec6\u8282\u548c\u9ad8\u5c42\u8bed\u4e49\u51b2\u7a81\u573a\u666f\u3002<\/li>\n<li>GLA\/Sobel branch\uff1a\u53ef\u5c1d\u8bd5\u5728 polyp \u7684\u8fb9\u754c\u589e\u5f3a\u4e2d\u4f7f\u7528\uff0c\u4f46\u9700\u6ce8\u610f colonoscopy highlights\/reflections \u53ef\u80fd\u4ea7\u751f\u5047\u8fb9\u7f18\u3002<\/li>\n<\/ol>\n<p>\u5bf9 related work\uff1a\u53ef\u653e\u5728 3D Mamba-based medical segmentation \u6bb5\u843d\uff0c\u548c U-Mamba\u3001SegMamba\u3001SegMamba-V2\u3001LS3M \u7b49\u4e00\u8d77\u8ba8\u8bba\uff1b\u4e0d\u5efa\u8bae\u4f5c\u4e3a polyp segmentation \u4e3b baseline\u3002<\/p>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p>\u5efa\u8bae\u9605\u8bfb\uff0c\u4f46\u4e0d\u5efa\u8bae\u628a\u5b83\u4f5c\u4e3a\u4eca\u65e5\u6700\u4f18\u5148\u590d\u73b0\u5bf9\u8c61\u3002\u5efa\u8bae\u91cd\u70b9\u8bfb Methods 3.1\u20133.2 \u548c Tables 1\u20134\uff1b\u5bf9\u7ed3\u8bba\u8981\u4fdd\u7559\uff0c\u56e0\u4e3a\u7f3a\u5c11\u4ee3\u7801\u3001\u6548\u7387\u8868\u548c\u5b98\u65b9 test server \u9a8c\u8bc1\u3002\u82e5\u7528\u6237\u5f53\u524d\u4e3b\u7ebf\u662f DAMamba\/polyp segmentation\uff0c\u5e94\u5148\u8bfb\u8bba\u6587 1\uff1b\u82e5\u4e3b\u7ebf\u6269\u5c55\u5230 3D Mamba\uff0c\u5219\u518d\u7cbe\u8bfb\u5e76\u590d\u73b0 MHM \u4e0e AGF\u3002<\/p>\n<hr \/>\n<h2>\u4eca\u65e5\u63a8\u8350\u4f18\u5148\u7ea7<\/h2>\n<ol>\n<li><strong>Patch-MoE Mamba<\/strong>\uff1a\u6700\u503c\u5f97\u4f18\u5148\u6df1\u5165\u8bfb\u3002\u5b83\u76f4\u63a5\u9762\u5411 polyp segmentation\uff0c\u5e76\u4e14\u63d0\u51fa\u7684 patch-ordered scanning \u4e0e MoE directional fusion \u5bf9 DAMamba\/VM-UNet \u7c7b\u6846\u67b6\u6539\u9020\u975e\u5e38\u76f4\u63a5\uff1b\u867d\u7136\u8ba1\u7b97\u6210\u672c\u504f\u9ad8\uff0c\u4f46\u673a\u5236\u53c2\u8003\u4ef7\u503c\u660e\u786e\u3002<\/li>\n<li><strong>MHMamba<\/strong>\uff1a\u9002\u5408 3D medical image segmentation \u65b9\u5411\u8ddf\u8fdb\u3002\u5b83\u5bf9 BraTS \u548c 3D Mamba backbone \u6709\u542f\u53d1\uff0c\u4f46\u4ee3\u7801\u3001\u6548\u7387\u548c\u5b98\u65b9\u6d4b\u8bd5\u9a8c\u8bc1\u4e0d\u8db3\uff0c\u4e14\u4e0e 2D polyp segmentation \u7684\u5173\u7cfb\u8f83\u95f4\u63a5\u3002<\/li>\n<\/ol>\n<h2>\u4eca\u65e5 PDF \u83b7\u53d6\u60c5\u51b5<\/h2>\n<ul>\n<li>\u8bba\u6587 1\uff1a\u5df2\u6210\u529f\u83b7\u53d6 PDF\uff1b\u53ef\u8bbf\u95ee\u94fe\u63a5\u4e3a https:\/\/arxiv.org\/pdf\/2605.17719 \u3002\u672c\u6b21\u73af\u5883\u5df2\u4e0b\u8f7d\u5230 <code>\/tmp\/medseg_papers\/patch_moe_mamba.pdf<\/code>\u3002MEDIA:\/tmp\/medseg_papers\/patch_moe_mamba.pdf<\/li>\n<li>\u8bba\u6587 2\uff1a\u5df2\u6210\u529f\u83b7\u53d6 PDF\uff1b\u53ef\u8bbf\u95ee\u94fe\u63a5\u4e3a https:\/\/arxiv.org\/pdf\/2605.16464 \u3002\u672c\u6b21\u73af\u5883\u5df2\u4e0b\u8f7d\u5230 <code>\/tmp\/medseg_papers\/mhmamba.pdf<\/code>\u3002MEDIA:\/tmp\/medseg_papers\/mhmamba.pdf<\/li>\n<\/ul>\n<h2>\u4eca\u65e5\u53ef\u6267\u884c\u5efa\u8bae<\/h2>\n<ol>\n<li>\u5982\u679c\u4eca\u5929\u53ea\u7cbe\u8bfb\u4e00\u7bc7\uff0c\u4f18\u5148\u8bfb <strong>Patch-MoE Mamba<\/strong> \u7684 Methods II-B\/II-C \u548c Tables III\/V\uff1a\u5148\u590d\u73b0\u4f4e\u6210\u672c\u7684 patch-ordered scanning\uff0c\u518d\u51b3\u5b9a\u662f\u5426\u52a0\u5165\u5b8c\u6574 MoE fusion\u3002<\/li>\n<li>\u5bf9 DAMamba \u6216 VM-UNet \u6539\u9020\uff0c\u5efa\u8bae\u5c1d\u8bd5\u201cpatch-ordered scan + \u8f7b\u91cf\u56db\u65b9\u5411 gating\u201d\uff0c\u907f\u514d\u76f4\u63a5\u91c7\u7528\u8bba\u6587 1 \u7684\u5b8c\u6574 concat expert\uff0c\u56e0\u4e3a\u5b8c\u6574\u6a21\u578b\u53c2\u6570\/FLOPs \u589e\u5e45\u8fc7\u5927\u3002<\/li>\n<li>\u82e5\u540e\u7eed\u6269\u5c55\u5230 3D medical image segmentation\uff0c\u53ef\u628a <strong>MHMamba<\/strong> \u7684 AGF skip fusion \u548c 4-head Mamba \u4f5c\u4e3a\u5019\u9009\u6a21\u5757\uff0c\u4f46\u5fc5\u987b\u8865\u505a nnU-Net\/MedNeXt\/SegMamba \u7684\u7edf\u4e00 pipeline \u5bf9\u6bd4\u548c\u6548\u7387\u7edf\u8ba1\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\u6ca1\u6709\u68c0\u7d22\u5230\u5df2\u7ecf\u660e\u786e\u8fdb\u5165 MICCAI\/CVPR\/ICCV\/ECCV\/NeurIPS &#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-1077","post","type-post","status-publish","format-standard","hentry","category-85"],"views":12,"_links":{"self":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1077","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=1077"}],"version-history":[{"count":0,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1077\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/media?parent=1077"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/categories?post=1077"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/tags?post=1077"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}