{"id":1076,"date":"2026-05-24T08:36:49","date_gmt":"2026-05-24T00:36:49","guid":{"rendered":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/24\/2026-05-24-%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%9asegguidednet-%e4%b8%8e-medcrp-cl\/"},"modified":"2026-05-24T08:36:49","modified_gmt":"2026-05-24T00:36:49","slug":"2026-05-24-%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%9asegguidednet-%e4%b8%8e-medcrp-cl","status":"publish","type":"post","link":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/24\/2026-05-24-%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%9asegguidednet-%e4%b8%8e-medcrp-cl\/","title":{"rendered":"2026-05-24 \u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\u7cbe\u8bfb\uff1aSegGuidedNet \u4e0e MedCRP-CL"},"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\u5728 2026-05-21 \u81f3 2026-05-19 \u7684\u6700\u65b0 arXiv\/\u4f1a\u8bae\u9884\u5370\u672c\u4e2d\uff0c\u7b5b\u9009\u51fa\u4e24\u7bc7\u6bd4\u201c\u5355\u7eaf\u5806\u6a21\u5757\u201d\u66f4\u503c\u5f97\u8ddf\u8e2a\u7684\u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\uff1a\u4e00\u7bc7\u662f\u9762\u5411 BraTS \u8111\u80bf\u7624 3D \u5206\u5272\u7684\u8f7b\u91cf sub-region attention supervision\uff0c\u53e6\u4e00\u7bc7\u662f ICML 2026 \u63a5\u6536\u7684 continual medical image segmentation \u6846\u67b6\u3002\u6574\u4f53\u8d8b\u52bf\u662f\uff1a\u8fd1\u671f\u533b\u5b66\u5206\u5272\u521b\u65b0\u660e\u663e\u4ece\u5355\u4e2a U-Net\/Transformer\/Mamba block \u6539\u9020\uff0c\u8f6c\u5411\u201c\u53ef\u89e3\u91ca\u76d1\u7763\u3001\u6301\u7eed\u5b66\u4e60\u3001\u4efb\u52a1\u7ed3\u6784\u53d1\u73b0\u3001\u9690\u79c1\u53cb\u597d\u9002\u914d\u201d\u7b49\u66f4\u63a5\u8fd1\u771f\u5b9e\u90e8\u7f72\u7684\u95ee\u9898\u3002<\/p>\n<h2>\u68c0\u7d22\u8bf4\u660e<\/h2>\n<p>\u672c\u6b21\u4f18\u5148\u68c0\u7d22 arXiv \u6700\u65b0\u63d0\u4ea4\u3001medical image segmentation\u30013D medical image segmentation\u3001nnU-Net\u3001polyp segmentation\u3001continual medical image segmentation\u3001Mamba\/Transformer\/U-Net \u7b49\u5173\u952e\u8bcd\uff0c\u5e76\u6838\u5bf9 2025 \u5e74\u4ee5\u540e\u533b\u5b66\u5206\u5272\u76f8\u5173\u5019\u9009\u3002\u5f53\u5929\u672a\u53d1\u73b0\u66f4\u591a\u5df2\u6b63\u5f0f\u4e0a\u7ebf\u9876\u520a\u5b98\u7f51\u4e14\u672a\u91cd\u590d\u7684\u5168\u65b0\u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\uff0c\u56e0\u6b64\u4ece 2026 \u5e74 5 \u6708\u6700\u65b0 arXiv\/ICML accepted preprint \u4e2d\u9009\u62e9\u4e24\u7bc7\uff1b\u6240\u6709\u5165\u9009\u8bba\u6587\u5747\u4e3a 2025 \u5e74\u53ca\u4ee5\u540e\u3002\u5df2\u68c0\u67e5\u5386\u53f2\u63a8\u8350\u8bb0\u5f55\u5e76\u6392\u9664\u4e86\u91cd\u590d\u8bba\u6587\uff1b\u672c\u6b21\u8df3\u8fc7\u7684\u5386\u53f2\u91cd\u590d\u5019\u9009\u5305\u62ec Patch-MoE Mamba\u3001DepthPolyp\u3001Semi-MedRef\u3001Beyond Euclidean Prototypes\u3001PanGuide3D\u3001FEFormer\u3001USEMA\u3001MedCore \u7b49\u3002<\/p>\n<h2>WordPress \u53d1\u5e03<\/h2>\n<ul>\n<li>WordPress \u6587\u7ae0\u94fe\u63a5\uff1a\u5f85\u53d1\u5e03\u540e\u56de\u586b<\/li>\n<li>WordPress Post ID\uff1a\u5f85\u53d1\u5e03\u540e\u56de\u586b<\/li>\n<\/ul>\n<hr \/>\n<h2>\u8bba\u6587 1\uff1aSegGuidedNet: Sub-Region-Aware Attention Supervision for Interpretable Brain Tumour Segmentation<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aSegGuidedNet: Sub-Region-Aware Attention Supervision for Interpretable Brain Tumour Segmentation<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aHasaan Maqsood, Saif Ur Rehman Khan, Sebastian Vollmer, Andreas Dengel, Muhammad Nabeel Asim \/ Hasaan Maqsood<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-21 arXiv v1<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint\uff0carXiv:2605.22572<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.22572<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1aMEDIA:\/tmp\/medseg_daily_2026-05-24\/segguidednet_2605.22572.pdf\uff1bhttps:\/\/arxiv.org\/pdf\/2605.22572<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1a\u672a\u516c\u5f00\uff1b\u8bba\u6587\u58f0\u660e \u201cwill be available on GitHub upon publication\u201d\uff0c\u5f53\u524d\u672a\u786e\u8ba4\u5b98\u65b9\u4ed3\u5e93<\/li>\n<li>\u4efb\u52a1\uff1a\u591a\u53c2\u6570 MRI \u8111\u80bf\u7624 3D \u5206\u5272\uff1bBraTS ET\/TC\/WT \u590d\u5408\u533a\u57df\u4e0e NCR\/ED\/ET \u5b50\u533a\u57df\u5206\u5272<\/li>\n<li>\u6570\u636e\u96c6\uff1aBraTS 2021\u3001BraTS 2023 GLI\uff1b\u6bcf\u4e2a\u6570\u636e\u96c6 1,251 subjects\uff0c70\/10\/20 \u5212\u5206\uff0cheld-out test \u4e3a 251 subjects<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1a3D residual U-Net \/ encoder-decoder + sub-region attention supervision + auxiliary attention loss\uff1b\u504f CNN-based interpretable 3D medical segmentation<\/li>\n<\/ul>\n<h3>paper-deep-reader \u7cbe\u8bfb\u7ed3\u679c<\/h3>\n<h4>1. \u4e00\u53e5\u8bdd\u7ed3\u8bba<\/h4>\n<p>SegGuidedNet \u7684\u4e3b\u8981\u4ef7\u503c\u4e0d\u5728\u4e8e\u53d1\u660e\u4e00\u4e2a\u5168\u65b0 backbone\uff0c\u800c\u5728\u4e8e\u7528\u6781\u8f7b\u91cf\u7684\u5b50\u533a\u57df attention \u8f85\u52a9\u76d1\u7763\uff0c\u628a BraTS \u8111\u80bf\u7624\u5206\u5272\u4e2d ET\u3001TC\u3001WT \u80cc\u540e\u7684 NCR\/ED\/ET \u53ef\u533a\u5206\u6027\u663e\u5f0f\u538b\u8fdb decoder\uff0c\u5f97\u5230\u4e00\u4e2a\u6bd4\u590d\u6742 ensemble \u66f4\u4fbf\u5b9c\u3001\u540c\u65f6\u5e26\u6709\u5185\u751f\u53ef\u89e3\u91ca\u56fe\u7684 3D U-Net \u53d8\u4f53\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u8bba\u6587\u7814\u7a76\u591a\u53c2\u6570 MRI \u8111\u80bf\u7624\u5206\u5272\u3002\u8f93\u5165\u662f 4 \u4e2a\u5171\u914d\u51c6 MRI \u5e8f\u5217 T1\u3001T1ce\u3001T2\u3001FLAIR\uff0c\u76ee\u6807\u662f\u9884\u6d4b voxel-wise label map\uff1abackground\u3001NCR\u3001ED\u3001ET\uff1b\u8bc4\u4f30\u65f6\u6309 BraTS \u6807\u51c6\u6c47\u603b\u4e3a Enhancing Tumour\uff08ET\uff09\u3001Tumour Core\uff08TC=NCR\u222aET\uff09\u3001Whole Tumour\uff08WT=NCR\u222aED\u222aET\uff09\u3002<\/p>\n<p>\u8fd9\u4e2a\u95ee\u9898\u91cd\u8981\u6709\u4e09\u70b9\uff1a\u7b2c\u4e00\uff0c\u8111\u80bf\u7624\u5b50\u533a\u57df\u76f4\u63a5\u670d\u52a1\u4e8e\u6cbb\u7597\u8ba1\u5212\u3001\u7597\u6548\u8bc4\u4f30\u548c\u795e\u7ecf\u80bf\u7624\u5b66\u51b3\u7b56\uff1b\u7b2c\u4e8c\uff0cBraTS \u5df2\u7ecf\u6210\u4e3a 3D \u533b\u5b66\u5206\u5272\u7684\u91cd\u8981\u6807\u51c6 benchmark\uff0c\u5f3a baseline \u5305\u62ec nnU-Net\u3001Swin UNETR\u3001HNF-Netv2 \u7b49\uff1b\u7b2c\u4e09\uff0cNCR\u3001ED\u3001ET \u4e0d\u662f\u7b80\u5355\u4e09\u7c7b\u50cf\u7d20\uff0c\u5b83\u4eec\u6709\u5d4c\u5957\u7ed3\u6784\u3001\u4f53\u79ef\u4e0d\u5747\u8861\u548c MRI \u4fe1\u53f7\u91cd\u53e0\uff0c\u5c24\u5176 NCR \u4e0e ET \u5728 T1ce \u4e2d\u5916\u89c2\u76f8\u8fd1\uff0cET \u53c8\u901a\u5e38\u53ea\u5360\u5c11\u91cf\u524d\u666f\u4f53\u7d20\u3002<\/p>\n<p>paper map\uff1a\u8bba\u6587\u7814\u7a76 BraTS \u591a\u53c2\u6570 MRI \u8111\u80bf\u7624 3D \u5b50\u533a\u57df\u5206\u5272\uff1b\u4e3b\u52a8\u4f5c\u662f\u5728 residual 3D U-Net decoder \u6700\u540e\u4e00\u5c42\u5e76\u8054\u4e00\u4e2a SegAttentionGate\uff0c\u7528\u6807\u51c6 segmentation mask \u751f\u6210 NCR\/ED\/ET \u4e09\u4e2a\u4e8c\u503c attention supervision\uff1b\u4e3b\u5f20\u662f\u8fd9\u80fd\u63d0\u5347\u5b50\u533a\u57df\u53ef\u5206\u6027\u5e76\u63d0\u4f9b\u65e0\u989d\u5916\u63a8\u7406\u6210\u672c\u7684\u53ef\u89e3\u91ca attention map\uff1b\u8bc1\u636e\u4e3b\u8981\u6765\u81ea BraTS 2021\/2023 GLI held-out test \u7684 Dice\/HD95\u3001\u4e0e nnU-Net\/HNF-Netv2\/Swin UNETR \u7684\u6bd4\u8f83\u3001attention map \u53ef\u89c6\u5316\u548c failure case\uff1b\u4e3b\u8981\u5931\u8d25\u98ce\u9669\u662f\u7f3a\u5c11\u4e25\u683c ablation\u3001\u7f3a\u5c11\u5916\u90e8\u8de8\u4e2d\u5fc3\u6d4b\u8bd5\/\u4ee3\u7801\u590d\u73b0\uff0c\u4e14\u4e0e SOTA \u7684\u6bd4\u8f83\u53ef\u80fd\u53d7\u8bad\u7ec3\u5212\u5206\u3001ensemble \u8bbe\u7f6e\u548c\u5f15\u7528\u6570\u503c\u4e00\u81f4\u6027\u5f71\u54cd\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u8ba4\u4e3a\u6807\u51c6 U-Net\/3D U-Net\/nnU-Net \u7c7b\u65b9\u6cd5\u867d\u7136\u5f3a\uff0c\u4f46\u53ea\u7528 voxel-wise segmentation loss \u8bad\u7ec3 decoder\uff0c\u6ca1\u6709\u663e\u5f0f\u8981\u6c42\u6700\u7ec8 decoder feature \u5bf9\u4e0d\u540c\u80bf\u7624\u5b50\u533a\u57df\u4fdd\u6301\u7a7a\u95f4\u53ef\u533a\u5206\u6027\u3002BraTS \u7684\u96be\u70b9\u5728\u4e8e\uff1a<\/p>\n<ul>\n<li><strong>\u5916\u89c2\u91cd\u53e0<\/strong>\uff1aNCR \u4e0e ET \u5728 T1ce \u4e2d\u53ef\u80fd\u90fd\u5448\u9ad8\u4fe1\u53f7\uff0cdecoder \u5bb9\u6613\u6df7\u6dc6\u3002<\/li>\n<li><strong>\u4e25\u91cd\u7c7b\u522b\u4e0d\u5747\u8861<\/strong>\uff1aET \u5c0f\uff0cNCR \u66f4\u5c0f\u4e14\u8fb9\u754c\u4e0d\u89c4\u5219\uff0c\u4ec5\u9760 Dice+CE \u4e0d\u4e00\u5b9a\u8db3\u591f\u3002<\/li>\n<li><strong>\u7ed3\u6784\u5d4c\u5957<\/strong>\uff1aWT\u3001TC\u3001ET \u662f\u5c42\u7ea7\/\u590d\u5408\u533a\u57df\uff0c\u6a21\u578b\u53ef\u80fd\u5b66\u5230\u4e00\u4e2a\u201c\u80bf\u7624\u6574\u4f53\u54cd\u5e94\u201d\uff0c\u800c\u4e0d\u662f\u53ef\u5206\u7684\u5b50\u533a\u57df\u54cd\u5e94\u3002<\/li>\n<li><strong>\u73b0\u6709 SOTA \u4f9d\u8d56 ensemble \u6216\u590d\u6742\u7ed3\u6784<\/strong>\uff1annU-Net\u3001HNF-Netv2\u3001Swin UNETR \u7b49\u5e38\u501f\u52a9\u591a\u6a21\u578b ensemble\u3001TTA \u6216\u8f83\u91cd\u7684 Transformer\/CNN \u6df7\u5408\u7ed3\u6784\uff0c\u63a8\u7406\u6210\u672c\u9ad8\uff0c\u4e14\u89e3\u91ca\u6027\u901a\u5e38\u4f9d\u8d56 post-hoc \u65b9\u6cd5\u3002<\/li>\n<\/ul>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>SegGuidedNet \u662f\u4e00\u4e2a residual 3D U-Net \u98ce\u683c\u7684 encoder-decoder\u3002\u8f93\u5165\u4e3a\u56db\u901a\u9053 MRI patch\uff0c\u5f62\u72b6\u53ef\u6982\u62ec\u4e3a <code>(B,4,128,128,128)<\/code>\uff1b\u8f93\u51fa\u5305\u62ec\uff1a<\/p>\n<ol>\n<li>\u56db\u7c7b segmentation logits\uff1a<code>Lseg \u2208 R^{B\u00d74\u00d7128^3}<\/code>\uff1b<\/li>\n<li>\u4e09\u4e2a\u5b50\u533a\u57df attention maps\uff1a<code>A \u2208 [0,1]^{B\u00d73\u00d7128^3}<\/code>\uff0c\u5206\u522b\u5bf9\u5e94 NCR\u3001ED\u3001ET\u3002<\/li>\n<\/ol>\n<p>\u4e3b\u5e72\u7f51\u7edc\uff1a<br \/>\n- encoder \u6709 4 \u4e2a EncoderBlocks\uff0c\u6bcf\u5c42\u4e3a residual 3\u00d73\u00d73 convolution + InstanceNorm + LeakyReLU + 2\u00d72\u00d72 max-pooling\uff1b<br \/>\n- \u901a\u9053\u5bbd\u5ea6\u7ea6\u4e3a <code>[32,64,128,256]<\/code>\uff0cbottleneck \u4e3a 320 channels\uff1b<br \/>\n- decoder \u7528 transposed convolution \u4e0a\u91c7\u6837\uff0c\u5e76\u4e0e\u5bf9\u5e94 encoder skip feature concat\uff1b<br \/>\n- \u4f7f\u7528 InstanceNorm \u800c\u975e BatchNorm\uff0c\u4ee5\u9002\u5e94 batch size=2 \u4e0e\u591a\u4e2d\u5fc3 MRI intensity variability\uff1b<br \/>\n- \u603b\u53c2\u6570\u91cf\u7ea6 7.8M\u3002<\/p>\n<p>\u6838\u5fc3\u76ee\u6807\u51fd\u6570\uff1a<\/p>\n<p>[<br \/>\nL = L_{seg} + \\lambda L_{attn}, \\quad \\lambda=0.1<br \/>\n]<\/p>\n<p>\u5176\u4e2d <code>Lseg = Dice loss + Cross entropy<\/code>\uff0cDice \u5728 foreground classes \u4e0a\u8ba1\u7b97\uff1b<code>Lattn<\/code> \u662f\u4e09\u4e2a attention map \u4e0e NCR\/ED\/ET \u4e8c\u503c mask \u7684\u5e73\u5747 BCE\u3002\u6ce8\u610f\uff1a<code>Lattn<\/code> \u4e0d\u9700\u8981\u989d\u5916\u6807\u6ce8\uff0c\u56e0\u4e3a\u4e8c\u503c\u5b50\u533a\u57df mask \u76f4\u63a5\u7531 BraTS \u6807\u51c6\u6807\u7b7e\u751f\u6210\u3002<\/p>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<p><strong>\u6a21\u5757 1\uff1aResidual 3D U-Net backbone<\/strong><br \/>\n- \u8f93\u5165\uff1a4 \u901a\u9053 mpMRI 3D patch\u3002<br \/>\n- \u8f93\u51fa\uff1a\u6700\u7ec8 decoder feature <code>d1 \u2208 R^{B\u00d732\u00d7128^3}<\/code> \u4e0e segmentation logits\u3002<br \/>\n- \u89e3\u51b3\u95ee\u9898\uff1a\u63d0\u4f9b\u5f3a\u800c\u7ecf\u5178\u7684 3D \u533b\u5b66\u5206\u5272\u57fa\u5ea7\u3002<br \/>\n- \u521b\u65b0\u6027\u5224\u65ad\uff1a\u4e3b\u5e72\u672c\u8eab\u4e0d\u65b0\uff0c\u5c5e\u4e8e\u53ef\u590d\u73b0\u3001\u53ef\u66ff\u6362\u7684\u5f3a baseline \u7ed3\u6784\u3002<br \/>\n- \u53ef\u8fc1\u79fb\u6027\uff1a\u9ad8\u3002\u53ef\u66ff\u6362\u4e3a nnU-Net\/MedNeXt\/SegResNet \u7684 decoder feature \u540e\u63a5\u7c7b\u4f3c attention supervision\u3002<\/p>\n<p><strong>\u6a21\u5757 2\uff1aSegAttentionGate<\/strong><br \/>\n- \u8f93\u5165\uff1a\u6700\u7ec8 decoder feature <code>d1<\/code>\u3002<br \/>\n- \u64cd\u4f5c\uff1a3\u00d73\u00d73 conv \u5c06 32 \u901a\u9053\u538b\u5230 16 \u901a\u9053\uff0cInstanceNorm + LeakyReLU\uff0c\u518d 1\u00d71\u00d71 conv \u8f93\u51fa 3 \u901a\u9053\uff0csigmoid \u5f97\u5230 NCR\/ED\/ET attention maps\u3002<br \/>\n- \u516c\u5f0f\uff1a<code>Lattn = W1(phi(IN(W3(d1))))<\/code>\uff0c<code>A = sigmoid(Lattn)<\/code>\u3002<br \/>\n- \u8f93\u51fa\uff1a\u4e09\u4e2a\u5b50\u533a\u57df\u7a7a\u95f4\u6982\u7387\/attention map\u3002<br \/>\n- \u89e3\u51b3\u95ee\u9898\uff1a\u7ed9 decoder \u4e00\u4e2a\u663e\u5f0f\u7684\u201c\u5b50\u533a\u57df\u5e94\u5728\u54ea\u91cc\u6fc0\u6d3b\u201d\u7684\u68af\u5ea6\u4fe1\u53f7\uff0c\u907f\u514d\u53ea\u5b66\u5230\u4e00\u4e2a\u6a21\u7cca\u80bf\u7624\u6574\u4f53\u54cd\u5e94\u3002<br \/>\n- \u521b\u65b0\u6027\u5224\u65ad\uff1a\u7ed3\u6784\u5f88\u7b80\u5355\uff0c\u521b\u65b0\u4e3b\u8981\u662f\u4efb\u52a1\u5f52\u7eb3\u504f\u7f6e\u548c\u76d1\u7763\u4f4d\u7f6e\uff0c\u800c\u4e0d\u662f\u590d\u6742\u6a21\u5757\u8bbe\u8ba1\u3002\u4f18\u70b9\u662f\u4f4e\u6210\u672c\u3001\u5bb9\u6613\u590d\u73b0\uff1b\u7f3a\u70b9\u662f\u82e5\u6ca1\u6709\u5145\u5206\u6d88\u878d\uff0c\u96be\u5224\u65ad\u6027\u80fd\u6765\u81ea attention supervision \u8fd8\u662f\u8bad\u7ec3 recipe\u3002<br \/>\n- \u5bf9 polyp segmentation \u7684\u8fc1\u79fb\uff1a\u82e5\u53ea\u6709 binary polyp mask\uff0c\u76f4\u63a5\u7684 NCR\/ED\/ET \u591a\u5b50\u533a\u57df\u76d1\u7763\u4e0d\u5b58\u5728\uff1b\u4f46\u53ef\u4ee5\u628a\u606f\u8089\u5206\u6210 boundary\/interior\/uncertain rim\uff0c\u6216\u7ed3\u5408 distance transform \u751f\u6210\u8fb9\u754c attention supervision\u3002<br \/>\n- \u5bf9 3D medical segmentation \u7684\u8fc1\u79fb\uff1a\u9002\u5408\u591a\u7ed3\u6784\u3001\u591a\u5b50\u533a\u57df\u3001\u5c42\u7ea7\u6807\u7b7e\u4efb\u52a1\uff0c\u5982\u809d\u80bf\u7624 core\/rim\u3001\u80f0\u817a\/\u80bf\u7624\u3001\u5fc3\u810f\u591a\u8154\u5ba4\u3001\u7259\u9f7f\/\u7259\u6839\/\u7259\u69fd\u9aa8\u7b49\u3002<\/p>\n<p><strong>\u6a21\u5757 3\uff1aattention auxiliary loss<\/strong><br \/>\n- \u8f93\u5165\uff1aattention logits\/maps \u4e0e ground-truth \u5b50\u533a\u57df\u4e8c\u503c mask\u3002<br \/>\n- \u8f93\u51fa\uff1aBCE loss\uff0c\u6743\u91cd \u03bb=0.1\u3002<br \/>\n- \u89e3\u51b3\u95ee\u9898\uff1a\u5728\u4e0d\u6539\u53d8\u4e3b\u5206\u5272\u5934\u7684\u60c5\u51b5\u4e0b\uff0c\u4e3a shared decoder feature \u6dfb\u52a0\u5b50\u533a\u57df\u5224\u522b\u7ea6\u675f\u3002<br \/>\n- \u662f\u5426\u771f\u6b63\u6709\u521b\u65b0\uff1a\u5c5e\u4e8e\u5c0f\u800c\u6709\u6548\u7684\u76d1\u7763\u8bbe\u8ba1\uff0c\u4e0d\u662f\u7406\u8bba\u7a81\u7834\u3002\u771f\u6b63\u4ef7\u503c\u5728\u4e8e\u5b83\u201c\u53ef\u63d2\u62d4\u3001\u4f4e\u53c2\u6570\u3001\u4f4e\u63a8\u7406\u6210\u672c\u201d\u3002<\/p>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p>\u5b9e\u9a8c\u8bbe\u7f6e\uff1a<br \/>\n- \u6570\u636e\uff1aBraTS 2021 \u548c BraTS 2023 GLI\uff0c\u6bcf\u4e2a 1,251 subjects\uff0c875 train \/ 125 val \/ 251 test\u3002<br \/>\n- \u8f93\u5165\uff1a128\u00b3 patch\uff0cforeground-centered random crop probability 0.8\uff1b\u6d4b\u8bd5\u4e3a deterministic centre-cropping\u3002<br \/>\n- \u8bad\u7ec3\uff1aAdamW\uff0cLR 1e-4\uff0cweight decay 1e-5\uff0ccosine annealing \u5230 1e-6\uff1b50 epochs\uff1bbatch size 2\uff1bAMP\uff1b\u5355\u5f20 NVIDIA A100 80GB\uff1bseed=42\uff1bdeterministic CuDNN\u3002<br \/>\n- \u589e\u5f3a\uff1aflips\u300190\u00b0 rotations\u3001elastic deformation\u3001intensity scaling\u3001brightness shift\u3001Gaussian noise\/blur\u3001channel dropout\u3002<br \/>\n- \u6307\u6807\uff1aDSC\u3001HD95\u3001sensitivity\u3001specificity\uff1b\u62a5\u544a ET\/TC\/WT\uff0c\u4e5f\u62a5\u544a attention \u5b50\u533a\u57df Dice\u3002<\/p>\n<p>\u4e3b\u8981\u7ed3\u679c\uff1a<br \/>\n- BraTS 2021 test\uff1amean Dice 0.905\uff1bET 0.873\uff0cTC 0.906\uff0cWT 0.935\uff1bHD95 \u5747\u4f4e\u4e8e 4.0 mm\u3002<br \/>\n- BraTS 2023 GLI test\uff1amean Dice 0.897\uff1bET 0.859\uff0cTC 0.902\uff0cWT 0.931\uff1bHD95 \u6700\u9ad8\u7ea6 5.12 mm\u3002<br \/>\n- \u4e0e SOTA \u6bd4\u8f83\uff1a\u8bba\u6587\u8868 10 \u62a5\u544a SegGuidedNet \u5728 BraTS 2021 \u4e0a\u4f18\u4e8e nnU-Net \u7684 ET\/TC\/WT\uff080.820\/0.851\/0.890\uff09\u548c HNF-Netv2 \u7684 TC\/WT\uff0c\u4f46\u4f4e\u4e8e Swin UNETR ensemble\uff080.920\/0.930\/0.940\uff09\uff1b\u4f5c\u8005\u5f3a\u8c03\u81ea\u5df1\u662f single model\uff0c\u65e0 ensemble\u3002<br \/>\n- \u5b9a\u6027\u7ed3\u679c\uff1abest\/median case \u5206\u5272\u8f83\u51c6\u786e\uff0cworst case \u4e3b\u8981\u5728 small\/diffuse NCR \u548c\u4e0d\u89c4\u5219\u8fb9\u754c\u4e0a\u51fa\u9519\uff1battention map \u4e0e GT \u5b50\u533a\u57df\u6709\u8f83\u597d\u7a7a\u95f4\u5bf9\u5e94\u3002<\/p>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u53ef\u4fe1\u70b9\uff1a<br \/>\n- \u4f7f\u7528\u4e24\u4e2a BraTS benchmark edition\uff0c\u4e14\u6bcf\u4e2a\u90fd\u6709 251 held-out test subjects\uff0c\u6bd4\u53ea\u505a\u5355\u4e00\u5c0f\u6570\u636e\u96c6\u66f4\u53ef\u4fe1\u3002<br \/>\n- \u6307\u6807\u8986\u76d6 Dice \u4e0e HD95\uff0c\u65e2\u770b overlap \u4e5f\u770b\u8fb9\u754c\u8bef\u5dee\u3002<br \/>\n- \u62a5\u544a\u8bad\u7ec3\u7ec6\u8282\u8f83\u5b8c\u6574\uff0c\u5305\u62ec patch size\u3001\u4f18\u5316\u5668\u3001\u589e\u5f3a\u3001\u786c\u4ef6\u3001seed\u3002<br \/>\n- \u6a21\u5757\u53c2\u6570\u6781\u5c11\uff08\u7ea6 14k\uff0c&lt;0.2%\uff09\uff0c\u5982\u679c\u7ed3\u679c\u53ef\u590d\u73b0\uff0c\u786e\u5b9e\u6709\u8f83\u9ad8\u5de5\u7a0b\u6027\u4ef7\u6bd4\u3002<\/p>\n<p>\u9700\u8981\u8c28\u614e\u7684\u70b9\uff1a<br \/>\n- \u8bba\u6587\u6ca1\u6709\u770b\u5230\u4e25\u683c\u7684 \u201cw\/o SegAttentionGate \/ w\/o Lattn \/ \u4e0d\u540c \u03bb \/ attention branch \u4f4d\u7f6e\u201d \u91cf\u5316\u6d88\u878d\u8868\u3002\u6ca1\u6709\u8fd9\u4e2a\u6d88\u878d\uff0c\u4e0d\u80fd\u5b8c\u5168\u786e\u8ba4\u63d0\u5347\u6765\u81ea\u6838\u5fc3\u6a21\u5757\uff0c\u800c\u4e0d\u662f\u8bad\u7ec3\u7b56\u7565\u3001\u5212\u5206\u6216\u5b9e\u73b0\u5dee\u5f02\u3002<br \/>\n- \u4e0e nnU-Net\/Swin UNETR\/HNF-Netv2 \u7684\u6bd4\u8f83\u4e3b\u8981\u662f\u5f15\u7528\/\u8868\u683c\u6bd4\u8f83\uff0c\u4e0d\u4e00\u5b9a\u662f\u540c\u4e00\u8bad\u7ec3\u5212\u5206\u3001\u540c\u4e00\u9884\u5904\u7406\u3001\u540c\u4e00 inference protocol \u4e0b\u7684\u516c\u5e73\u590d\u8dd1\u3002<br \/>\n- \u4ee3\u7801\u5c1a\u672a\u516c\u5f00\uff0c\u590d\u73b0\u6027\u6682\u65f6\u53d7\u9650\u3002<br \/>\n- Data availability \u5199 \u201cavailable from corresponding author upon request\u201d\uff0c\u800c BraTS \u6570\u636e\u672c\u8eab\u53ef\u7533\u8bf7\uff0c\u4f46\u5177\u4f53\u5212\u5206\u4e0e\u5b9e\u73b0\u9700\u8981\u4ee3\u7801\u652f\u6301\u3002<br \/>\n- \u53ea\u9a8c\u8bc1\u8111\u80bf\u7624 mpMRI\uff0c\u4e0d\u80fd\u76f4\u63a5\u63a8\u5e7f\u5230\u606f\u8089\u3001\u8179\u90e8\u591a\u5668\u5b98\u3001CT \u80bf\u7624\u6216\u8de8\u57df\u6cdb\u5316\u3002<\/p>\n<p>\u7ed3\u8bba\u5f3a\u5ea6\uff1a\u53ef\u4ee5\u76f8\u4fe1\u201c\u5b50\u533a\u57df attention supervision \u662f\u4e00\u4e2a\u4f4e\u6210\u672c\u3001\u503c\u5f97\u5c1d\u8bd5\u7684 3D \u5206\u5272\u8f85\u52a9\u76d1\u7763\u601d\u8def\u201d\uff1b\u6682\u4e0d\u5e94\u628a\u5b83\u89c6\u4e3a\u5df2\u5145\u5206\u8bc1\u660e\u7684 BraTS SOTA \u6216\u901a\u7528\u533b\u5b66\u5206\u5272 backbone\u3002<\/p>\n<h4>8. \u4e0e\u4e3b\u6d41\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u7684\u5173\u7cfb<\/h4>\n<ul>\n<li>\u4e0e <strong>U-Net\/3D U-Net\/nnU-Net<\/strong>\uff1aSegGuidedNet \u672c\u8d28\u662f residual 3D U-Net \u7684\u76d1\u7763\u589e\u5f3a\u7248\uff0c\u4e0d\u662f\u63a8\u7ffb U-Net\u3002\u5b83\u66f4\u50cf\u4e00\u4e2a\u53ef\u4ee5\u63d2\u5230 nnU-Net decoder \u672b\u7aef\u7684 auxiliary supervision head\u3002<\/li>\n<li>\u4e0e <strong>MedNeXt \/ CNN-based segmentation<\/strong>\uff1a\u601d\u8def\u53ef\u8fc1\u79fb\u5230 MedNeXt\u3001SegResNet\u3001ConvNeXt-like 3D encoder-decoder\uff0c\u56e0\u4e3a\u5b83\u4f9d\u8d56\u6700\u7ec8 decoder feature\uff0c\u800c\u975e\u7279\u5b9a backbone\u3002<\/li>\n<li>\u4e0e <strong>UNETR\/Swin-UNETR\/TransUNet<\/strong>\uff1a\u8bba\u6587\u9009\u62e9\u8f7b\u91cf CNN \u8def\u7ebf\uff0c\u5f3a\u8c03\u4f4e\u63a8\u7406\u6210\u672c\uff1b\u76f8\u6bd4 Transformer\uff0c\u5b83\u4e0d\u8ffd\u6c42\u5168\u5c40\u5efa\u6a21\uff0c\u800c\u662f\u663e\u5f0f\u589e\u5f3a\u5c40\u90e8\/\u5b50\u533a\u57df\u5224\u522b\u3002<\/li>\n<li>\u4e0e <strong>Mamba\/VMamba\/SegMamba\/DAMamba<\/strong>\uff1a\u6ca1\u6709\u72b6\u6001\u7a7a\u95f4\u6a21\u5757\uff1b\u4f46\u5bf9 DAMamba \u6539\u9020\u6709\u542f\u53d1\uff1a\u53ef\u5728 Mamba decoder \u6216 hybrid CNN-Mamba decoder \u540e\u63a5 task-specific attention supervision\uff0c\u4ee5\u5f25\u8865\u957f\u7a0b\u5efa\u6a21\u4e4b\u5916\u7684\u5b50\u7ed3\u6784\u53ef\u5206\u6027\u3002<\/li>\n<li>\u4e0e <strong>foundation model for medical segmentation<\/strong>\uff1a\u4e0d\u662f foundation model\uff1b\u4f46 attention maps \u53ef\u4f5c\u4e3a prompt\/\u89e3\u91ca\/QA \u7684\u4e2d\u95f4\u76d1\u7763\u4fe1\u53f7\uff0c\u672a\u6765\u53ef\u4e0e MedSAM\/SAM-Med3D \u7684 mask decoder \u7ed3\u5408\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<ul>\n<li><strong>\u5bf9 polyp segmentation<\/strong>\uff1a\u76f4\u63a5\u4ef7\u503c\u4e2d\u7b49\u3002\u606f\u8089\u901a\u5e38\u662f binary segmentation\uff0c\u6ca1\u6709\u5929\u7136 NCR\/ED\/ET \u5b50\u533a\u57df\uff1b\u4f46\u53ef\u4ee5\u501f\u9274\u4e3a boundary\/interior \u6216 uncertain-boundary attention supervision\uff0c\u5c24\u5176\u9002\u5408\u5f31\u8fb9\u754c\u3001\u4f4e\u5bf9\u6bd4\u606f\u8089\u3002<\/li>\n<li><strong>\u5bf9 DAMamba \u6539\u9020<\/strong>\uff1a\u4ef7\u503c\u8f83\u9ad8\u3002DAMamba \u82e5\u5df2\u7ecf\u89e3\u51b3\u957f\u7a0b\u4f9d\u8d56\uff0c\u53ef\u518d\u52a0\u5165\u8f7b\u91cf auxiliary attention head\uff0c\u8ba9 decoder \u663e\u5f0f\u5b66\u4e60 lesion core\/boundary\/background \u6216 multi-scale anatomical region\u3002<\/li>\n<li><strong>\u5bf9 3D medical image segmentation<\/strong>\uff1a\u4ef7\u503c\u8f83\u9ad8\uff0c\u5c24\u5176\u662f\u591a\u7ed3\u6784\u3001\u591a\u5b50\u533a\u57df\u4efb\u52a1\u3002\u6a21\u5757\u7b80\u5355\u3001\u53c2\u6570\u5c11\uff0c\u9002\u5408\u4f5c\u4e3a ablation-friendly \u63d2\u4ef6\u3002<\/li>\n<li><strong>\u5bf9 related work<\/strong>\uff1a\u53ef\u653e\u5728\u201cinterpretable \/ auxiliary-supervised 3D segmentation\u201d\u6216\u201cbrain tumor segmentation efficient single-model design\u201d\u76f8\u5173\u6bb5\u843d\u3002<\/li>\n<li><strong>\u5bf9 baseline<\/strong>\uff1a\u82e5\u4ee3\u7801\u516c\u5f00\u540e\uff0c\u503c\u5f97\u4f5c\u4e3a BraTS\/3D tumor segmentation baseline\uff1b\u5f53\u524d\u53ea\u80fd\u4f5c\u4e3a\u65b9\u6cd5\u53c2\u8003\uff0c\u4e0d\u80fd\u7acb\u5373\u4f5c\u4e3a\u53ef\u590d\u73b0\u5b9e\u9a8c\u57fa\u7ebf\u3002<\/li>\n<\/ul>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5efa\u8bae\u7cbe\u8bfb\uff0c\u4f46\u5e26\u7740\u6000\u7591\u8bfb\u5b9e\u9a8c\u90e8\u5206\u3002<\/strong>\u65b9\u6cd5\u90e8\u5206\u5f88\u77ed\uff0c\u6700\u503c\u5f97\u8bfb\u7684\u662f problem formalization\u3001SegAttentionGate \u548c loss \u8bbe\u8ba1\uff1b\u5b9e\u9a8c\u90e8\u5206\u91cd\u70b9\u68c0\u67e5\u662f\u5426\u8865\u5145\u4e86 ablation\u3002\u82e5\u4f60\u8981\u505a DAMamba \u6216 U-Net \u7c7b\u533b\u5b66\u5206\u5272\u6539\u9020\uff0c\u8fd9\u7bc7\u9002\u5408\u4f5c\u4e3a\u201c\u4f4e\u6210\u672c auxiliary supervision\u201d\u7075\u611f\u6765\u6e90\uff0c\u800c\u4e0d\u662f\u76f4\u63a5\u4f5c\u4e3a SOTA \u8bc1\u636e\u3002<\/p>\n<hr \/>\n<h2>\u8bba\u6587 2\uff1aMedCRP-CL: Continual Medical Image Segmentation via Bayesian Nonparametric Semantic Modality Discovery<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aMedCRP-CL: Continual Medical Image Segmentation via Bayesian Nonparametric Semantic Modality Discovery<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aZiyuan Gao \/ Ziyuan Gao<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-19 arXiv v1<\/li>\n<li>\u6765\u6e90\uff1aICML 2026 accepted paper \/ arXiv preprint\uff0carXiv:2605.20297<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.20297<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1aMEDIA:\/tmp\/medseg_daily_2026-05-24\/medcrp_cl_2605.20297.pdf\uff1bhttps:\/\/arxiv.org\/pdf\/2605.20297<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/zygao930\/MedCRP-CL<\/li>\n<li>\u4efb\u52a1\uff1aContinual medical image segmentation\uff1b\u591a\u4efb\u52a1\u3001\u8de8\u6a21\u6001\u3001\u9690\u79c1\u53cb\u597d\u3001replay-free sequential segmentation learning<\/li>\n<li>\u6570\u636e\u96c6\uff1a16 \u4e2a\u4efb\u52a1\uff0c\u8986\u76d6 endoscopy polyp\uff08Kvasir\u3001ClinicDB\u3001ETIS\u3001CVC-300\u3001ColonDB\uff09\u3001dermoscopy ISIC\u3001ultrasound\uff08CAMUS\u3001BUSI benign\/malignant\uff09\u3001chest X-ray pathology localization\/segmentation subsets\uff08Airspace Opacity\u3001Atelectasis\u3001Cardiomegaly\u3001Edema\u3001Pleural Effusion\u3001Enlarged Cardiomediastinum\u3001Support Devices\uff09<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1aVision-language segmentation backbone\uff08CLIPSeg\uff09+ prompt-based CRP semantic modality discovery + modality-specific LoRA + intra-modality EWC\uff1b\u504f continual learning \/ foundation adaptation \/ method-algorithm<\/li>\n<\/ul>\n<h3>paper-deep-reader \u7cbe\u8bfb\u7ed3\u679c<\/h3>\n<h4>1. \u4e00\u53e5\u8bdd\u7ed3\u8bba<\/h4>\n<p>MedCRP-CL \u6700\u503c\u5f97\u5173\u6ce8\u7684\u5730\u65b9\u662f\u628a\u201c\u533b\u5b66\u5206\u5272\u6301\u7eed\u5b66\u4e60\u4e2d\u54ea\u4e9b\u4efb\u52a1\u8be5\u5171\u4eab\u3001\u54ea\u4e9b\u4efb\u52a1\u8be5\u9694\u79bb\u201d\u5f62\u5f0f\u5316\u4e3a prompt embedding \u4e0a\u7684 Bayesian nonparametric clustering\uff0c\u5e76\u7528 CRP \u81ea\u52a8\u5206\u914d LoRA \u4e13\u5bb6\uff0c\u4ece\u800c\u5728\u4e0d\u5b58\u50a8\u5386\u53f2\u60a3\u8005\u6570\u636e\u7684\u60c5\u51b5\u4e0b\u663e\u8457\u964d\u4f4e\u9057\u5fd8\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u8bba\u6587\u7814\u7a76\u7684\u662f continual medical image segmentation\uff1a\u533b\u5b66\u5206\u5272\u4efb\u52a1\u4f1a\u4ece\u4e0d\u540c\u533b\u9662\u3001\u4e0d\u540c\u8bbe\u5907\u3001\u4e0d\u540c\u89e3\u5256\u90e8\u4f4d\u548c\u4e0d\u540c\u75c5\u7406\u573a\u666f\u8fde\u7eed\u5230\u6765\uff0c\u6a21\u578b\u9700\u8981\u5b66\u4e60\u65b0\u4efb\u52a1\uff0c\u540c\u65f6\u4e0d\u80fd\u5fd8\u8bb0\u65e7\u4efb\u52a1\u3002\u4f20\u7edf\u4e00\u6b21\u6027\u8bad\u7ec3\u5047\u8bbe\u6240\u6709\u6570\u636e\u540c\u65f6\u53ef\u7528\uff0c\u800c\u771f\u5b9e\u4e34\u5e8a\u90e8\u7f72\u4e2d\u5e38\u89c1\u9650\u5236\u5305\u62ec\u6570\u636e\u9690\u79c1\u3001\u8de8\u4e2d\u5fc3\u6570\u636e\u4e0d\u53ef\u96c6\u4e2d\u3001\u4efb\u52a1\u987a\u5e8f\u5230\u8fbe\u3001\u75be\u75c5\/\u5668\u5b98\/\u6a21\u6001\u4e0d\u65ad\u6269\u5c55\u3002<\/p>\n<p>\u4f5c\u8005\u6307\u51fa\u6838\u5fc3\u77db\u76fe\u662f <strong>parameter sharing vs. parameter isolation<\/strong>\uff1a<br \/>\n- \u5982\u679c\u6240\u6709\u4efb\u52a1\u5171\u4eab\u540c\u4e00\u5957\u53c2\u6570\uff0c\u5dee\u5f02\u5f88\u5927\u7684\u4efb\u52a1\u4f1a\u4e92\u76f8\u5e72\u6270\uff0c\u5bfc\u81f4 catastrophic forgetting\u3002<br \/>\n- \u5982\u679c\u6bcf\u4e2a\u4efb\u52a1\u5b8c\u5168\u72ec\u7acb\u8bad\u7ec3\uff0c\u53c2\u6570\u7ebf\u6027\u589e\u957f\uff0c\u4e14\u76f8\u4f3c\u4efb\u52a1\u4e4b\u95f4\u4e0d\u80fd\u6b63\u8fc1\u79fb\u3002<br \/>\n- \u7269\u7406\u6a21\u6001\u6807\u7b7e\uff08\u5982 ultrasound\u3001X-ray\u3001endoscopy\uff09\u7c92\u5ea6\u592a\u7c97\uff1a\u5fc3\u810f\u8d85\u58f0\u4e0e\u4e73\u817a\u8d85\u58f0\u540c\u5c5e ultrasound\uff0c\u4f46\u89e3\u5256\u7ed3\u6784\u3001\u7eb9\u7406\u548c\u75c5\u7076\u6a21\u5f0f\u5b8c\u5168\u4e0d\u540c\u3002<\/p>\n<p>paper map\uff1a\u8bba\u6587\u7814\u7a76 sequential heterogeneous medical segmentation\uff1b\u4e3b\u52a8\u4f5c\u662f\u7528 clinical text prompts \u7684 frozen text embedding \u901a\u8fc7 Chinese Restaurant Process \u52a8\u6001\u53d1\u73b0 semantic modalities\uff0c\u518d\u4e3a\u6bcf\u4e2a semantic modality \u5206\u914d LoRA adapter\uff0c\u5e76\u53ea\u5728\u540c modality \u5185\u7528 EWC \u9632\u9057\u5fd8\uff1b\u4e3b\u5f20\u662f\u8be5\u7ed3\u6784\u80fd\u5728 16 \u4e2a\u8de8\u6a21\u6001\u4efb\u52a1\u4e0a\u63d0\u9ad8 Dice\u3001\u964d\u4f4e forgetting\u3001\u51cf\u5c11\u53c2\u6570\u5e76\u4fdd\u6301 replay-free\uff1b\u8bc1\u636e\u6765\u81ea\u4e0e Sequential\/EWC\/RAPF\/CL-LoRA\/MoE-Adapters\/Individual \u7684\u5bf9\u6bd4\u3001\u4efb\u52a1\u987a\u5e8f\u654f\u611f\u6027\u5206\u6790\u3001\u6a21\u5757\u6d88\u878d\u3001physical vs CRP grouping\u3001text vs visual clustering \u548c prompt robustness\uff1b\u4e3b\u8981\u5931\u8d25\u98ce\u9669\u662f\u4efb\u52a1\u96c6\u5408\u4ecd\u504f 2D\/CLIPSeg prompt-driven segmentation\uff0cprompt \u8d28\u91cf\u4e0e\u4efb\u52a1\u8fb9\u754c\u5047\u8bbe\u53ef\u80fd\u51b3\u5b9a\u6548\u679c\uff0c\u4e14\u7406\u8bba\u4fdd\u8bc1\u4f9d\u8d56\u8f83\u5f3a\u7684 Gaussian separation \u5047\u8bbe\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u6279\u8bc4\u51e0\u7c7b\u73b0\u6709\u65b9\u6cd5\uff1a<\/p>\n<ul>\n<li><strong>\u5168\u5c40 EWC \/ \u6b63\u5219\u5316\u65b9\u6cd5<\/strong>\uff1a\u5bf9\u6240\u6709\u4efb\u52a1\u65bd\u52a0\u7edf\u4e00\u53c2\u6570\u7ea6\u675f\u3002\u5982\u679c\u4efb\u52a1\u5dee\u5f02\u5927\uff0c\u6a21\u578b\u4f1a\u5728\u4e0d\u76f8\u5bb9\u76ee\u6807\u4e4b\u95f4\u6298\u4e2d\uff0c\u65e2\u5b66\u4e0d\u597d\u65b0\u4efb\u52a1\uff0c\u4e5f\u53ef\u80fd\u5fd8\u65e7\u4efb\u52a1\u3002<\/li>\n<li><strong>Replay-based continual learning<\/strong>\uff1a\u9700\u8981\u4fdd\u5b58\u65e7\u6837\u672c\u6216\u7279\u5f81\uff0c\u5728\u533b\u7597\u573a\u666f\u4e2d\u53ef\u80fd\u4e0e HIPAA\/GDPR \u6216\u533b\u9662\u6570\u636e\u7ba1\u7406\u51b2\u7a81\u3002<\/li>\n<li><strong>\u56fa\u5b9a MoE \u6216\u6bcf\u4efb\u52a1 LoRA<\/strong>\uff1a\u9700\u8981\u9884\u8bbe expert \u6570\u91cf\uff0c\u6216\u8005\u4e3a\u6bcf\u4e2a\u4efb\u52a1\u589e\u52a0\u4e00\u4e2a adapter\uff0c\u53c2\u6570\u589e\u957f\u5927\uff0c\u5e76\u4e14\u4e0d\u80fd\u81ea\u52a8\u53d1\u73b0\u54ea\u4e9b\u4efb\u52a1\u5e94\u5171\u4eab\u3002<\/li>\n<li><strong>\u6309\u7269\u7406\u6a21\u6001\u5206\u7ec4<\/strong>\uff1a\u628a\u6240\u6709 ultrasound \u5408\u5e76\u4f1a\u6df7\u6dc6\u5fc3\u810f\u8d85\u58f0\u548c\u4e73\u817a\u8d85\u58f0\uff1b\u628a\u6240\u6709 endoscopy \u5206\u5230\u4e00\u7c7b\u4e5f\u672a\u5fc5\u80fd\u5904\u7406\u4e0d\u540c polyp \u6570\u636e\u96c6\u7684 domain shift\u3002<\/li>\n<li><strong>\u89c6\u89c9\u7279\u5f81\u805a\u7c7b<\/strong>\uff1a\u9ad8\u7ef4\u3001\u53d7\u7ad9\u70b9\/\u8bbe\u5907\/\u6210\u50cf\u98ce\u683c\u5f71\u54cd\u5927\uff0c\u8bba\u6587\u5b9e\u9a8c\u4e2d visual-only clustering \u5bf9\u5fc3\u810f\/\u4e73\u817a ultrasound \u7684\u533a\u5206\u53cd\u800c\u5f31\u4e8e text prompt\u3002<\/li>\n<\/ul>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>MedCRP-CL \u7684\u6d41\u7a0b\u5982\u4e0b\uff1a<\/p>\n<ol>\n<li>\u6bcf\u4e2a\u65b0\u4efb\u52a1 <code>T_t<\/code> \u5305\u542b\u56fe\u50cf <code>x_i^t<\/code>\u3001mask <code>y_i^t<\/code> \u548c clinical text prompt <code>p_i^t<\/code>\u3002<\/li>\n<li>\u7528 frozen CLIP text encoder \u63d0\u53d6\u4efb\u52a1\u7ea7 prompt embedding\uff1a\u5bf9\u4efb\u52a1\u5185 unique prompts \u7684\u5f52\u4e00\u5316 embedding \u53d6\u5747\u503c\uff0c\u5f97\u5230 <code>e_t<\/code>\u3002<\/li>\n<li>\u5bf9\u5df2\u6709 semantic modality \u7ef4\u62a4 centroid <code>\u03bc_k<\/code>\uff0c\u8ba1\u7b97 similarity <code>s_{t,k}=&lt;e_t, \u03bc_k&gt;<\/code>\u3002<\/li>\n<li>\u4f7f\u7528 CRP prior \u51b3\u5b9a\u52a0\u5165\u5df2\u6709 cluster \u8fd8\u662f\u521b\u5efa\u65b0 cluster\uff1a\u5df2\u6709 cluster \u6982\u7387\u4e0e\u5176\u5df2\u6709\u4efb\u52a1\u6570 <code>n_k<\/code> \u6210\u6b63\u6bd4\uff0c\u65b0 cluster \u6982\u7387\u7531 concentration parameter <code>\u03b1<\/code> \u63a7\u5236\u3002<\/li>\n<li>\u4f7f\u7528 similarity likelihood \u4fee\u6b63 prior\uff1a\u76f8\u4f3c\u4efb\u52a1\u503e\u5411\u52a0\u5165\u5df2\u6709 modality\uff0c\u4e0d\u76f8\u4f3c\u4efb\u52a1\u521b\u5efa\u65b0 modality\u3002<\/li>\n<li>MAP inference \u5f97\u5230 <code>z_t<\/code>\uff0c\u5373\u5f53\u524d\u4efb\u52a1\u6240\u5c5e semantic modality\u3002<\/li>\n<li>\u6fc0\u6d3b\u8be5 modality \u5bf9\u5e94\u7684 LoRA adapter\uff0c\u53ea\u8bad\u7ec3\u8be5 adapter\uff1b\u5982\u679c\u662f\u65b0 modality\uff0c\u5219\u5206\u914d\u65b0 LoRA\u3002<\/li>\n<li>\u540c\u4e00 semantic modality \u5185\u4f7f\u7528 EWC \u4fdd\u5b58\u65e7\u4efb\u52a1\u91cd\u8981\u53c2\u6570\u65b9\u5411\uff1b\u4e0d\u540c modality \u4e4b\u95f4\u53c2\u6570\u9694\u79bb\uff0c\u4e0d\u53d1\u751f\u68af\u5ea6\u5e72\u6270\u3002<\/li>\n<li>\u4e0d\u5b58\u50a8\u5386\u53f2\u539f\u59cb\u56fe\u50cf\uff0c\u53ea\u5b58\u50a8 cluster centroid\u3001similarity statistics\u3001Fisher\/anchor parameters \u7b49\u805a\u5408\u7edf\u8ba1\u3002<\/li>\n<\/ol>\n<p>\u6574\u4f53 backbone \u662f frozen CLIPSeg\uff0cLoRA rank=8\uff0c\u03b1LoRA=16\uff0c\u5e94\u7528\u4e8e vision\/text encoder \u7684 Q\/K\/V\/O projection\uff1b\u56fe\u50cf resize \u5230 352\u00d7352\uff1b\u8bad\u7ec3\u4f7f\u7528 AdamW\u3001LR 1e-3\u3001weight decay 8e-5\uff0c\u5355\u5f20 RTX 4090\u3002<\/p>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<p><strong>\u6a21\u5757 1\uff1aPrompt-based semantic modality embedding<\/strong><br \/>\n- \u8f93\u5165\uff1a\u4efb\u52a1\u7684 clinical prompts\u3002<br \/>\n- \u8f93\u51fa\uff1a\u4efb\u52a1\u7ea7 embedding <code>e_t<\/code>\u3002<br \/>\n- \u89e3\u51b3\u95ee\u9898\uff1a\u7528\u6587\u672c\u6355\u83b7\u201c\u89e3\u5256\u90e8\u4f4d + \u75c5\u7406\u4e0a\u4e0b\u6587\u201d\uff0c\u6bd4\u7269\u7406\u6a21\u6001\u6807\u7b7e\u66f4\u7ec6\u3002<br \/>\n- \u521b\u65b0\u6027\u5224\u65ad\uff1a\u7528 text prompt \u505a continual segmentation routing \u662f\u672c\u6587\u6700\u5173\u952e\u7684\u673a\u5236\u4e4b\u4e00\u3002\u5b83\u9002\u5408 vision-language segmentation \u6846\u67b6\uff0c\u4f46\u4f9d\u8d56 prompt \u8d28\u91cf\u3002<br \/>\n- \u5bf9 polyp segmentation\uff1a\u6709\u4ef7\u503c\u3002\u5982\u679c\u591a\u4e2a\u606f\u8089\u6570\u636e\u96c6 prompt \u76f8\u8fd1\uff0c\u53ef\u80fd\u88ab\u805a\u4e3a\u540c\u4e00 endoscopy\/polyp semantic modality\uff0c\u5b9e\u73b0\u8de8\u6570\u636e\u96c6\u5171\u4eab\u3002<\/p>\n<p><strong>\u6a21\u5757 2\uff1aCRP Bayesian nonparametric modality discovery<\/strong><br \/>\n- \u8f93\u5165\uff1a\u5f53\u524d\u4efb\u52a1 embedding\u3001\u5df2\u6709 cluster centroid\u3001cluster size\u3001similarity likelihood\u3002<br \/>\n- \u8f93\u51fa\uff1a\u52a0\u5165\u5df2\u6709 semantic modality \u6216\u65b0\u5efa modality\u3002<br \/>\n- \u5173\u952e\u516c\u5f0f\uff1a\u5df2\u6709 cluster \u7684 log posterior \u7ea6\u4e3a <code>log n_k - log(t-1+\u03b1) + \u2113(s_{t,k})<\/code>\uff1b\u65b0 cluster \u7ea6\u4e3a <code>log \u03b1 - log(t-1+\u03b1) - \u2113(max_k s_{t,k})<\/code>\u3002<br \/>\n- \u89e3\u51b3\u95ee\u9898\uff1a\u4e0d\u9700\u8981\u9884\u8bbe\u4efb\u52a1\u7c07\u6570\u91cf K\uff0c\u4efb\u52a1\u6d41\u4e2d\u65b0\u9886\u57df\u5230\u6765\u65f6\u53ef\u6269\u5c55\u3002<br \/>\n- \u521b\u65b0\u6027\u5224\u65ad\uff1aCRP \u672c\u8eab\u4e0d\u662f\u65b0\u7406\u8bba\uff0c\u4f46\u653e\u5230\u533b\u5b66\u5206\u5272\u6301\u7eed\u5b66\u4e60\u7684 task routing \u4e0a\u5f88\u5408\u9002\uff0c\u4e14\u6bd4\u56fa\u5b9a K \u7684 MoE \u66f4\u81ea\u7136\u3002<\/p>\n<p><strong>\u6a21\u5757 3\uff1aAdaptive similarity likelihood<\/strong><br \/>\n- \u8f93\u5165\uff1asimilarity score <code>s<\/code>\u3002<br \/>\n- \u5047\u8bbe\uff1asame-modality similarity \u4e0e different-modality similarity \u5206\u522b\u670d\u4ece Gaussian\uff0c\u5728\u7ebf\u4f30\u8ba1 <code>\u03bc_intra, \u03c3_intra, \u03bc_inter, \u03c3_inter<\/code>\u3002<br \/>\n- \u8f93\u51fa\uff1alog-likelihood ratio <code>\u2113(s)<\/code>\u3002<br \/>\n- \u89e3\u51b3\u95ee\u9898\uff1a\u907f\u514d\u624b\u5de5 similarity threshold\u3002<br \/>\n- \u98ce\u9669\uff1aGaussian \u5047\u8bbe\u548c separation \u5047\u8bbe\u5728\u66f4\u590d\u6742\u771f\u5b9e\u533b\u9662\u4efb\u52a1\u6d41\u4e2d\u672a\u5fc5\u6210\u7acb\uff1b\u5982\u679c prompt \u9000\u5316\u6210 generic prompt\uff0c\u5219\u8bba\u6587\u5b9e\u9a8c\u663e\u793a K \u4f1a\u9000\u5316\u4e3a 1\u3002<\/p>\n<p><strong>\u6a21\u5757 4\uff1aModality-specific LoRA<\/strong><br \/>\n- \u8f93\u5165\uff1aCLIPSeg frozen backbone \u4e2d\u7684 linear layer\u3002<br \/>\n- \u64cd\u4f5c\uff1a\u6bcf\u4e2a semantic modality \u6709\u81ea\u5df1\u7684\u4f4e\u79e9\u66f4\u65b0 <code>W_k = W_0 + (\u03b1LoRA\/r) B_k A_k<\/code>\u3002<br \/>\n- \u8f93\u51fa\uff1a\u6309\u4efb\u52a1 cluster \u6fc0\u6d3b\u5bf9\u5e94 adapter\u3002<br \/>\n- \u89e3\u51b3\u95ee\u9898\uff1a\u8de8 modality \u53c2\u6570\u9694\u79bb\uff0c\u540c modality \u53c2\u6570\u5171\u4eab\u3002<br \/>\n- \u8fc1\u79fb\u4ef7\u503c\uff1a\u9ad8\u3002\u53ef\u4ee5\u501f\u9274\u5230 MedSAM\u3001SAM-Med2D\u3001\u533b\u5b66 VLM segmentation \u6216 promptable 3D segmentation\u3002<\/p>\n<p><strong>\u6a21\u5757 5\uff1aIntra-modality EWC<\/strong><br \/>\n- \u8f93\u5165\uff1a\u540c\u4e00 modality \u5185\u65e7\u4efb\u52a1 Fisher information \u4e0e anchor parameters\u3002<br \/>\n- \u8f93\u51fa\uff1aregularization penalty\uff0c\u9650\u5236\u91cd\u8981\u53c2\u6570\u6f02\u79fb\u3002<br \/>\n- \u89e3\u51b3\u95ee\u9898\uff1a\u5373\u4f7f\u540c\u4e00 modality \u5185\u5171\u4eab adapter\uff0c\u4e5f\u8981\u9632\u6b62\u65b0\u4efb\u52a1\u8986\u76d6\u65e7\u4efb\u52a1\u3002<br \/>\n- \u521b\u65b0\u6027\u5224\u65ad\uff1aEWC \u662f\u7ecf\u5178\u65b9\u6cd5\uff1b\u672c\u6587\u4ef7\u503c\u5728\u4e8e\u53ea\u5728 semantic modality \u5185\u65bd\u52a0\uff0c\u800c\u4e0d\u662f\u5168\u5c40\u65bd\u52a0\u3002<\/p>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p>\u5b9e\u9a8c\u4efb\u52a1\uff1a16 \u4e2a\u533b\u5b66\u5206\u5272\u4efb\u52a1\uff0c\u8986\u76d6\u56db\u7c7b\u6210\u50cf\uff1a<br \/>\n- Endoscopy \/ Colon\uff1aKvasir\u3001ClinicDB\u3001ETIS\u3001CVC-300\u3001ColonDB\u3002<br \/>\n- Dermoscopy \/ Skin\uff1aISIC\u3002<br \/>\n- Ultrasound\uff1aCAMUS \u5fc3\u810f\u3001BUSI benign\/malignant \u4e73\u817a\u3002<br \/>\n- X-ray \/ Chest\uff1aCheXlocalize \u6d3e\u751f\u7684 7 \u4e2a pathology localization\/segmentation subsets\u3002<\/p>\n<p>\u6307\u6807\uff1a<br \/>\n- Average Dice\uff1a\u6240\u6709\u4efb\u52a1\u6700\u7ec8\u5206\u5272\u6027\u80fd\u5e73\u5747\u3002<br \/>\n- Forgetting Rate\uff1a\u6bcf\u4e2a\u65e7\u4efb\u52a1 peak validation performance \u4e0e\u6700\u7ec8 performance \u7684\u5dee\u503c\u5e73\u5747\u3002<br \/>\n- \u8fd8\u62a5\u544a trainable parameters\u3001GPU memory\u3001relative time\u3002<\/p>\n<p>\u4e3b\u8981\u7ed3\u679c\uff08Table 2\uff09\uff1a<br \/>\n- Individual upper bound\uff1aDice 77.9%\u3002<br \/>\n- Sequential\uff1aDice 48.0\u00b17.1\uff0cForgetting 28.3\u00b17.7\u3002<br \/>\n- EWC\uff1aDice 56.8\u00b13.7\uff0cForgetting 11.3\u00b13.5\u3002<br \/>\n- RAPF\uff1aDice 58.4\u00b11.7\uff0cForgetting 7.2\u00b12.6\u3002<br \/>\n- CL-LoRA\uff1aDice 60.7\u00b12.0\uff0cForgetting 9.7\u00b11.4\uff0c\u53c2\u6570 0.05M\u3002<br \/>\n- MoE-Adapters\uff1aDice 65.3\u00b13.4\uff0cForgetting 7.1\u00b13.2\uff0c\u53c2\u6570 51.9M\u3002<br \/>\n- MedCRP-CL\uff1aDice 73.3\u00b11.0\uff0cForgetting 4.1\u00b10.8\uff0c\u53c2\u6570 8.6M\u3002<\/p>\n<p>\u6d88\u878d\u7ed3\u679c\uff1a<br \/>\n- Full model\uff1aDice 73.33\uff0cForgetting 4.09\u3002<br \/>\n- w\/o EWC\uff1aDice 71.92\uff0cForgetting 5.41\uff0c\u8bf4\u660e intra-modality consolidation \u6709\u5e2e\u52a9\u4f46\u4e0d\u662f\u6700\u5927\u8d21\u732e\u3002<br \/>\n- w\/o CRP\uff1aDice 57.59\uff0cForgetting 15.55\uff0c\u8bf4\u660e task grouping\/routing \u662f\u5173\u952e\u3002<br \/>\n- Single LoRA\uff1aDice 46.94\uff0cForgetting 27.34\uff0c\u8bf4\u660e\u6240\u6709\u4efb\u52a1\u5171\u4eab\u5355 adapter \u4f1a\u4e25\u91cd\u9057\u5fd8\u3002<br \/>\n- w\/o LoRA\uff1aDice 45.39\uff0cForgetting 0.03\uff0c\u8bf4\u660e\u5b8c\u5168\u51bb\u7ed3 backbone \u867d\u4e0d\u9057\u5fd8\u4f46\u9002\u5e94\u80fd\u529b\u5dee\u3002<\/p>\n<p>semantic modality \u5206\u6790\uff1a<br \/>\n- CRP \u81ea\u52a8\u53d1\u73b0 K=5\uff1b\u76f8\u6bd4 physical imaging type grouping\uff08K=4\uff09\uff0cDice \u4ece 65.75 \u63d0\u9ad8\u5230 73.33\uff0cForgetting \u4ece 9.23 \u964d\u5230 4.09\u3002<br \/>\n- \u5173\u952e\u53d1\u73b0\uff1aCRP \u628a cardiac ultrasound \u4e0e breast ultrasound \u5206\u5f00\uff0c\u56e0\u4e3a\u6587\u672c\u8bed\u4e49\u4e0d\u540c\uff1b\u7269\u7406\u6a21\u6001\u5206\u7ec4\u4f1a\u628a\u5b83\u4eec\u9519\u8bef\u5408\u5e76\u3002<br \/>\n- Text-only clustering \u7684 intra\/inter gap \u7ea6 0.50\uff1bvisual-only gap \u7ea6 0.22\uff0c\u5e76\u4e14 visual-only K \u4e0d\u7a33\u5b9a\u3002<br \/>\n- 10 \u4e2a contrastive text encoders \u5728 \u03b1=5 \u4e0b\u5747\u53d1\u73b0 K=5\uff1bSigLIP \u548c S-PubMedBERT \u8fd9\u7c7b\u975e contrastive encoder \u9000\u5316\u4e3a K=1\u3002<br \/>\n- prompt robustness\uff1a\u4e34\u5e8a\u7f29\u5199\u300110\u201320% typo\u300120\u201330% keyword drop\u3001word shuffle \u4ecd\u4fdd\u6301 K=5\uff1b30% typo\u300150% keyword drop \u6216 generic prompt \u4f1a\u9000\u5316\u3002<\/p>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u53ef\u4fe1\u70b9\uff1a<br \/>\n- \u8bba\u6587\u4e0d\u662f\u53ea\u5728\u5355\u4e00\u6570\u636e\u96c6\u4e0a\u505a\u5c0f\u6539\u52a8\uff0c\u800c\u662f\u8de8 16 \u4e2a\u4efb\u52a1\u30014 \u7c7b\u6210\u50cf\u3001\u591a\u4e2a\u5668\u5b98\/\u75c5\u7076\u505a continual learning\u3002<br \/>\n- \u5bf9\u6bd4\u4e86 sequential\u3001EWC\u3001RAPF\u3001CL-LoRA\u3001MoE-Adapters\u3001individual upper bound\uff0c\u57fa\u7ebf\u8986\u76d6\u8f83\u5408\u7406\u3002<br \/>\n- \u6d88\u878d\u76f4\u63a5\u652f\u6301\u6838\u5fc3\u4e3b\u5f20\uff1aCRP routing \u662f\u6700\u5927\u8d21\u732e\uff0cLoRA \u63d0\u4f9b plasticity\uff0cEWC \u63d0\u4f9b stability\u3002<br \/>\n- \u4efb\u52a1\u987a\u5e8f\u654f\u611f\u6027\u3001text vs visual clustering\u3001encoder sensitivity\u3001prompt robustness \u90fd\u662f\u5bf9\u4e3b\u5f20\u975e\u5e38\u5173\u952e\u7684\u68c0\u67e5\u3002<br \/>\n- \u4ee3\u7801\u4ed3\u5e93\u5df2\u53ef\u8bbf\u95ee\uff0c\u4fbf\u4e8e\u540e\u7eed\u590d\u73b0\u3002<\/p>\n<p>\u9700\u8981\u8c28\u614e\u7684\u70b9\uff1a<br \/>\n- backbone \u662f CLIPSeg\uff0c\u4efb\u52a1\u591a\u4e3a 2D segmentation \/ prompt-guided segmentation\uff1b\u5b83\u4e0d\u7b49\u4ef7\u4e8e nnU-Net \u5f0f 3D medical segmentation\uff0c\u4e5f\u4e0d\u76f4\u63a5\u8986\u76d6 CT\/MRI \u591a\u5668\u5b98 3D \u4f53\u6570\u636e\u3002<br \/>\n- \u90e8\u5206 chest X-ray \u4efb\u52a1\u6765\u81ea localization\/bounding-box \u98ce\u683c\u6570\u636e\uff0c\u548c\u4e25\u683c voxel\/pixel-level medical segmentation \u7684\u6807\u6ce8\u8d28\u91cf\u53ef\u80fd\u4e0d\u540c\u3002<br \/>\n- prompt \u662f\u65b9\u6cd5\u6210\u529f\u7684\u5173\u952e\uff1b\u5982\u679c\u771f\u5b9e\u4e34\u5e8a\u4efb\u52a1\u6ca1\u6709\u9ad8\u8d28\u91cf target prompts\uff0c\u6216 prompt \u4e0e\u56fe\u50cf\u4efb\u52a1\u4e0d\u4e00\u81f4\uff0crouting \u4f1a\u9000\u5316\u3002<br \/>\n- CRP \u7406\u8bba\u4fdd\u8bc1\u4f9d\u8d56 same\/different similarity \u7684 Gaussian separation\uff0c\u771f\u5b9e\u591a\u4e2d\u5fc3\u4efb\u52a1\u6d41\u4e2d\u53ef\u80fd\u6709\u8fde\u7eed\u8c31\u800c\u975e\u6e05\u6670\u7c07\u3002<br \/>\n- \u53c2\u6570 8.6M \u5c0f\u4e8e MoE-Adapters\uff0c\u4f46\u5927\u4e8e CL-LoRA\uff1b\u5982\u679c\u4efb\u52a1\u6570\u7ee7\u7eed\u589e\u957f\uff0cadapter \u6570\u4e5f\u4f1a\u589e\u957f\uff0c\u53ea\u662f\u6309 semantic modality \u800c\u975e\u6309 task \u7ebf\u6027\u589e\u957f\u3002<br \/>\n- \u6ca1\u6709\u8bc1\u660e\u5bf9 3D volumes\u3001nnU-Net\/MedNeXt backbone \u6216 SAM\/MedSAM \u7c7b\u5927\u578b foundation segmenter \u540c\u6837\u6709\u6548\u3002<\/p>\n<p>\u7ed3\u8bba\u5f3a\u5ea6\uff1a\u53ef\u4ee5\u8f83\u5f3a\u5730\u76f8\u4fe1\u201cprompt-based semantic modality discovery \u5bf9 heterogeneous continual medical segmentation \u6709\u4ef7\u503c\u201d\uff1b\u4f46\u5e94\u628a\u7ed3\u8bba\u9650\u5236\u5728 CLIPSeg\/VLSM \u4e0e 2D \u591a\u4efb\u52a1\u8bbe\u7f6e\uff0c\u4e0d\u80fd\u76f4\u63a5\u63a8\u65ad\u5230\u6240\u6709 3D \u533b\u5b66\u5206\u5272\u90e8\u7f72\u3002<\/p>\n<h4>8. \u4e0e\u4e3b\u6d41\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u7684\u5173\u7cfb<\/h4>\n<ul>\n<li>\u4e0e <strong>U-Net\/nnU-Net<\/strong>\uff1aMedCRP-CL \u4e0d\u662f nnU-Net \u66ff\u4ee3\u54c1\uff0c\u800c\u662f\u6301\u7eed\u5b66\u4e60\/\u4efb\u52a1\u8def\u7531\u6846\u67b6\u3002\u82e5\u8981\u8fc1\u79fb\u5230 nnU-Net\uff0c\u9700\u8981\u628a LoRA\/adapter \u6982\u5ff5\u6362\u6210 decoder\/encoder adapter \u6216 task-specific normalization\u3002<\/li>\n<li>\u4e0e <strong>MedNeXt \/ CNN-based segmentation<\/strong>\uff1a\u53ef\u4ee5\u501f\u9274 semantic modality routing\uff0c\u628a\u4e0d\u540c\u5668\u5b98\/\u6a21\u6001\u4efb\u52a1\u5206\u914d\u5230\u4e0d\u540c adapter \u6216 normalization branch\u3002<\/li>\n<li>\u4e0e <strong>UNETR\/Swin-UNETR\/TransUNet<\/strong>\uff1aTransformer linear projections \u5929\u7136\u9002\u5408 LoRA\uff0c\u56e0\u6b64\u6bd4\u7eaf CNN \u66f4\u5bb9\u6613\u79fb\u690d MedCRP-CL\u3002<\/li>\n<li>\u4e0e <strong>Mamba\/VMamba\/SegMamba\/DAMamba<\/strong>\uff1aMamba block \u4e5f\u6709\u53ef\u63d2\u5165\u4f4e\u79e9\u9002\u914d\u6216 selective scan \u53c2\u6570\u8c03\u5236\u7684\u7a7a\u95f4\u3002MedCRP-CL \u5bf9 DAMamba \u7684\u542f\u53d1\u662f\uff1a\u591a\u6570\u636e\u96c6\/\u591a\u5668\u5b98\u8bad\u7ec3\u65f6\u4e0d\u8981\u53ea\u505a\u7edf\u4e00\u5171\u4eab\uff0c\u53ef\u7528 prompt\/metadata \u81ea\u52a8\u51b3\u5b9a\u5171\u4eab\u6216\u9694\u79bb\u3002<\/li>\n<li>\u4e0e <strong>foundation model for medical segmentation<\/strong>\uff1a\u5173\u7cfb\u5bc6\u5207\u3002\u5b83\u5efa\u7acb\u5728 CLIPSeg\/VLSM \u4e0a\uff0c\u662f\u4e00\u79cd medical segmentation foundation\/adaptation \u7684 continual learning \u65b9\u6848\uff1b\u672a\u6765\u53ef\u63a5\u5230 MedSAM\u3001SAM-Med3D\u3001universal medical segmentation \u6a21\u578b\u4e0a\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<ul>\n<li><strong>\u5bf9 polyp segmentation<\/strong>\uff1a\u4ef7\u503c\u8f83\u9ad8\u3002\u8bba\u6587\u5305\u542b Kvasir\u3001ClinicDB\u3001ETIS\u3001CVC-300\u3001ColonDB \u7b49\u606f\u8089\u6570\u636e\u96c6\uff0c\u53ef\u7528\u4e8e\u601d\u8003\u8de8\u606f\u8089\u6570\u636e\u96c6\u6301\u7eed\u5b66\u4e60\u3001domain shift \u548c\u6570\u636e\u96c6\u987a\u5e8f\u8bad\u7ec3\u3002<\/li>\n<li><strong>\u5bf9 DAMamba \u6539\u9020<\/strong>\uff1a\u4ef7\u503c\u4e2d\u5230\u9ad8\u3002\u82e5\u4f60\u7684 DAMamba \u9762\u5411\u591a\u6570\u636e\u96c6\u6216 universal medical segmentation\uff0c\u53ef\u4ee5\u501f\u9274 semantic modality routing\uff0c\u4e3a\u4e0d\u540c\u6570\u636e\u96c6\/\u5668\u5b98\u6fc0\u6d3b\u4e0d\u540c lightweight adapter\uff0c\u800c\u4e0d\u662f\u5168\u5171\u4eab\u3002<\/li>\n<li><strong>\u5bf9\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u9009\u62e9<\/strong>\uff1a\u5b83\u63d0\u9192\u6211\u4eec\u201c\u591a\u4efb\u52a1\u8054\u5408\u8bad\u7ec3\u201d\u4e0d\u662f\u7b80\u5355\u62fc\u6570\u636e\uff1b\u4efb\u52a1\u7ed3\u6784\u53d1\u73b0\u672c\u8eab\u53ef\u80fd\u662f\u8d21\u732e\u70b9\u3002<\/li>\n<li><strong>\u5bf9 introduction\/related work<\/strong>\uff1a\u53ef\u653e\u5728 continual medical image segmentation\u3001vision-language medical segmentation\u3001parameter-efficient adaptation\u3001privacy-preserving\/replay-free learning \u6bb5\u843d\u3002<\/li>\n<li><strong>\u5bf9\u590d\u73b0\u5b9e\u9a8c<\/strong>\uff1a\u4ee3\u7801\u5df2\u516c\u5f00\uff0c\u503c\u5f97 clone \u8dd1\u4e00\u904d\uff0c\u5c24\u5176\u662f polyp task sequence\u3002\u5982\u679c\u505a DAMamba\uff0c\u53ef\u4ee5\u5148\u590d\u73b0\u5176 Kvasir\/ClinicDB\/ETIS\/ColonDB continual order\uff0c\u518d\u66ff\u6362 backbone\u3002<\/li>\n<\/ul>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5f3a\u70c8\u5efa\u8bae\u7cbe\u8bfb\u3002<\/strong>\u76f8\u6bd4\u8bb8\u591a\u53ea\u6539 U-Net \u6a21\u5757\u7684 preprint\uff0c\u8fd9\u7bc7\u7684\u7814\u7a76\u95ee\u9898\u66f4\u63a5\u8fd1\u771f\u5b9e\u90e8\u7f72\uff1a\u8de8\u4efb\u52a1\u3001\u8de8\u6a21\u6001\u3001\u987a\u5e8f\u5230\u8fbe\u3001\u9690\u79c1\u9650\u5236\u3001\u53c2\u6570\u6548\u7387\u3002\u5efa\u8bae\u91cd\u70b9\u8bfb Method 3.1\/3.2\u3001Table 2\/4\/6\/7\/8\/9 \u548c Appendix A \u7684\u5047\u8bbe\uff1b\u5982\u679c\u4f60\u7684\u7814\u7a76\u6682\u65f6\u53ea\u505a\u5355\u6570\u636e\u96c6\u606f\u8089\u5206\u5272\uff0c\u53ef\u4ee5\u5148\u7565\u8bfb\u7406\u8bba\uff0c\u91cd\u70b9\u501f\u9274 task routing \u4e0e polyp \u591a\u6570\u636e\u96c6\u8bbe\u7f6e\u3002<\/p>\n<hr \/>\n<h2>\u4eca\u65e5\u63a8\u8350\u4f18\u5148\u7ea7<\/h2>\n<ol>\n<li><strong>MedCRP-CL<\/strong>\uff1a\u6700\u503c\u5f97\u4f18\u5148\u6df1\u5165\u8bfb\u3002\u5b83\u662f ICML 2026 accepted\uff0c\u95ee\u9898\u8bbe\u5b9a\u6bd4\u666e\u901a\u7ed3\u6784\u5806\u53e0\u66f4\u6709\u7814\u7a76\u4ef7\u503c\uff0c\u5e76\u4e14\u5305\u542b polyp segmentation \u591a\u6570\u636e\u96c6\uff0c\u548c\u540e\u7eed\u505a universal\/continual medical segmentation\u3001DAMamba \u591a\u4efb\u52a1\u6269\u5c55\u3001related work \u5199\u4f5c\u90fd\u66f4\u76f8\u5173\u3002<\/li>\n<li><strong>SegGuidedNet<\/strong>\uff1a\u5efa\u8bae\u4f5c\u4e3a\u8f7b\u91cf\u6a21\u5757\u7075\u611f\u9605\u8bfb\u3002\u5b83\u5bf9 3D U-Net\/brain tumor segmentation \u5f88\u5b9e\u7528\uff0c\u4f46\u76ee\u524d\u7f3a\u5c11\u5173\u952e ablation \u548c\u4ee3\u7801\uff0c\u9002\u5408\u501f\u9274\u5176 auxiliary attention supervision\uff0c\u800c\u4e0d\u662f\u76f4\u63a5\u5f53\u4f5c\u5f3a SOTA \u8bc1\u636e\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\uff0c\u8def\u5f84\uff1aMEDIA:\/tmp\/medseg_daily_2026-05-24\/segguidednet_2605.22572.pdf\uff1bPDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.22572<\/li>\n<li>\u8bba\u6587 2\uff1a\u5df2\u9644 PDF\uff0c\u8def\u5f84\uff1aMEDIA:\/tmp\/medseg_daily_2026-05-24\/medcrp_cl_2605.20297.pdf\uff1bPDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.20297<\/li>\n<\/ul>\n<h2>\u4eca\u65e5\u53ef\u6267\u884c\u5efa\u8bae<\/h2>\n<ol>\n<li>\u5982\u679c\u4eca\u5929\u53ea\u80fd\u8bfb\u4e00\u7bc7\uff0c\u5148\u8bfb <strong>MedCRP-CL<\/strong>\uff1a\u91cd\u70b9\u770b CRP task routing\u3001LoRA adapter \u5206\u914d\u3001Table 4 \u6d88\u878d\u548c polyp \u6570\u636e\u96c6\u4efb\u52a1\u5e8f\u5217\uff0c\u5224\u65ad\u80fd\u5426\u8fc1\u79fb\u5230 DAMamba \u7684\u591a\u6570\u636e\u96c6\u8bad\u7ec3\u3002<\/li>\n<li>\u5bf9 DAMamba\/U-Net \u7c7b\u6846\u67b6\uff0c\u53ef\u5c1d\u8bd5\u4ece <strong>SegGuidedNet<\/strong> \u501f\u9274\u4e00\u4e2a\u4f4e\u6210\u672c auxiliary head\uff1a\u628a binary polyp mask \u6d3e\u751f\u4e3a interior\/boundary\/uncertain rim \u4e09\u7c7b\u76d1\u7763\uff0c\u6d4b\u8bd5\u662f\u5426\u6539\u5584\u8fb9\u754c Dice\u3001HD95 \u6216 mIoU\u3002<\/li>\n<li>\u5199 related work \u65f6\uff0c\u628a MedCRP-CL \u653e\u5728 continual\/foundation medical segmentation\uff0c\u628a SegGuidedNet \u653e\u5728 interpretable auxiliary-supervised 3D segmentation\uff1b\u4e24\u8005\u90fd\u4e0d\u8981\u5199\u6210\u201c\u901a\u7528\u533b\u5b66\u56fe\u50cf\u5206\u5272 SOTA backbone\u201d\u3002<\/li>\n<\/ol>\n<h2>\u53c2\u8003\u94fe\u63a5<\/h2>\n<ul>\n<li>SegGuidedNet arXiv\uff1ahttps:\/\/arxiv.org\/abs\/2605.22572<\/li>\n<li>SegGuidedNet PDF\uff1ahttps:\/\/arxiv.org\/pdf\/2605.22572<\/li>\n<li>MedCRP-CL arXiv\uff1ahttps:\/\/arxiv.org\/abs\/2605.20297<\/li>\n<li>MedCRP-CL PDF\uff1ahttps:\/\/arxiv.org\/pdf\/2605.20297<\/li>\n<li>MedCRP-CL code\uff1ahttps:\/\/github.com\/zygao930\/MedCRP-CL<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u4eca\u65e5\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6700\u65b0\u8bba\u6587\u7cbe\u8bfb\u8ffd\u8e2a \u4eca\u65e5\u7ed3\u8bba \u4eca\u5929\u5728 2026-05-21 \u81f3 2026-05-19 \u7684\u6700\u65b0 arXiv\/\u4f1a\u8bae\u9884\u5370\u672c &#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-1076","post","type-post","status-publish","format-standard","hentry","category-85"],"views":32,"_links":{"self":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1076","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=1076"}],"version-history":[{"count":0,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1076\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/media?parent=1076"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/categories?post=1076"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/tags?post=1076"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}