{"id":1070,"date":"2026-05-21T08:39:02","date_gmt":"2026-05-21T00:39:02","guid":{"rendered":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/21\/2026-05-21-%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%9a%e9%a2%91%e5%9f%9f%e5%8e%9f%e5%9e%8b%e5%8c%b9%e9%85%8d%e4%b8%8e%e5%a4%9a%e4%b8%93%e5%ae%b6\/"},"modified":"2026-05-21T08:39:02","modified_gmt":"2026-05-21T00:39:02","slug":"2026-05-21-%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%9a%e9%a2%91%e5%9f%9f%e5%8e%9f%e5%9e%8b%e5%8c%b9%e9%85%8d%e4%b8%8e%e5%a4%9a%e4%b8%93%e5%ae%b6","status":"publish","type":"post","link":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/21\/2026-05-21-%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%9a%e9%a2%91%e5%9f%9f%e5%8e%9f%e5%9e%8b%e5%8c%b9%e9%85%8d%e4%b8%8e%e5%a4%9a%e4%b8%93%e5%ae%b6\/","title":{"rendered":"2026-05-21 \u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\u7cbe\u8bfb\uff1a\u9891\u57df\u539f\u578b\u5339\u914d\u4e0e\u591a\u4e13\u5bb6\u4e0d\u786e\u5b9a\u6027\u5206\u5272"},"content":{"rendered":"<h1>\u4eca\u65e5\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6700\u65b0\u8bba\u6587\u7cbe\u8bfb\u8ffd\u8e2a<\/h1>\n<h2>\u4eca\u65e5\u7ed3\u8bba<\/h2>\n<p>\u4eca\u5929\u68c0\u7d22\u5230\u7684\u6700\u65b0\u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\u4ecd\u4ee5 <strong>2026 arXiv preprint \/ CVPR 2026 \u65b9\u5411<\/strong>\u4e3a\u4e3b\uff0c\u771f\u6b63\u6765\u81ea\u671f\u520a\u6b63\u5f0f\u5377\u671f\u6216\u9876\u4f1a\u5b98\u7f51\u7684\u65b0\u8bba\u6587\u8f83\u5c11\uff1b\u7b5b\u9009\u540e\u6700\u503c\u5f97\u5173\u6ce8\u7684\u4e24\u7bc7\u5206\u522b\u4ee3\u8868\u4e24\u4e2a\u4e0d\u540c\u8d8b\u52bf\uff1a<strong>few-shot medical segmentation \u7684\u9891\u57df\/\u6d41\u5f62\u5339\u914d\u6539\u9020<\/strong>\uff0c\u4ee5\u53ca <strong>multi-rater \/ uncertainty-aware segmentation \u7684\u9891\u57df\u4e2a\u6027\u5316\u4e0e\u566a\u58f0\u89e3\u8026<\/strong>\u3002\u6574\u4f53\u4e0a\uff0c\u533b\u5b66\u5206\u5272\u8fd1\u671f\u7684\u521b\u65b0\u4e0d\u518d\u53ea\u662f\u5806\u53e0 U-Net\/Transformer\/Mamba block\uff0c\u800c\u66f4\u5f3a\u8c03\uff1a\u7279\u5f81\u9891\u8c31\u3001\u8fb9\u754c\u4e0d\u786e\u5b9a\u6027\u3001\u4e13\u5bb6\u6807\u6ce8\u5dee\u5f02\u3001\u4ee5\u53ca\u4f4e\u6807\u6ce8\/\u8de8\u57df\u6761\u4ef6\u4e0b\u7684\u53ef\u4fe1\u6cdb\u5316\u3002<\/p>\n<h2>\u68c0\u7d22\u8bf4\u660e<\/h2>\n<p>\u672c\u6b21\u4f18\u5148\u68c0\u7d22 arXiv\u3001Semantic Scholar\/\u641c\u7d22\u5f15\u64ce\u53ef\u89c1\u8bb0\u5f55\u4ee5\u53ca\u5386\u53f2 cron \u8f93\u51fa\u4e2d\u7684\u533b\u5b66\u56fe\u50cf\u5206\u5272\u76f8\u5173\u8bba\u6587\uff0c\u91cd\u70b9\u5173\u6ce8 2025 \u5e74\u53ca\u4ee5\u540e\u7684 medical image segmentation\u3001few-shot segmentation\u3001multi-rater segmentation\u3001frequency-domain segmentation\u3001polyp\/noisy-label\/3D\/universal segmentation \u7b49\u65b9\u5411\u3002\u4eca\u5929\u6ca1\u6709\u53d1\u73b0\u66f4\u9ad8\u8d28\u91cf\u4e14\u672a\u91cd\u590d\u7684\u6b63\u5f0f\u9876\u520a\/\u9876\u4f1a\u5b98\u7f51\u65b0\u8bba\u6587\uff0c\u56e0\u6b64\u4ece 2026 \u5e74 5 \u6708 arXiv \/ CVPR 2026 accepted preprint \u4e2d\u7b5b\u9009\uff1b\u4e24\u7bc7\u5165\u9009\u8bba\u6587\u5747\u4e3a 2025 \u5e74\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\uff1aPatch-MoE Mamba\u3001DepthPolyp\u3001Semi-MedRef\u3001CMFDNet\u3001Topo-VM-UNetV2\u3001FEFormer\u3001USEMA\u3001MedCore\u3001TopoMamba\u3001ESICA \u7b49\u3002<\/p>\n<h2>WordPress \u53d1\u5e03<\/h2>\n<ul>\n<li>WordPress \u6587\u7ae0\u94fe\u63a5\uff1a\u53d1\u5e03\u4e2d<\/li>\n<li>WordPress Post ID\uff1a\u53d1\u5e03\u4e2d<\/li>\n<\/ul>\n<hr \/>\n<h2>\u8bba\u6587 1\uff1aBeyond Euclidean Prototypes: Spectral Disentanglement and Geodesic Matching for Few-Shot Medical Image Segmentation<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aBeyond Euclidean Prototypes: Spectral Disentanglement and Geodesic Matching for Few-Shot Medical Image Segmentation<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aPenghao Jia, Zhiyong Huang, Mingyang Hou, Zhi Yu, Shuai Miao, Jiahong Wang, Yan Yan \/ Penghao Jia<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-18 submitted<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint\uff0carXiv:2605.17904<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.17904<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.17904<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/naivejph\/SGP-Net.git<\/li>\n<li>\u4efb\u52a1\uff1aFew-Shot Medical Image Segmentation\uff1b1-way 1-shot organ segmentation<\/li>\n<li>\u6570\u636e\u96c6\uff1aAbd-MRI \/ CHAOS-T2\u3001Abd-CT \/ SABS\u3001CMR \/ MICCAI 2019 Multi-sequence Cardiac MRI<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1aprototype-based few-shot segmentation\uff1bfrequency-domain prototype disentanglement\uff1bgeodesic \/ heat-diffusion matching<\/li>\n<\/ul>\n<h3>paper-deep-reader \u7cbe\u8bfb\u7ed3\u679c<\/h3>\n<h4>1. \u4e00\u53e5\u8bdd\u7ed3\u8bba<\/h4>\n<p>SGP-Net \u7684\u6700\u5927\u4ef7\u503c\u5728\u4e8e\u628a few-shot \u533b\u5b66\u5206\u5272\u4e2d\u5e38\u89c1\u7684\u201c\u5355\u4e00 prototype + cosine matching\u201d\u66ff\u6362\u4e3a <strong>\u9891\u6bb5\u89e3\u8026 prototype + \u7279\u5f81\u6d41\u5f62\u4e0a\u7684 heat-diffusion geodesic matching<\/strong>\uff0c\u5bf9\u4f4e\u5bf9\u6bd4\u5668\u5b98\u3001\u8fb9\u754c\u6cc4\u6f0f\u548c support-query mismatch \u6709\u660e\u786e\u673a\u5236\u89e3\u91ca\uff0c\u4e5f\u7ed9\u9891\u57df\u6a21\u5757\u5982\u4f55\u5d4c\u5165\u533b\u5b66\u5206\u5272\u63d0\u4f9b\u4e86\u53ef\u590d\u7528\u8303\u5f0f\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u8bba\u6587\u7814\u7a76\u7684\u662f few-shot medical image segmentation\uff1a\u6d4b\u8bd5\u65f6\u76ee\u6807\u5668\u5b98\u7c7b\u522b\u5728\u8bad\u7ec3\u9636\u6bb5\u4e0d\u53ef\u89c1\uff0c\u53ea\u7ed9 1 \u5f20\u6216\u5c11\u91cf support image\/mask\uff0c\u9700\u8981\u5206\u5272 query image \u4e2d\u7684\u65b0\u5668\u5b98\u3002\u8fd9\u4e2a\u95ee\u9898\u91cd\u8981\uff0c\u56e0\u4e3a\u533b\u5b66\u56fe\u50cf\u50cf\u7d20\u7ea7\u6807\u6ce8\u6602\u8d35\uff0c\u7f55\u89c1\u75c5\u6216\u65b0\u5668\u5b98\/\u65b0\u4e2d\u5fc3\u573a\u666f\u4e2d\u4e0d\u53ef\u80fd\u4e3a\u6bcf\u4e2a\u76ee\u6807\u91cd\u65b0\u8bad\u7ec3\u5b8c\u6574\u5206\u5272\u6a21\u578b\u3002<\/p>\n<p>\u4f5c\u8005\u6307\u51fa\uff0c\u5f53\u524d FSMIS \u4e3b\u6d41\u65b9\u6cd5\u591a\u4ee5 prototype learning \u4e3a\u6838\u5fc3\uff1a\u4ece support mask \u533a\u57df\u901a\u8fc7 masked average pooling \u5f97\u5230\u7c7b\u522b\u539f\u578b\uff0c\u518d\u4e0e query feature \u505a cosine similarity\u3002\u4f46\u533b\u5b66\u56fe\u50cf\u4e2d\u5668\u5b98\u7eb9\u7406\u76f8\u4f3c\u3001\u4f4e\u5bf9\u6bd4\u3001\u8fb9\u754c\u6a21\u7cca\uff0c\u7b80\u5355\u4f59\u5f26\u76f8\u4f3c\u5ea6\u5bb9\u6613\u628a\u76f8\u90bb\u7ec4\u7ec7\u8bef\u8ba4\u4e3a\u76ee\u6807\uff0c\u6216\u8005\u5728\u76ee\u6807\u5185\u90e8\u4ea7\u751f\u65ad\u88c2\u54cd\u5e94\u3002<\/p>\n<p><strong>\u5185\u90e8 paper map\uff1a<\/strong> \u672c\u6587\u7814\u7a76\u5c11\u6837\u672c\u533b\u5b66\u5206\u5272\u4e2d\u7684\u539f\u578b\u5339\u914d\u95ee\u9898\uff0c\u8bbe\u5b9a\u662f 1-way 1-shot episodic segmentation\u3002\u4e3b\u62db\u662f\u7528 Spectral Prototype Bank \u5c06 support\/query feature \u5206\u89e3\u4e3a low\/mid\/high \u4e09\u4e2a\u9891\u6bb5\u5e76\u5206\u522b\u63d0\u53d6 prototype\uff0c\u518d\u7528 Geodesic Matcher \u901a\u8fc7 heat diffusion \u5728 8-neighbor feature affinity graph \u4e0a\u4f20\u64ad\u5339\u914d\u4fe1\u53f7\u3002\u5b83\u58f0\u79f0\u5728 Abd-MRI\u3001Abd-CT\u3001CMR \u4e0a\u4f18\u4e8e\u73b0\u6709 FSMIS \u65b9\u6cd5\uff0c\u8bc1\u636e\u4e3b\u8981\u662f Dice \u8868\u683c\u3001Setting 1\/2 \u6cdb\u5316\u3001\u6d88\u878d\u548c\u53ef\u89c6\u5316\u3002\u771f\u6b63\u8d1f\u8f7d\u5728\u201c\u9891\u6bb5\u662f\u5426\u771f\u7684\u5bf9\u5e94 silhouette\/texture\/boundary\u201d\u548c\u201cheat diffusion geodesic \u662f\u5426\u6bd4 cosine \u66f4\u7a33\u201d\u3002\u4e3b\u8981\u98ce\u9669\u662f\u4ecd\u662f 2D slice \u7ea7\u30011-way 1-shot\u3001\u6807\u51c6\u5c0f\u6570\u636e\u96c6\u8bc4\u4f30\uff0c\u4e14\u6ca1\u6709\u8de8\u4e2d\u5fc3\u5927\u89c4\u6a21\u4e34\u5e8a\u9a8c\u8bc1\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u5f52\u7eb3\u4e86\u4e24\u4e2a\u5173\u952e\u4e0d\u8db3\uff1a<\/p>\n<ol>\n<li><strong>Cue entanglement\uff1a<\/strong> \u5355\u4e2a prototype \u540c\u65f6\u627f\u8f7d\u5668\u5b98\u8f6e\u5ed3\u3001\u5185\u90e8\u7eb9\u7406\u548c\u8fb9\u754c\u4fe1\u606f\uff1bsupport \u4e0e query \u5728\u4efb\u4e00 cue \u4e0a\u4e0d\u5339\u914d\uff0c\u90fd\u4f1a\u6c61\u67d3\u6574\u4f53\u539f\u578b\u5339\u914d\u3002\u591a prototype \u65b9\u6cd5\u901a\u5e38\u53ea\u662f\u6309\u7a7a\u95f4\u5b50\u533a\u57df\u62c6\u5206\uff0c\u5e76\u6ca1\u6709\u628a shape\/texture\/boundary \u4ece\u8868\u793a\u4e0a\u89e3\u8026\u3002<\/li>\n<li><strong>Topology-blind matching\uff1a<\/strong> cosine similarity \u5728 ambient Euclidean feature space \u4e2d\u8ba1\u7b97\u8ddd\u79bb\uff0c\u4e0d\u8003\u8651\u7279\u5f81\u6d41\u5f62\u8fde\u901a\u6027\u3002\u7ed3\u679c\u662f\uff1a\u4e00\u4e2a\u6b27\u6c0f\u8ddd\u79bb\u8fd1\u4f46\u4e0d\u5728\u540c\u4e00 manifold \u4e0a\u7684\u76f8\u90bb\u7ec4\u7ec7\u53ef\u80fd\u5f97\u9ad8\u5206\uff0c\u800c\u4e00\u4e2a\u4f4d\u4e8e\u540c\u4e00\u5668\u5b98 manifold \u4f46\u5c40\u90e8\u5916\u89c2\u5dee\u5f02\u8f83\u5927\u7684\u50cf\u7d20\u53ef\u80fd\u5f97\u4f4e\u5206\u3002<\/li>\n<\/ol>\n<p>\u8fd9\u4e24\u4e2a\u95ee\u9898\u90fd\u4e0e\u533b\u5b66\u56fe\u50cf\u5206\u5272\u7684\u5178\u578b\u5931\u8d25\u6a21\u5f0f\u76f4\u63a5\u76f8\u5173\uff1a\u4f4e\u5bf9\u6bd4\u5668\u5b98\u5185\u90e8\u54cd\u5e94\u7834\u788e\u3001\u8fb9\u754c\u5904\u5411\u76f8\u4f3c\u90bb\u8fd1\u7ec4\u7ec7\u6cc4\u6f0f\u3002<\/p>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>SGP-Net \u7684\u7ed3\u6784\u53ef\u4ee5\u62c6\u6210\u56db\u5c42\uff1a<\/p>\n<ol>\n<li><strong>Shared feature encoder<\/strong>\uff1asupport image \u548c query image \u7ecf\u5171\u4eab encoder \u5f97\u5230 $F_s, F_q \\in \\mathbb{R}^{B\\times C\\times h\\times w}$\u3002\u5b9e\u73b0\u4e2d\u4f7f\u7528 ResNet-101 backbone\uff0cCOCO \u9884\u8bad\u7ec3\uff0c\u8f93\u51fa stride 8\u3002<\/li>\n<li><strong>Spectral Prototype Bank, SPB<\/strong>\uff1a\u5bf9 support\/query feature \u505a 2D FFT\uff0c\u7528\u53ef\u5b66\u4e60 radial cutoff \u5c06\u9891\u57df\u5206\u6210 low\/mid\/high \u4e09\u6bb5\u3002\u6bcf\u4e2a\u9891\u6bb5 inverse FFT \u56de\u7a7a\u95f4\u57df\uff0c\u518d\u5728 support mask \u5185\u505a masked average pooling\uff0c\u5f97\u5230 $P^{low},P^{mid},P^{high}$ \u4e09\u4e2a prototype\u3002<\/li>\n<li><strong>Geodesic Matcher, GM<\/strong>\uff1a\u5bf9\u6bcf\u4e2a\u9891\u6bb5\u5148\u8ba1\u7b97 cosine map\uff0c\u518d\u901a\u8fc7 soft seeding \u548c heat diffusion \u5728 8-neighbor feature affinity graph \u4e0a\u4f20\u64ad\uff0c\u5f97\u5230 geodesic reachability score\uff1b\u6700\u540e\u4ee5 learnable gate \u878d\u5408 cosine\/geodesic score\uff0c\u5e76\u6309\u50cf\u7d20\u5bf9\u4e09\u6bb5 prototype \u505a softmax \u52a0\u6743\u878d\u5408\u3002<\/li>\n<li><strong>Dual decoder<\/strong>\uff1a\u540c\u4e00 Spectral-Geodesic Prototype Module \u5206\u522b\u7528 foreground mask \u548c background mask \u8c03\u7528\u4e24\u6b21\uff0c\u5171\u4eab\u53c2\u6570\uff0c\u5f97\u5230 foreground\/background matched features\uff0c\u7ecf\u4e24\u4e2a decoder \u4ea7\u751f logits\uff0csoftmax \u5f97\u5230\u6700\u7ec8 mask\u3002<\/li>\n<\/ol>\n<p>\u8bad\u7ec3\u635f\u5931\u5305\u62ec primary segmentation loss\u3001boundary-aware loss\uff0c\u4ee5\u53ca support\/query role-swapped alignment loss\u3002\u540e\u8005\u628a query prediction \u4e8c\u503c\u5316\u4e3a pseudo-mask\uff0c\u53cd\u5411\u7528 query \u4f5c\u4e3a support \u6765\u9884\u6d4b\u539f support mask\uff0c\u9f13\u52b1 prototype matching \u7684\u53cc\u5411\u4e00\u81f4\u6027\u3002<\/p>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<p><strong>\u6a21\u5757 A\uff1aSpectral Prototype Bank<\/strong><br \/>\n- \u8f93\u5165\uff1asupport\/query feature maps $F_s,F_q$ \u548c support mask $M_s$\u3002<br \/>\n- \u64cd\u4f5c\uff1a\u5bf9 $F_s,F_q$ \u505a real-input FFT\uff1b\u7528\u4e24\u4e2a\u53ef\u5b66\u4e60\u534a\u5f84 $r_1,r_2$ \u5207\u5206 low\/mid\/high radial frequency bands\u3002\u4e3a\u4e86\u53ef\u5fae\uff0chard mask \u88ab sigmoid roll-off \u66ff\u4ee3\uff0c\u5e76\u901a\u8fc7 softplus \u53c2\u6570\u5316\u786e\u4fdd $r_1&gt;0,r_2&gt;r_1$\u3002<br \/>\n- \u8f93\u51fa\uff1a\u4e09\u6bb5 query band features \u548c\u4e09\u4e2a support prototypes\u3002<br \/>\n- \u89e3\u51b3\u7684\u95ee\u9898\uff1a\u5c06\u5168\u5c40\u8f6e\u5ed3\u3001\u5668\u5b98\u5185\u90e8\u7eb9\u7406\u3001\u8fb9\u754c\u7ec6\u8282\u62c6\u5f00\uff0c\u907f\u514d\u5355\u4e00 prototype \u7684 cue entanglement\u3002<br \/>\n- \u521b\u65b0\u6027\u5224\u65ad\uff1a\u533b\u5b66\u5206\u5272\u4e2d\u9891\u57df\u589e\u5f3a\u4e0d\u65b0\uff0c\u4f46\u201c\u5728 prototype \u7ea7\u522b\u505a\u9891\u6bb5\u89e3\u8026\uff0c\u800c\u4e0d\u662f\u53ea\u505a feature enhancement\u201d\u662f\u6bd4\u8f83\u6e05\u6670\u7684\u6539\u9020\u70b9\u3002<br \/>\n- \u53ef\u8fc1\u79fb\u6027\uff1a\u53ef\u4ee5\u79fb\u690d\u5230 prototype-based polyp segmentation\u3001few-shot organ segmentation\u3001\u751a\u81f3 support-query matching \u7684 SAM\/MedSAM prompt refinement\uff1b\u4f46\u5bf9\u666e\u901a fully-supervised U-Net \u9700\u8981\u91cd\u65b0\u8bbe\u8ba1\uff0c\u56e0\u4e3a\u6ca1\u6709 support prototype\u3002<\/p>\n<p><strong>\u6a21\u5757 B\uff1aGeodesic Matcher<\/strong><br \/>\n- \u8f93\u5165\uff1araw query feature\u3001\u4e09\u6bb5 query features\u3001\u4e09\u6bb5 prototypes\u3002<br \/>\n- \u64cd\u4f5c\uff1a\u6bcf\u6bb5\u5148\u7b97 cosine similarity\uff1b\u53d6 0.85 quantile \u4f5c\u4e3a soft seed \u9608\u503c\uff1b\u57fa\u4e8e query band feature \u6784\u5efa 8-neighbor affinity\uff1b\u8fed\u4ee3 $T=5$ \u6b21 heat diffusion\uff0c\u5f97\u5230 geodesic reachability\uff1b\u7528 learnable $\\alpha^{(k)}$ \u878d\u5408 cosine \u548c geo score\u3002<br \/>\n- \u8f93\u51fa\uff1a$F_{matched}\\in\\mathbb{R}^{B\\times(2C+3)\\times h\\times w}$\uff0c\u7531 raw query feature\u3001blended prototype feature\u3001\u4e09\u6bb5 score stack \u62fc\u63a5\u800c\u6765\u3002<br \/>\n- \u89e3\u51b3\u7684\u95ee\u9898\uff1a\u4e0d\u518d\u8ba9\u5339\u914d\u4fe1\u53f7\u8de8\u8d8a feature manifold \u7684\u201c\u6377\u5f84\u201d\uff0c\u4ece\u800c\u6291\u5236\u5916\u89c2\u76f8\u4f3c\u4f46\u4e0d\u8fde\u901a\u7684 off-manifold \u533a\u57df\u3002<br \/>\n- \u521b\u65b0\u6027\u5224\u65ad\uff1a\u628a heat method \/ Varadhan-style geodesic approximation \u5f15\u5165 prototype matching \u662f\u8bba\u6587\u6700\u6838\u5fc3\u7684\u6280\u672f\u70b9\uff1b\u4f46\u5b83\u4ecd\u662f\u79bb\u6563\u5c40\u90e8 diffusion heuristic\uff0c\u4e0d\u7b49\u4e8e\u4e25\u683c\u51e0\u4f55\u8ddd\u79bb\u4f30\u8ba1\u3002<br \/>\n- \u5bf9 3D\/\u606f\u8089\u4efb\u52a1\u8fc1\u79fb\uff1a\u5bf9 3D medical segmentation \u53ef\u6539\u6210 6\/18\/26-neighbor volumetric affinity\uff0c\u4f46\u8ba1\u7b97\u548c\u663e\u5b58\u4f1a\u660e\u663e\u589e\u52a0\uff1b\u5bf9 polyp segmentation\uff0c\u82e5\u505a few-shot \u6216 domain adaptation\uff0c\u8fb9\u754c\u6cc4\u6f0f\u4e0e\u4f4e\u5bf9\u6bd4\u95ee\u9898\u76f8\u4f3c\uff0c\u56e0\u6b64\u6709\u8fc1\u79fb\u4ef7\u503c\u3002<\/p>\n<p><strong>\u6a21\u5757 C\uff1aForeground-background symmetric design<\/strong><br \/>\n\u524d\u666f\u7528 $M_s$\uff0c\u80cc\u666f\u7528 $1-M_s$\uff0c\u540c\u4e00\u6a21\u5757\u5171\u4eab\u53c2\u6570\u8c03\u7528\u4e24\u6b21\uff0c\u518d\u7528\u53cc decoder \u8f93\u51fa foreground\/background logits\u3002\u8fd9\u4e2a\u8bbe\u8ba1\u53ef\u4ee5\u964d\u4f4e\u80cc\u666f\u8bef\u5339\u914d\uff0c\u5c24\u5176\u9002\u5408\u533b\u5b66\u56fe\u50cf\u4e2d\u76ee\u6807\u5c0f\u3001\u80cc\u666f\u590d\u6742\u7684\u8bbe\u5b9a\u3002<\/p>\n<p><strong>\u6a21\u5757 D\uff1arole-swapped alignment loss<\/strong><br \/>\n\u628a query prediction \u4f5c\u4e3a pseudo support mask \u53cd\u5411\u9884\u6d4b support\uff0c\u53ef\u4ee5\u7406\u89e3\u4e3a episodic consistency regularization\u3002\u4f18\u70b9\u662f\u589e\u5f3a support-query \u5bf9\u79f0\u6027\uff1b\u98ce\u9669\u662f\u5982\u679c query pseudo-mask \u521d\u671f\u8d28\u91cf\u5dee\uff0c\u53ef\u80fd\u5f15\u5165\u566a\u58f0\uff0c\u56e0\u6b64\u4f5c\u8005\u52a0\u5165\u7a7a mask fallback\u3002<\/p>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p>\u5b9e\u9a8c\u9075\u5faa 1-way 1-shot episodic protocol\uff0c\u5728\u4e09\u4e2a\u516c\u5f00\u6570\u636e\u96c6\u4e0a\u8bc4\u4f30\uff1a<\/p>\n<ul>\n<li>Abd-MRI \/ CHAOS-T2\uff1a20 \u4e2a T2-SPIR \u8179\u90e8 MRI\uff0c4 \u4e2a\u5668\u5b98\uff1aleft kidney\u3001right kidney\u3001liver\u3001spleen\u3002<\/li>\n<li>Abd-CT \/ SABS\uff1a30 \u4e2a abdominal CT\uff0c\u540c\u6837 4 \u4e2a\u5668\u5b98\u3002<\/li>\n<li>CMR\uff1a35 \u4e2a cardiac MRI\uff0c3 \u4e2a\u7ed3\u6784\uff1aLV-BP\u3001LV-MYO\u3001RV\u3002<\/li>\n<\/ul>\n<p>\u8bc4\u4f30\u8bbe\u7f6e\u5305\u62ec\uff1a<br \/>\n- Setting 1\uff1a\u5305\u542b test-class organ \u7684 slice \u53ef\u51fa\u73b0\u5728\u8bad\u7ec3\u4e2d\uff0c\u4f46\u5bf9\u5e94\u50cf\u7d20\u4e0d\u6807\u6ce8\u4e3a\u6d4b\u8bd5\u7c7b\u3002<br \/>\n- Setting 2\uff1a\u66f4\u4e25\u683c\uff0c\u5305\u542b test class \u7684 slice \u4ece\u8bad\u7ec3\u4e2d\u5b8c\u5168\u79fb\u9664\uff1b\u66f4\u63a5\u8fd1 truly unseen class\u3002<br \/>\n- 5-fold cross-validation\uff0c\u6bcf\u6b21 held-out \u4e00\u4e2a\u5668\u5b98\u4f5c\u4e3a novel class\u3002<br \/>\n- \u6307\u6807\uff1aDice Similarity Coefficient\u3002<\/p>\n<p>\u4e3b\u8981\u7ed3\u679c\uff1a<br \/>\n- Abd-MRI Setting 1\uff1aSGP-Net mean Dice 84.83%\uff0c\u9ad8\u4e8e UPRE-Net 84.05%\u3002<br \/>\n- Abd-CT Setting 1\uff1aSGP-Net 81.22%\uff0c\u9ad8\u4e8e AVT-ProNet 80.60%\u3002<br \/>\n- Abd-MRI Setting 2\uff1aSGP-Net 81.41%\uff0c\u9ad8\u4e8e UPRE-Net 81.05%\u3002<br \/>\n- Abd-CT Setting 2\uff1aSGP-Net 80.94%\uff0c\u9ad8\u4e8e UPRE-Net 79.47%\u3002<br \/>\n- CMR\uff1aSGP-Net mean Dice 81.16%\uff0c\u9ad8\u4e8e UPRE-Net 80.42%\u3002<\/p>\n<p>\u6d88\u878d\u5f88\u5173\u952e\uff1a<br \/>\n- baseline single prototype + cosine\uff1a79.42%\u3002<br \/>\n- \u53ea\u52a0 SPB\uff1a82.18%\u3002<br \/>\n- \u53ea\u52a0 GM\uff1a82.53%\u3002<br \/>\n- SPB+GM\uff1a84.83%\u3002<br \/>\n\u8fd9\u652f\u6301\u4f5c\u8005\u5173\u4e8e\u4e8c\u8005\u89e3\u51b3\u4e0d\u540c\u5931\u8d25\u6a21\u5f0f\u7684\u8bba\u70b9\u3002<\/p>\n<p>\u8d85\u53c2\u6570\u6d88\u878d\uff1a<br \/>\n- \u9891\u6bb5\u6570 $K=3$ \u6700\u597d\uff0cK=4\/5 \u4e0b\u964d\uff0c\u8bf4\u660e\u8fc7\u7ec6\u9891\u6bb5\u53ef\u80fd\u5197\u4f59\/\u91cd\u53e0\u3002<br \/>\n- diffusion steps $T=5$ \u6700\u597d\uff0cT=7\/8 \u7565\u964d\uff0c\u8bf4\u660e\u6269\u6563\u8fc7\u591a\u53ef\u80fd\u5e73\u6ed1\u8fc7\u5ea6\u3002<\/p>\n<p>\u8ba1\u7b97\u6210\u672c\uff1aSGP-Net 51.82M \u53c2\u6570\u3001279.63G FLOPs\u300114.6 FPS\uff1b\u6bd4 UPRE-Net\/DIFD \u53c2\u6570\u7565\u5c11\uff0c\u901f\u5ea6\u63a5\u8fd1 DSPNet\uff0c\u4f46\u4ecd\u4e0d\u662f\u8f7b\u91cf\u6a21\u578b\u3002<\/p>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u53ef\u4fe1\u70b9\uff1a<br \/>\n- \u6570\u636e\u96c6\u8986\u76d6 MRI\u3001CT\u3001cardiac MRI\uff0c\u4e14\u5305\u542b Setting 2 \u8fd9\u79cd\u66f4\u4e25\u683c unseen-class \u8bc4\u4f30\u3002<br \/>\n- baseline \u6bd4\u8f83\u5305\u62ec PA-Net\u3001SSL-ALPNet\u3001ADNet\u3001Q-Net\u3001PAMI\u3001DSPNet\u3001AVT-ProNet\u3001FAMNet\u3001UPRE-Net\u3001DIFD\u3001MACCFA\uff0c\u8986\u76d6\u7ecf\u5178\u548c\u8fd1\u5e74 FSMIS \u65b9\u6cd5\u3002<br \/>\n- \u6709\u6a21\u5757\u6d88\u878d\u3001K\/T \u8d85\u53c2\u6570\u6d88\u878d\u3001\u53ef\u89c6\u5316\u3001fold variance\u3001Setting 1\u21922 performance drop\u3001\u590d\u6742\u5ea6\u8868\u3002<br \/>\n- \u90e8\u5206\u65b9\u6cd5\u6807\u6ce8\u4e3a\u5b98\u65b9\u4ee3\u7801\u7edf\u4e00\u590d\u73b0\uff0c\u907f\u514d\u5b8c\u5168\u4f9d\u8d56\u8bba\u6587\u5f15\u7528\u7ed3\u679c\u3002<\/p>\n<p>\u9700\u8981\u8c28\u614e\u7684\u70b9\uff1a<br \/>\n- \u4e3b\u8981\u4ecd\u662f\u5c0f\u89c4\u6a21\u516c\u5f00\u6570\u636e\u96c6\uff0c\u4e14 3D volumes \u88ab reformatted into 2D axial slices\uff1b\u8fd9\u4e0d\u80fd\u5145\u5206\u8bc1\u660e\u771f\u5b9e 3D \u8fde\u7eed\u4f53\u5206\u5272\u80fd\u529b\u3002<br \/>\n- \u7edd\u5bf9\u63d0\u5347\u5bf9\u6700\u5f3a baseline \u5f80\u5f80\u5728 0.3\u20131.5 Dice \u70b9\uff0c\u867d\u7136\u7a33\u5b9a\u4f46\u4e0d\u662f\u538b\u5012\u6027\u63d0\u5347\uff1b\u7f3a\u5c11\u7edf\u8ba1\u663e\u8457\u6027\u68c0\u9a8c\u3002<br \/>\n- backbone \u4f7f\u7528 COCO \u9884\u8bad\u7ec3 ResNet-101\uff0c\u4e0e\u73b0\u4ee3\u533b\u5b66 foundation models \/ SAM\/MedSAM\/nnU-Net \u7cfb\u5217\u5e76\u975e\u540c\u4e00\u7ade\u4e89\u8303\u5f0f\u3002<br \/>\n- \u4ee3\u7801\u867d\u7ed9\u51fa GitHub \u94fe\u63a5\uff0c\u4f46\u9700\u8981\u5b9e\u9645\u8fd0\u884c\u624d\u80fd\u786e\u8ba4\u53ef\u590d\u73b0\u6027\uff1b\u672c\u6587\u672a\u62a5\u544a\u591a\u968f\u673a\u79cd\u5b50\u8bef\u5dee\u6761\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>\uff1aSGP-Net \u4e0d\u662f\u81ea\u52a8\u914d\u7f6e\u5f0f\u5168\u76d1\u7763\u6846\u67b6\uff0c\u800c\u662f episodic few-shot prototype segmentation\uff1bdecoder \u53ef\u501f\u9274 U-Net \u601d\u8def\uff0c\u4f46\u6838\u5fc3\u4e0d\u662f encoder-decoder \u67b6\u6784\uff0c\u800c\u662f support-query matching\u3002<\/li>\n<li>\u4e0e <strong>MedNeXt \/ ConvNeXt-like segmentation<\/strong>\uff1a\u6ca1\u6709\u76f4\u63a5\u5bf9\u6bd4\uff1b\u5176\u9891\u57df\/\u6d41\u5f62\u6a21\u5757\u7406\u8bba\u4e0a\u53ef\u4f5c\u4e3a feature matching head\uff0c\u800c\u4e0d\u662f\u66ff\u4ee3 ConvNeXt backbone\u3002<\/li>\n<li>\u4e0e <strong>UNetR \/ Swin-UNet \/ TransUNet \/ TransFuse<\/strong>\uff1a\u8fd9\u4e9b\u591a\u662f\u5168\u76d1\u7763 encoder \u6539\u9020\uff1bSGP-Net \u7684\u521b\u65b0\u53d1\u751f\u5728 prototype representation \u548c matching metric\u3002<\/li>\n<li>\u4e0e <strong>Mamba \/ VMamba \/ SegMamba \/ DAMamba<\/strong>\uff1a\u672c\u6587\u4e0d\u662f Mamba \u6a21\u578b\uff0c\u4f46 geodesic diffusion \u7684\u201c\u6cbf\u5c40\u90e8 affinity \u4f20\u64ad\u5339\u914d\u4fe1\u53f7\u201d\u4e0e Mamba \u7684\u957f\u7a0b\u9009\u62e9\u6027\u626b\u63cf\u53ef\u5f62\u6210\u4e92\u8865\uff1aMamba \u66f4\u504f\u5e8f\u5217\/\u5168\u5c40\u4f9d\u8d56\uff0cSGP-Net \u66f4\u504f support-query manifold matching\u3002<\/li>\n<li>\u4e0e <strong>foundation model \/ MedSAM<\/strong>\uff1aSGP-Net \u6ca1\u6709\u4f9d\u8d56 SAM prompt\uff1b\u4f46\u5b83\u5bf9 few-shot support mask \u7684\u5229\u7528\u65b9\u5f0f\uff0c\u53ef\u542f\u53d1 MedSAM prompt refinement \u6216 prototype-conditioned SAM decoder\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u5982\u679c\u4f60\u7684\u65b9\u5411\u5305\u62ec polyp segmentation\u3001DAMamba \u6539\u9020\u6216\u65b0\u533b\u5b66\u5206\u5272\u6846\u67b6\u8bbe\u8ba1\uff0c\u8fd9\u7bc7\u8bba\u6587\u6709\u4e09\u7c7b\u4ef7\u503c\uff1a<\/p>\n<ol>\n<li><strong>\u65b9\u6cd5\u6a21\u5757\u4ef7\u503c\uff1a<\/strong> SPB \u7684 low\/mid\/high frequency prototype decomposition \u53ef\u4ee5\u501f\u9274\u5230 polyp segmentation \u7684 boundary-aware decoder \u4e2d\uff0c\u5c24\u5176\u662f\u628a high-frequency branch \u4e13\u95e8\u7528\u4e8e\u8fb9\u754c refinement\u3002<\/li>\n<li><strong>\u673a\u5236\u8868\u8fbe\u4ef7\u503c\uff1a<\/strong> \u8bba\u6587\u628a\u201c\u9891\u57df = silhouette\/texture\/boundary\u201d\u548c\u201cgeodesic matching = suppress off-manifold look-alikes\u201d\u8bb2\u5f97\u6bd4\u8f83\u6e05\u695a\uff0c\u53ef\u7528\u4e8e related work \u6216 introduction \u4e2d\u89e3\u91ca\u4e3a\u4ec0\u4e48\u533b\u5b66\u56fe\u50cf\u9700\u8981\u8d85\u8d8a\u666e\u901a cosine \/ attention matching\u3002<\/li>\n<li><strong>\u5bf9 DAMamba \u7684\u542f\u53d1\uff1a<\/strong> \u53ef\u8003\u8651\u628a Mamba scanning \u540e\u7684\u7279\u5f81\u6784\u5efa local affinity graph\uff0c\u518d\u505a\u8f7b\u91cf diffusion \/ geodesic refinement\uff1b\u6216\u8005\u628a Mamba \u5206\u652f\u6309\u9891\u6bb5\u7ea6\u675f\uff0c\u8ba9\u4e0d\u540c scan \u5206\u652f\u5206\u522b\u5173\u6ce8\u5668\u5b98\u5185\u90e8\u3001\u7eb9\u7406\u548c\u8fb9\u754c\u3002<\/li>\n<\/ol>\n<p>\u4f46\u5982\u679c\u4f60\u505a\u7684\u662f fully-supervised polyp segmentation baseline\uff0cSGP-Net \u4e0d\u80fd\u76f4\u63a5\u4f5c\u4e3a\u540c\u7c7b SOTA \u5bf9\u6bd4\uff0c\u56e0\u4e3a\u5b83\u7684\u95ee\u9898\u8bbe\u5b9a\u662f few-shot\uff1b\u66f4\u9002\u5408\u4f5c\u4e3a\u6a21\u5757\u7075\u611f\u6216 few-shot\/low-label \u65b9\u5411\u53c2\u8003\u3002<\/p>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5f3a\u70c8\u5efa\u8bae\u7cbe\u8bfb\u6280\u672f\u90e8\u5206\u548c\u6d88\u878d\u90e8\u5206\u3002<\/strong> \u8fd9\u7bc7\u8bba\u6587\u7684\u8d21\u732e\u70b9\u6e05\u695a\uff0c\u673a\u5236\u6bd4\u8bb8\u591a\u201c\u5806\u6a21\u5757\u201d\u5f0f\u5206\u5272\u8bba\u6587\u66f4\u53ef\u89e3\u91ca\uff1b\u5efa\u8bae\u91cd\u70b9\u8bfb Section III-C\/D\u3001Algorithm 1\u3001Table III\u2013VI\u3001Fig. 7\u201310\u3002\u82e5\u65f6\u95f4\u6709\u9650\uff0cRelated Work \u53ef\u7565\u8bfb\u3002\u590d\u73b0\u524d\u9700\u6ce8\u610f\uff1a\u5b83\u4f9d\u8d56 episodic sampling\u3001supervoxel pseudo-mask\u3001Setting 1\/2 \u6570\u636e\u5212\u5206\u548c chunk-based 3D evaluation\uff0c\u590d\u73b0\u5b9e\u9a8c\u7ec6\u8282\u6bd4\u666e\u901a U-Net \u8bad\u7ec3\u66f4\u590d\u6742\u3002<\/p>\n<hr \/>\n<h2>\u8bba\u6587 2\uff1aHarmonized Feature Conditioning and Frequency-Prompt Personalization for Multi-Rater Medical Segmentation<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aHarmonized Feature Conditioning and Frequency-Prompt Personalization for Multi-Rater Medical Segmentation<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aSanaz Karimijafarbigloo, Armin Khosravi, Alireza Kheyrkhah, Reza Azad, Mauricio Reyes, Dorit Merhof \/ Sanaz Karimijafarbigloo<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-06 submitted<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint\uff1barXiv \u9875\u9762\u5907\u6ce8 Accepted in main CVPR 2026<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.08210<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.08210<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1a\u8bba\u6587\u6458\u8981\u5904\u5199 \u201cGitHub code\u201d\uff0c\u4f46\u6b63\u6587\/\u9875\u9762\u4e2d\u672a\u80fd\u786e\u8ba4\u5177\u4f53 URL\uff1b\u6682\u8bb0\u4e3a\u672a\u83b7\u53d6<\/li>\n<li>\u4efb\u52a1\uff1aMulti-rater medical image segmentation\uff1bprobabilistic \/ personalized segmentation\uff1buncertainty-aware segmentation<\/li>\n<li>\u6570\u636e\u96c6\uff1aLIDC-IDRI\u3001NPC-170\uff1b\u8865\u5145\u6750\u6599\u542b Kvasir-SEG noisy-label \u6269\u5c55\u5b9e\u9a8c<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1aProbabilistic U-Net backbone\uff1bNoise Harmonizer\uff1bfrequency-domain rater-aware prompts\uff1bGED regularization<\/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 multi-rater \u533b\u5b66\u5206\u5272\u4e2d\u7684\u4e24\u7c7b\u4e0d\u786e\u5b9a\u6027\u660e\u786e\u62c6\u5f00\uff1a\u7528 <strong>Noise Harmonizer<\/strong> \u5904\u7406 scanner\/acquisition artifact\uff0c\u7528 <strong>high-frequency prompt personalization<\/strong> \u5904\u7406\u4e13\u5bb6\u8fb9\u754c\u98ce\u683c\u5dee\u5f02\uff0c\u5e76\u7528 GED \u7ea6\u675f\u9884\u6d4b\u5206\u5e03\u8d34\u8fd1\u771f\u5b9e\u591a\u6807\u6ce8\u5206\u5e03\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u533b\u5b66\u5206\u5272\u7684\u201cground truth\u201d\u5e76\u4e0d\u603b\u662f\u552f\u4e00\uff1a\u80ba\u7ed3\u8282\u3001\u9f3b\u54bd\u764c GTV\u3001\u606f\u8089\u8fb9\u754c\u7b49\u4efb\u52a1\u4e2d\uff0c\u4e0d\u540c\u4e13\u5bb6\u53ef\u80fd\u56e0\u4e3a\u4e34\u5e8a\u7ecf\u9a8c\u3001\u8fb9\u754c\u5224\u65ad\u3001\u5f71\u50cf\u8d28\u91cf\u548c\u75c5\u7076\u6a21\u7cca\u7a0b\u5ea6\u4ea7\u751f\u4e0d\u540c\u6807\u6ce8\u3002\u4f20\u7edf majority voting\u3001STAPLE \u6216\u5e73\u5747 soft label \u4f1a\u628a\u8fd9\u79cd\u5dee\u5f02\u538b\u6210\u5355\u4e00\u6807\u7b7e\uff0c\u5bfc\u81f4\u6a21\u578b\u8fc7\u5ea6\u81ea\u4fe1\uff0c\u5e76\u4e22\u5931\u4e34\u5e8a\u4e0a\u6709\u610f\u4e49\u7684\u4e0d\u786e\u5b9a\u6027\u3002<\/p>\n<p>\u672c\u6587\u7814\u7a76 multi-rater medical segmentation\uff1a\u7ed9\u5b9a\u540c\u4e00\u56fe\u50cf\u548c\u591a\u4e2a\u4e13\u5bb6\u6807\u6ce8\uff0c\u4e0d\u4ec5\u8981\u8f93\u51fa\u51c6\u786e mask\uff0c\u8fd8\u8981\u80fd\u8868\u8fbe\u591a\u79cd plausible segmentations\uff0c\u5e76\u80fd\u751f\u6210 rater-specific personalized prediction\u3002<\/p>\n<p><strong>\u5185\u90e8 paper map\uff1a<\/strong> \u672c\u6587\u7814\u7a76\u591a\u4e13\u5bb6\u533b\u5b66\u5206\u5272\u4e2d scanner noise \u4e0e annotator variability \u6df7\u6742\u7684\u95ee\u9898\uff0c\u8bbe\u5b9a\u662f LIDC-IDRI \u80ba\u7ed3\u8282 CT\u3001NPC-170 \u591a\u6a21\u6001 MRI \u7b49\u591a\u6807\u6ce8\u5206\u5272\u3002\u4e3b\u62db\u662f\u4ee5 Probabilistic U-Net \u4e3a backbone\uff0c\u5148\u7528 Noise Harmonizer \u8c03\u5236 latent\/decoder feature \u4ee5\u7a33\u5b9a acquisition-induced artifacts\uff0c\u518d\u7528 DWT high-frequency prompt module \u8868\u8fbe\u4e13\u5bb6\u8fb9\u754c\u98ce\u683c\uff0c\u5e76\u7528 GED regularization \u5bf9\u9f50\u9884\u6d4b\u5206\u5e03\u548c empirical annotation distribution\u3002\u5b83\u58f0\u79f0\u5728 GED\u3001soft Dice\u3001personalized Dice\u3001noise robustness\u3001domain shift \u4e0a\u4f18\u4e8e Prob. U-Net \u548c D-Persona \u7b49\u65b9\u6cd5\u3002\u771f\u6b63\u8d1f\u8f7d\u5728\u201c\u566a\u58f0 harmonization \u4e0e\u4e13\u5bb6\u98ce\u683c personalization \u662f\u5426\u771f\u7684\u88ab\u5206\u79bb\u201d\u4ee5\u53ca\u201cfrequency prompt \u662f\u5426\u8db3\u4ee5\u89e3\u91ca rater style\u201d\u3002\u4e3b\u8981\u98ce\u9669\u662f\u65b9\u6cd5\u590d\u6742\u3001\u4e24\u9636\u6bb5\u8bad\u7ec3\u3001\u4ee3\u7801\u94fe\u63a5\u672a\u786e\u8ba4\uff0c\u4e14\u90e8\u5206\u5b9e\u9a8c\/\u6d88\u878d\u5728 supplement \u4e2d\uff0c\u4e3b\u6587\u5bf9\u6a21\u5757\u72ec\u7acb\u8d21\u732e\u5448\u73b0\u4e0d\u591f\u96c6\u4e2d\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u6279\u8bc4\u73b0\u6709\u65b9\u6cd5\u7684\u4e0d\u8db3\u4e3b\u8981\u6709\u4e09\u7c7b\uff1a<\/p>\n<ol>\n<li><strong>Consensus \/ label fusion \u4e22\u5931\u5dee\u5f02\uff1a<\/strong> majority voting\u3001STAPLE\u3001soft label fusion \u4ecd\u7136\u503e\u5411\u4e8e\u6784\u9020\u4e00\u4e2a\u201c\u53ef\u9760\u5355\u771f\u503c\u201d\uff0c\u4e0d\u80fd\u4fdd\u7559\u4e13\u5bb6\u7279\u5f02\u7684 boundary style\u3002<\/li>\n<li><strong>Probabilistic latent model \u5bb9\u6613\u6b20\u6b63\u5219\uff1a<\/strong> Probabilistic U-Net\u3001PHiSeg \u7b49\u53ef\u4ee5\u91c7\u6837\u591a\u79cd mask\uff0c\u4f46 latent distribution \u53ef\u80fd\u7a00\u758f\u6216 under-regularized\uff0c\u751f\u6210\u6837\u672c\u5197\u4f59\u6216\u4e0e\u771f\u5b9e\u6807\u6ce8\u5206\u5e03\u4e0d\u5339\u914d\u3002<\/li>\n<li><strong>\u4e2a\u6027\u5316\u65b9\u6cd5\u6ca1\u6709\u533a\u5206\u566a\u58f0\u4e0e\u4e34\u5e8a\u5dee\u5f02\uff1a<\/strong> D-Persona\u3001DiffOSeg \u7b49\u80fd\u505a expert prompt\uff0c\u4f46\u591a\u5728 spatial feature \u4e0a\u8c03\u5236\uff1bscanner artifact\u3001acquisition noise\u3001annotation quality \u53ef\u80fd\u548c rater variability \u6df7\u5728\u540c\u4e00 latent space \u4e2d\uff0c\u5bfc\u81f4\u6a21\u578b\u628a\u566a\u58f0\u8bef\u8ba4\u4e3a\u4e34\u5e8a\u4e0d\u786e\u5b9a\u6027\u3002<\/li>\n<\/ol>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>\u65b9\u6cd5\u5efa\u7acb\u5728 Probabilistic U-Net \u4e0a\u3002\u57fa\u7840\u516c\u5f0f\u662f\uff1a<\/p>\n<p>$$p_\\theta(y|x)=\\int p_\\theta(y|x,z)p_\\theta(z|x)dz,$$<\/p>\n<p>\u5176\u4e2d prior\/posterior network \u9884\u6d4b Gaussian latent parameters\uff0c\u91c7\u6837 $z$ \u540e\u4e0e encoder feature \u62fc\u63a5\uff0cdecoder \u8f93\u51fa segmentation hypothesis\u3002<\/p>\n<p>\u672c\u6587\u52a0\u5165\u4e09\u90e8\u5206\uff1a<\/p>\n<ol>\n<li><strong>Noise Harmonizer\uff1a<\/strong> \u5728 decoder \u591a\u5c42 feature \u4e0a\u9884\u6d4b affine modulation \u53c2\u6570 $\\gamma_l,\\beta_l$\uff0c\u901a\u8fc7 $\\tilde f_l=\\gamma_l\\odot f_l+\\beta_l$ \u7a33\u5b9a scanner\/acquisition artifact\u3002<\/li>\n<li><strong>Frequency-domain Personalization Module\uff1a<\/strong> \u5bf9 feature \u505a Haar DWT\uff0c\u5206\u51fa $X_{LL},X_{LH},X_{HL},X_{HH}$\uff1b\u9ad8\u9891\u5206\u91cf\u88ab\u89c6\u4f5c\u8fb9\u754c\u7cbe\u5ea6\u3001\u7eb9\u7406\u654f\u611f\u6027\u7b49 rater style \u7684\u4e3b\u8981\u627f\u8f7d\u8005\uff0c\u7528 rater-aware prompt projection \u548c attention \u8c03\u5236\u9ad8\u9891\u7279\u5f81\uff0c\u518d IDWT \u56de full-spectrum representation\uff0c\u751f\u6210 rater-adaptive latent code $z'$\u3002<\/li>\n<li><strong>GED regularization\uff1a<\/strong> \u7528 Generalized Energy Distance \u5bf9\u9f50\u6a21\u578b\u9884\u6d4b\u5206\u5e03 $P(y|x)$ \u4e0e\u7ecf\u9a8c\u6807\u6ce8\u5206\u5e03 $A(y|x)$\uff0c\u65e2\u9f13\u52b1 fidelity\uff0c\u4e5f\u9632\u6b62\u9884\u6d4b collapse \u5230\u5355\u4e00 mask\u3002<\/li>\n<\/ol>\n<p>\u8bad\u7ec3\u662f\u4e24\u9636\u6bb5\uff1a<br \/>\n- Phase 1\uff1a\u8bad\u7ec3 encoder\u3001decoder\u3001Noise Harmonizer\uff0c\u6392\u9664 personalization head\uff0c\u91cd\u70b9\u5b66\u4e60 artifact-invariant latent representation\uff1b100 epochs\uff0cAdam\uff0clr=1e-4\u3002<br \/>\n- Phase 2\uff1a\u51bb\u7ed3 encoder\/decoder\/harmonizer\uff0c\u53ea\u8bad\u7ec3 Personalization Module\uff1b150 epochs\uff0clr=5e-5\u3002<\/p>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<p><strong>\u6a21\u5757 A\uff1aNoise Harmonizer<\/strong><br \/>\n- \u8f93\u5165\uff1adecoder layer feature $f_l$ \u4e0e learnable artifact tokens\u3002<br \/>\n- \u64cd\u4f5c\uff1a\u7528 attention \u4ece token-feature \u4ea4\u4e92\u4e2d\u5f97\u5230 $f'_l$\uff0c\u7ecf GAP \u548c MLP \u9884\u6d4b $[\\gamma_l,\\beta_l]$\uff0c\u7136\u540e affine \u8c03\u5236 $\\tilde f_l=\\gamma_l\\odot f_l+\\beta_l$\u3002<br \/>\n- \u8f93\u51fa\uff1aharmonized feature\u3002<br \/>\n- \u89e3\u51b3\u7684\u95ee\u9898\uff1a\u6291\u5236 scanner noise\u3001intensity drift\u3001motion artifact\u3001domain bias\uff0c\u4f7f latent uncertainty \u66f4\u63a5\u8fd1 anatomy \/ rater disagreement\uff0c\u800c\u4e0d\u662f acquisition artifact\u3002<br \/>\n- \u521b\u65b0\u6027\u5224\u65ad\uff1a\u7c7b\u4f3c conditional normalization \/ FiLM \u7684\u601d\u60f3\uff0c\u4f46\u660e\u786e\u9762\u5411 multi-rater segmentation \u4e2d\u201c\u566a\u58f0\u4e0e\u4e13\u5bb6\u5dee\u5f02\u6df7\u6742\u201d\u7684\u95ee\u9898\uff1b\u771f\u6b63\u521b\u65b0\u5728\u95ee\u9898\u5206\u89e3\u548c\u5b9e\u9a8c\u8bc1\u636e\uff0c\u800c\u4e0d\u662f affine modulation \u672c\u8eab\u3002<\/p>\n<p><strong>\u6a21\u5757 B\uff1aHigh-Frequency Prompt Personalization<\/strong><br \/>\n- \u8f93\u5165\uff1aharmonized feature $X$ \u548c rater-specific learnable weights \/ prompt components\u3002<br \/>\n- \u64cd\u4f5c\uff1a\u7ebf\u6027\u964d\u7ef4\u540e\u505a Haar DWT\uff0c\u5f97\u5230 LL\/LH\/HL\/HH\uff1b\u5c06 LH\/HL\/HH high-frequency maps \u62fc\u63a5\u4e3a $X_H$\uff0c\u901a\u8fc7 Rater-Aware Prompt Projection \u751f\u6210 prompt $P$\uff1b\u7528 Large Kernel Attention \u5bf9 $X_H$ \u505a rater-aware recalibration\uff1b\u518d\u4e0e low-frequency map \u7ed3\u5408\u5e76 IDWT \u91cd\u6784\u3002<br \/>\n- \u8f93\u51fa\uff1arater-adaptive latent vector $z'$\uff0c\u7528\u4e8e personalized decoding\u3002<br \/>\n- \u89e3\u51b3\u7684\u95ee\u9898\uff1a\u4e13\u5bb6\u5dee\u5f02\u901a\u5e38\u4f53\u73b0\u5728\u8fb9\u754c\u539a\u5ea6\u3001\u8fb9\u754c\u9510\u5ea6\u3001\u75c5\u7076\u8303\u56f4\u3001\u7eb9\u7406\u654f\u611f\u6027\uff0c\u8fd9\u4e9b\u66f4\u504f\u9ad8\u9891\u3002<br \/>\n- \u521b\u65b0\u6027\u5224\u65ad\uff1a\u628a rater style \u663e\u5f0f\u7ed1\u5b9a\u5230\u9ad8\u9891\u5206\u91cf\u662f\u6709\u89e3\u91ca\u529b\u7684\u8bbe\u8ba1\uff0c\u5c24\u5176\u9002\u5408\u8fb9\u754c\u4e3b\u5bfc\u7684\u533b\u5b66\u5206\u5272\uff1b\u4f46\u201c\u9ad8\u9891 = rater style\u201d\u5e76\u975e\u603b\u6210\u7acb\uff0c\u4f8b\u5982\u6709\u4e9b\u4e13\u5bb6\u5dee\u5f02\u53ef\u80fd\u662f\u8bed\u4e49\u7ea7\/\u533a\u57df\u7ea7\uff0c\u800c\u4e0d\u53ea\u662f\u8fb9\u754c\u7ea7\u3002<\/p>\n<p><strong>\u6a21\u5757 C\uff1aGED regularization<\/strong><br \/>\nGED loss \u4f7f\u7528\u8ddd\u79bb $d=1-IoU$\uff1a<\/p>\n<p>$$L_{GED}=\\frac{2}{KN}\\sum_{k=1}^K\\sum_{i=1}^N d(P_k,A_i)-\\frac{2}{K(K-1)}\\sum_{k&lt;k'}d(P_k,P_{k'}).$$<\/p>\n<p>\u7b2c\u4e00\u9879\u8ba9\u9884\u6d4b\u6837\u672c\u63a5\u8fd1\u771f\u5b9e\u4e13\u5bb6\u6807\u6ce8\u96c6\u5408\uff0c\u7b2c\u4e8c\u9879\u9f13\u52b1\u9884\u6d4b\u6837\u672c\u4e4b\u95f4\u4fdd\u6301\u591a\u6837\u6027\uff0c\u907f\u514d\u6240\u6709 samples collapse \u5230 consensus\u3002<\/p>\n<p><strong>\u6a21\u5757 D\uff1a\u4e24\u9636\u6bb5\u8bad\u7ec3<\/strong><br \/>\n\u5148\u5b66\u7a33\u5b9a\u7684 shared anatomical manifold\uff0c\u518d\u5728\u5176\u4e0a\u5b66 rater-specific spectral prompts\u3002\u4f18\u70b9\u662f\u51cf\u5c11\u566a\u58f0\u548c\u4e13\u5bb6\u98ce\u683c\u4e92\u76f8\u6c61\u67d3\uff1b\u7f3a\u70b9\u662f\u8bad\u7ec3\u6d41\u7a0b\u8f83\u590d\u6742\uff0cPhase 1\/Phase 2 \u7684\u51bb\u7ed3\u7b56\u7565\u53ef\u80fd\u5f71\u54cd\u7aef\u5230\u7aef\u6700\u4f18\u6027\u3002<\/p>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p>\u4e3b\u5b9e\u9a8c\u6570\u636e\u96c6\uff1a<\/p>\n<ul>\n<li><strong>LIDC-IDRI\uff1a<\/strong> \u80ba\u7ed3\u8282 CT\uff0c\u591a\u8fbe 4 \u4f4d radiologists \u6807\u6ce8\uff1b\u4f5c\u8005\u63d0\u53d6 214 patients \u7684 1,609 axial slices\uff0c128\u00d7128 nodule-centered patches\uff0cpatient-level 4-fold cross-validation\u3002<\/li>\n<li><strong>NPC-170\uff1a<\/strong> 170 \u4f4d\u9f3b\u54bd\u764c\u60a3\u8005\uff0c\u591a\u6a21\u6001 MRI\uff08T1\/T2\/T1c\uff09\uff0c4 \u4f4d radiation oncologists \u6807\u6ce8 GTVp\uff1b100\/20\/50 train\/val\/test split\u3002<\/li>\n<\/ul>\n<p>\u6307\u6807\uff1aGED\u3001Dicesoft\u3001Dicemax\u3001Dicematch\u3001per-rater DiceA(i)\u3001Dicemean\u3001ECE\u3001Brier\u3001robustness under noise\u3002<\/p>\n<p>\u4e3b\u8981\u7ed3\u679c\uff1a<br \/>\n- Distribution fitting\uff1a\u5728 LIDC-IDRI\uff0cHarmonizer Network #50 GED=0.1048\uff0cDicesoft=91.81\uff0c\u9ad8\u4e8e D-Persona #50 GED=0.1358\uff0cDicesoft=90.45\u3002NPC-170 \u4e0a Harmonizer #50 GED=0.1758\uff0cDicesoft=84.83\uff0c\u9ad8\u4e8e D-Persona #50 GED=0.1978\uff0cDicesoft=84.01\u3002<br \/>\n- Personalized segmentation\uff1aLIDC-IDRI \u4e0a Dicemean=90.78\uff0c\u4f18\u4e8e D-Persona 89.17\uff1bNPC-170 \u4e0a Dicemean=81.63\uff0c\u4f18\u4e8e D-Persona 80.40\u3002<br \/>\n- Calibration\uff1aLIDC per-rater ECE \u7ea6 0.003\u20130.005\uff0cBrier \u7ea6 0.003\u20130.005\u3002<br \/>\n- Noise robustness\uff1a\u5728 LIDC \u5f3a Gaussian noise $\\sigma=0.25$ \u4e0b Harmonizer DSC=84.27\uff0cDice drop=6.53\uff1bD-Persona DSC=71.11\uff0cdrop=18.06\uff1bProb. U-Net DSC=73.22\uff0cdrop=15.87\u3002<br \/>\n- Acquisition domain shift\uff1aLIDC scanner manufacturer split \u4e2d All except Siemens\u2192Siemens\uff0cHarmonizer DSC=85.30\uff0cdrop=5.48\uff1bD-Persona DSC=83.02\uff0cdrop=6.15\u3002<br \/>\n- Kvasir supplement\uff1a\u5728 simulated noisy polyp masks \u4e0b\uff0cSR\/SE Dice \u5206\u522b\u4e3a 85.13\/82.96\uff0c\u4f18\u4e8e D-Persona 84.69\/81.77\uff1bSDE \u4e0a\u7565\u4f4e\u4e8e D-Persona 78.89 vs 78.93\u3002<\/p>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u53ef\u4fe1\u70b9\uff1a<br \/>\n- \u4efb\u52a1\u8bbe\u5b9a\u5f88\u6709\u73b0\u5b9e\u610f\u4e49\uff1a\u591a\u4e13\u5bb6\u6807\u6ce8\u5dee\u5f02\u662f\u533b\u5b66\u5206\u5272\u90e8\u7f72\u4e2d\u7684\u771f\u5b9e\u95ee\u9898\u3002<br \/>\n- \u540c\u65f6\u62a5\u544a distributional metrics\u3001personalized metrics\u3001calibration\u3001noise robustness\u3001domain shift\uff0c\u6bd4\u53ea\u62a5 Dice \u66f4\u5168\u9762\u3002<br \/>\n- LIDC-IDRI \u548c NPC-170 \u5206\u522b\u8986\u76d6 CT lung nodule \u4e0e MRI tumor GTV\uff0c\u591a\u6a21\u6001\/\u591a\u573a\u666f\u652f\u6491\u6bd4\u5355\u6570\u636e\u96c6\u5f3a\u3002<br \/>\n- supplement \u4e2d\u8865\u5145\u4e86 Kvasir noisy polyp\u3001size-stratified robustness\u3001uncertainty vs rater agreement\u3001frequency visualization\u3001complexity \u7b49\u8bc1\u636e\u3002<\/p>\n<p>\u9700\u8981\u8c28\u614e\u7684\u70b9\uff1a<br \/>\n- \u4e3b\u6587\u5199 \u201cGitHub code\u201d\uff0c\u4f46\u6211\u672a\u80fd\u5728\u8bba\u6587\u9875\u9762\u6216\u6b63\u6587\u4e2d\u786e\u8ba4\u5177\u4f53\u4ee3\u7801 URL\uff1b\u53ef\u590d\u73b0\u6027\u6682\u65f6\u4e0d\u786e\u5b9a\u3002<br \/>\n- \u90e8\u5206\u91cd\u8981\u6d88\u878d\u5206\u6563\u5728 supplement\uff0c\u4e3b\u6587\u5bf9\u6bcf\u4e2a\u6a21\u5757\u53bb\u9664\u540e\u7684\u76f4\u63a5\u91cf\u5316\u8d21\u732e\u4e0d\u591f\u96c6\u4e2d\uff1b\u4f8b\u5982 Noise Harmonizer \u4e0e Personalizer \u7684\u72ec\u7acb\u589e\u76ca\u6700\u597d\u6709\u7edf\u4e00 ablation table\u3002<br \/>\n- \u4e24\u9636\u6bb5\u8bad\u7ec3\u51bb\u7ed3 backbone \u53ef\u80fd\u4f7f\u6027\u80fd\u4f9d\u8d56\u8bad\u7ec3 schedule\uff1b\u4e0d\u540c\u4efb\u52a1\u4e0a\u662f\u5426\u9700\u8981\u91cd\u65b0\u8c03\u53c2\u5c1a\u4e0d\u6e05\u695a\u3002<br \/>\n- \u9ad8\u9891 prompt \u7684\u89e3\u91ca\u5f88\u5408\u7406\uff0c\u4f46\u4e13\u5bb6\u5dee\u5f02\u4e0d\u4e00\u5b9a\u90fd\u5728\u9ad8\u9891\uff1b\u5bf9\u201c\u6807\u54ea\u4e2a\u7ed3\u6784\/\u662f\u5426\u5305\u542b\u90bb\u8fd1\u53ef\u7591\u533a\u57df\u201d\u8fd9\u79cd\u8bed\u4e49\u7ea7\u5206\u6b67\uff0c\u6a21\u578b\u4ecd\u53ef\u80fd\u96be\u4ee5\u5904\u7406\u3002\u4f5c\u8005\u4e5f\u5728 limitations \u4e2d\u627f\u8ba4\uff1a\u5f53\u4e00\u4e2a annotator \u6807\u5f97\u8fdc\u5927\u4e8e\u5176\u4ed6\u4eba\uff0c\u6216\u4e24\u4e2a\u5bf9\u79f0\u7ed3\u6784\u90fd\u53ef\u80fd\u662f\u76ee\u6807\u65f6\uff0c\u4e2a\u6027\u5316\u89e3\u7801\u4f1a\u51fa\u73b0\u622a\u65ad\u6216\u65e0\u6cd5\u6d88\u6b67\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 \/ Probabilistic U-Net<\/strong>\uff1a\u672c\u6587\u76f4\u63a5\u4ee5 Probabilistic U-Net \u4e3a base\uff0c\u6539\u9020\u7684\u662f latent distribution\u3001decoder feature harmonization \u548c rater personalization\u3002<\/li>\n<li>\u4e0e <strong>nnU-Net<\/strong>\uff1annU-Net \u504f\u81ea\u52a8\u914d\u7f6e\u7684 deterministic supervised segmentation\uff1b\u672c\u6587\u5173\u6ce8 multi-rater uncertainty \u548c personalized outputs\uff0c\u53ef\u4f5c\u4e3a nnU-Net \u7c7b\u6a21\u578b\u4e4b\u5916\u7684\u4e0d\u786e\u5b9a\u6027\u5efa\u6a21\u8865\u5145\u3002<\/li>\n<li>\u4e0e <strong>MedNeXt \/ Transformer \/ Swin-UNet \/ UNetR<\/strong>\uff1a\u8fd9\u4e9b\u4e3b\u8981\u6539 backbone\uff1b\u672c\u6587 backbone \u4e0d\u662f\u91cd\u70b9\uff0cNoise Harmonizer \/ Personalizer \u7406\u8bba\u4e0a\u53ef\u63d2\u5165\u5230\u4e0d\u540c encoder-decoder \u67b6\u6784\u4e2d\u3002<\/li>\n<li>\u4e0e <strong>Mamba \/ DAMamba<\/strong>\uff1a\u6ca1\u6709\u4f7f\u7528 Mamba\uff0c\u4f46 frequency prompt \u7684\u601d\u60f3\u53ef\u4e0e Mamba \u7684 selective scan \u7ed3\u5408\uff1a\u4f8b\u5982\u5728 DAMamba \u4e2d\u52a0\u5165\u9ad8\u9891\u8fb9\u754c prompt branch \u6216 uncertainty-conditioned scan\u3002<\/li>\n<li>\u4e0e <strong>foundation model \/ MedSAM<\/strong>\uff1a\u672c\u6587\u4e0d\u5c5e\u4e8e foundation model\uff1b\u4f46 multi-rater personalization \u5bf9 MedSAM\/interactive segmentation \u5f88\u91cd\u8981\uff0c\u56e0\u4e3a\u7528\u6237\u63d0\u793a\u548c\u4e13\u5bb6\u98ce\u683c\u5dee\u5f02\u672c\u8d28\u4e0a\u76f8\u5173\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u8fd9\u7bc7\u8bba\u6587\u5bf9\u4f60\u7684 polyp segmentation \/ DAMamba \/ \u533b\u5b66\u5206\u5272\u6846\u67b6\u7814\u7a76\u6709\u660e\u663e\u53c2\u8003\u4ef7\u503c\uff1a<\/p>\n<ol>\n<li><strong>Polyp segmentation\uff1a<\/strong> supplement \u7684 Kvasir noisy-label \u5b9e\u9a8c\u76f4\u63a5\u76f8\u5173\u3002\u5b83\u8bf4\u660e\u5728\u606f\u8089\u4efb\u52a1\u4e2d\uff0c\u8fb9\u754c\u566a\u58f0\u548c\u4e13\u5bb6\u5dee\u5f02\u53ef\u4ee5\u901a\u8fc7 harmonization + frequency prompt \u6539\u5584\uff0c\u5c24\u5176\u662f SR\/SE noisy masks \u4e0b\u8868\u73b0\u8f83\u597d\u3002<\/li>\n<li><strong>DAMamba \u6539\u9020\uff1a<\/strong> \u53ef\u5c1d\u8bd5\u628a Mamba branch \u5206\u6210 shared anatomy branch \u4e0e rater\/boundary prompt branch\uff1b\u6216\u8005\u5728 decoder \u672b\u7aef\u52a0\u5165 DWT \u9ad8\u9891\u8fb9\u754c prompt \u6765\u589e\u5f3a\u4e2a\u6027\u5316\/\u8fb9\u754c\u9c81\u68d2\u6027\u3002<\/li>\n<li><strong>\u8bba\u6587\u5199\u4f5c\u4ef7\u503c\uff1a<\/strong> \u5982\u679c\u4f60\u7684\u8bba\u6587\u6d89\u53ca\u4e0d\u786e\u5b9a\u6027\u3001\u6807\u6ce8\u566a\u58f0\u3001\u591a\u4e13\u5bb6\u5dee\u5f02\uff0c\u8fd9\u7bc7\u53ef\u4f5c\u4e3a related work \u4e2d\u8fde\u63a5 Probabilistic U-Net\u3001D-Persona\u3001DiffOSeg\u3001multi-rater calibration \u7684\u8fd1\u671f\u53c2\u8003\u3002<\/li>\n<li><strong>\u5b9e\u9a8c\u8bbe\u8ba1\u4ef7\u503c\uff1a<\/strong> \u5b83\u7684\u6307\u6807\u4f53\u7cfb\u5f88\u503c\u5f97\u501f\u9274\uff1a\u4e0d\u8981\u53ea\u62a5 Dice\uff0c\u53ef\u4ee5\u52a0\u5165 GED\u3001Dicematch\u3001rater agreement\u3001ECE\/Brier\u3001noise perturbation \u548c scanner manufacturer split\u3002<\/li>\n<\/ol>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5efa\u8bae\u7cbe\u8bfb\uff0c\u4f46\u4f18\u5148\u7ea7\u7565\u4f4e\u4e8e SGP-Net\u3002<\/strong> \u5982\u679c\u4f60\u5f53\u524d\u91cd\u70b9\u662f\u901a\u7528\/\u606f\u8089\u5206\u5272\u6846\u67b6\u3001\u8fb9\u754c\u6a21\u5757\u3001\u6807\u6ce8\u566a\u58f0\u6216 uncertainty-aware segmentation\uff0c\u8fd9\u7bc7\u503c\u5f97\u8bfb Section 3\u3001Table 1\u20133\u3001supplement \u7684 noise\/Kvasir\/domain shift \u90e8\u5206\u3002\u82e5\u4f60\u7684\u77ed\u671f\u76ee\u6807\u53ea\u662f\u6784\u5efa deterministic polyp SOTA\uff0c\u5219\u53ef\u5148\u7565\u8bfb\u4e3b\u6587\uff0c\u628a\u5b83\u4f5c\u4e3a\u201c\u591a\u6807\u6ce8\u4e0d\u786e\u5b9a\u6027\u4e0e\u566a\u58f0\u9c81\u68d2\u201d\u65b9\u5411\u50a8\u5907\u3002<\/p>\n<hr \/>\n<h2>\u4eca\u65e5\u63a8\u8350\u4f18\u5148\u7ea7<\/h2>\n<ol>\n<li><strong>Beyond Euclidean Prototypes \/ SGP-Net<\/strong>\uff1a\u6700\u503c\u5f97\u5148\u8bfb\u3002\u673a\u5236\u6e05\u695a\u3001\u6a21\u5757\u53ef\u590d\u7528\uff0cSPB + GM \u5bf9\u4f4e\u6807\u6ce8\u533b\u5b66\u5206\u5272\u3001support-query matching\u3001\u8fb9\u754c\u6cc4\u6f0f\u95ee\u9898\u6709\u76f4\u63a5\u542f\u53d1\uff1b\u4e5f\u66f4\u5bb9\u6613\u8f6c\u5316\u6210\u65b0\u6a21\u5757\u6216 ablation idea\u3002<\/li>\n<li><strong>Harmonized Feature Conditioning and Frequency-Prompt Personalization<\/strong>\uff1a\u9002\u5408\u5173\u6ce8 multi-rater\u3001uncertainty\u3001noisy label\u3001polyp boundary ambiguity \u7684\u7814\u7a76\u3002\u5b83\u5bf9 deterministic segmentation \u4e3b\u7ebf\u4e0d\u662f\u76f4\u63a5 baseline\uff0c\u4f46\u5bf9\u201c\u5982\u4f55\u5904\u7406\u6807\u6ce8\u5dee\u5f02\u548c\u8fb9\u754c\u4e0d\u786e\u5b9a\u6027\u201d\u5f88\u6709\u4ef7\u503c\u3002<\/li>\n<\/ol>\n<h2>\u4eca\u65e5 PDF \u83b7\u53d6\u60c5\u51b5<\/h2>\n<ul>\n<li>\u8bba\u6587 1\uff1a\u5df2\u9644 PDF\uff1b\u672c\u5730\u6587\u4ef6 PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.17904<\/li>\n<li>\u8bba\u6587 2\uff1a\u5df2\u9644 PDF\uff1b\u672c\u5730\u6587\u4ef6 PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.08210<\/li>\n<\/ul>\n<h2>\u4eca\u65e5\u53ef\u6267\u884c\u5efa\u8bae<\/h2>\n<ol>\n<li>\u5148\u7cbe\u8bfb\u5e76\u590d\u73b0 <strong>SGP-Net \u7684 SPB + GM \u601d\u8def<\/strong>\uff1a\u5373\u4f7f\u4e0d\u505a few-shot\uff0c\u4e5f\u53ef\u4ee5\u628a low\/mid\/high frequency branch \u548c boundary-aware geodesic\/diffusion refinement \u6539\u9020\u6210 polyp segmentation decoder \u6a21\u5757\u3002<\/li>\n<li>\u5982\u679c\u4f60\u5728\u505a <strong>DAMamba \/ Mamba-based medical segmentation<\/strong>\uff0c\u5efa\u8bae\u5c1d\u8bd5\u4e00\u4e2a\u5c0f\u6539\u9020\uff1aMamba backbone \u8f93\u51fa\u540e\uff0c\u589e\u52a0 DWT\/FFT high-frequency boundary branch\uff0c\u518d\u7528\u5c40\u90e8 affinity diffusion \u6216 selective scan refinement \u4fee\u6b63\u8fb9\u754c\u6cc4\u6f0f\u3002<\/li>\n<li>\u5982\u679c\u8bba\u6587\u5b9e\u9a8c\u6d89\u53ca noisy masks \u6216\u591a\u4e13\u5bb6\u6807\u6ce8\uff0c\u5efa\u8bae\u628a\u7b2c\u4e8c\u7bc7\u7684 GED\u3001Dicematch\u3001ECE\/Brier\u3001noise perturbation\u3001scanner split \u4f5c\u4e3a\u8bc4\u4f30\u8bbe\u8ba1\u53c2\u8003\uff1b\u5b83\u6bd4\u5355\u7eaf Dice \u66f4\u80fd\u8bf4\u660e\u6a21\u578b\u662f\u5426\u771f\u7684\u53ef\u4fe1\u3002<\/li>\n<\/ol>\n<h2>\u53c2\u8003\u94fe\u63a5<\/h2>\n<ul>\n<li>SGP-Net arXiv\uff1ahttps:\/\/arxiv.org\/abs\/2605.17904<\/li>\n<li>SGP-Net PDF\uff1ahttps:\/\/arxiv.org\/pdf\/2605.17904<\/li>\n<li>SGP-Net code\uff1ahttps:\/\/github.com\/naivejph\/SGP-Net.git<\/li>\n<li>Harmonizer Network arXiv\uff1ahttps:\/\/arxiv.org\/abs\/2605.08210<\/li>\n<li>Harmonizer Network PDF\uff1ahttps:\/\/arxiv.org\/pdf\/2605.08210<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u4eca\u65e5\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6700\u65b0\u8bba\u6587\u7cbe\u8bfb\u8ffd\u8e2a \u4eca\u65e5\u7ed3\u8bba \u4eca\u5929\u68c0\u7d22\u5230\u7684\u6700\u65b0\u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\u4ecd\u4ee5 2026 arXiv preprint \/ CV &#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-1070","post","type-post","status-publish","format-standard","hentry","category-85"],"views":212,"_links":{"self":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1070","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=1070"}],"version-history":[{"count":0,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1070\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/media?parent=1070"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/categories?post=1070"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/tags?post=1070"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}