{"id":1047,"date":"2026-05-13T08:35:16","date_gmt":"2026-05-13T00:35:16","guid":{"rendered":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/13\/2026-05-13-%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%9ageoproto-%e4%b8%8e-xtinyu-net\/"},"modified":"2026-05-13T08:35:16","modified_gmt":"2026-05-13T00:35:16","slug":"2026-05-13-%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%9ageoproto-%e4%b8%8e-xtinyu-net","status":"publish","type":"post","link":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/13\/2026-05-13-%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%9ageoproto-%e4%b8%8e-xtinyu-net\/","title":{"rendered":"2026-05-13 \u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\u7cbe\u8bfb\uff1aGeoProto \u4e0e XTinyU-Net"},"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 arXiv \u5728 2026-05-10 \u5230 2026-05-11 \u65b0\u589e\u4e86\u591a\u7bc7\u533b\u5b66\u56fe\u50cf\u5206\u5272\u76f8\u5173\u9884\u5370\u672c\uff1b\u53bb\u91cd\u540e\uff0c\u6700\u503c\u5f97\u5173\u6ce8\u7684\u662f\u4e00\u7bc7\u8de8\u57df few-shot segmentation \u65b9\u6cd5\u8bba\u6587 GeoProto\uff0c\u4ee5\u53ca\u4e00\u7bc7\u975e\u5e38\u8d34\u8fd1 U-Net \/ nnU-Net \u5de5\u7a0b\u5b9e\u8df5\u7684\u8f7b\u91cf\u5316\u9009\u62e9\u8bba\u6587 XTinyU-Net\u3002\u6574\u4f53\u8d8b\u52bf\u662f\uff1a\u4e00\u65b9\u9762\u7ee7\u7eed\u56f4\u7ed5\u8de8\u57df\u6cdb\u5316\u3001\u5c11\u6807\u6ce8\u548c foundation model \u9002\u914d\u5c55\u5f00\uff0c\u53e6\u4e00\u65b9\u9762\u4e5f\u5f00\u59cb\u91cd\u65b0\u5ba1\u89c6\u201c\u6807\u51c6 U-Net\/nnU-Net \u7ecf\u5408\u7406\u7f29\u653e\u662f\u5426\u6bd4\u590d\u6742\u8f7b\u91cf\u6a21\u5757\u66f4\u6709\u6548\u201d\u3002<\/p>\n<h2>\u68c0\u7d22\u8bf4\u660e<\/h2>\n<p>\u4eca\u5929\u4f18\u5148\u68c0\u7d22\u4e86 arXiv 2026-05-01 \u81f3 2026-05-13 \u7684 <code>medical image segmentation<\/code>\u3001<code>segmentation + medical\/CT\/MRI<\/code>\u3001<code>polyp segmentation<\/code>\u3001<code>Mamba + medical image segmentation<\/code>\u3001<code>nnUNet + segmentation<\/code> \u7b49\u67e5\u8be2\uff0c\u5e76\u53c2\u8003\u4e86\u8bba\u6587 PDF \u6b63\u6587\u3001arXiv \u5143\u6570\u636e\u4e0e\u5b98\u65b9\u4ee3\u7801\u94fe\u63a5\u53ef\u8fbe\u6027\u3002\u5f53\u5929\u6ca1\u6709\u68c0\u7d22\u5230\u5df2\u6b63\u5f0f\u6807\u6ce8\u4e3a MICCAI\/CVPR\/ICCV\/ECCV\/NeurIPS\/ICLR\/AAAI\/IJCAI\/ISBI \u6216\u9876\u520a\u63a5\u6536\u7684\u5168\u65b0\u5206\u5272\u8bba\u6587\uff0c\u56e0\u6b64\u4eca\u65e5\u5165\u9009\u5747\u4e3a 2026 \u5e74 arXiv preprint\uff1b\u4e24\u7bc7\u5747\u6ee1\u8db3\u201c2025 \u5e74\u53ca\u4ee5\u540e\u201d\u8981\u6c42\u3002\u5df2\u68c0\u67e5\u5386\u53f2\u63a8\u8350\u8bb0\u5f55\u5e76\u6392\u9664\u4e86\u91cd\u590d\u8bba\u6587\uff1b\u672c\u6b21\u8df3\u8fc7\u7684\u5386\u53f2\u63a8\u8350\u5019\u9009\u5305\u62ec ZScribbleSeg\u3001Topology-Constrained Quantized nnUNet\u3001One Sequence to Segment Them All\u3001Sharpening Lightweight Models for Generalized Polyp Segmentation \u7b49\u3002<\/p>\n<hr \/>\n<h2>\u8bba\u6587 1\uff1aGeometry-aware Prototype Learning for Cross-domain Few-shot Medical Image Segmentation<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aGeometry-aware Prototype Learning for Cross-domain Few-shot Medical Image Segmentation<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aFeifan Song, Yuntian Bo, Haofeng Zhang \/ \u7b2c\u4e00\u4f5c\u8005 Feifan Song<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-11<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint, arXiv:2605.10885v1<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.10885<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.10885v1 \uff08\u5df2\u4e0b\u8f7d\uff1aMEDIA:\/tmp\/medseg_daily_20260513\/geometry_prototype.pdf\uff09<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/FeifanSong\/Geoproto \uff08\u9875\u9762\u53ef\u8bbf\u95ee\uff1b\u8bba\u6587\u6b63\u6587\u5199\u4f5c <code>https:\/\/github.com\/FeifanSong\/Geoproto.git<\/code>\uff09<\/li>\n<li>\u4efb\u52a1\uff1aCross-domain few-shot medical image segmentation\uff1b1-way 1-shot \u8bbe\u7f6e\uff1b\u8de8\u6a21\u6001\u3001\u8de8\u5e8f\u5217\u3001\u8de8\u4e0a\u4e0b\u6587\u5206\u5272<\/li>\n<li>\u6570\u636e\u96c6\uff1aAbdominal CT\uff08MICCAI 2015 Multi-Atlas Labeling Challenge\uff09\u3001Abdominal MRI\uff08ISBI 2019 CHAOS T2-SPIR\uff09\u3001Cardiac b-SSFP \/ LGE\uff08MICCAI 2019 Multi-sequence Cardiac MR Segmentation Challenge\uff09\u3001MI-PMR prostate MRI\u3001Chest X-Ray\u3001ISIC2018<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1aprototype-based few-shot segmentation\uff1bgeometry-aware prototype enrichment\uff1bdomain generalization \/ cross-domain few-shot<\/li>\n<\/ul>\n<h3>paper-deep-reader \u7cbe\u8bfb\u7ed3\u679c<\/h3>\n<h4>1. \u4e00\u53e5\u8bdd\u7ed3\u8bba<\/h4>\n<p>GeoProto \u7684\u6838\u5fc3\u4ef7\u503c\u5728\u4e8e\u628a\u201c\u5668\u5b98\u4ece\u8fb9\u754c\u5230\u4e2d\u5fc3\u7684 ordinal geometry\u201d\u663e\u5f0f\u6ce8\u5165 few-shot prototype matching\uff0c\u4f7f\u8de8\u6a21\u6001\/\u8de8\u5e8f\u5217\u65f6\u7684 support-query \u5339\u914d\u4e0d\u53ea\u4f9d\u8d56\u5bb9\u6613\u53d7\u57df\u504f\u79fb\u5f71\u54cd\u7684\u5916\u89c2\u7279\u5f81\uff0c\u800c\u591a\u4e86\u4e00\u4e2a\u6765\u81ea mask\u3001\u65e0\u9700\u989d\u5916\u6807\u6ce8\u7684\u7ed3\u6784\u951a\u70b9\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u8bba\u6587\u7814\u7a76 cross-domain few-shot medical image segmentation\uff08CD-FSMIS\uff09\uff1a\u6a21\u578b\u5728\u6e90\u57df base categories \u4e0a episodic \u8bad\u7ec3\uff0c\u63a8\u7406\u65f6\u8981\u5728\u672a\u89c1\u8fc7\u7684\u76ee\u6807\u57df\u548c novel categories \u4e0a\uff0c\u4ec5\u7528\u5c11\u91cf support mask \u5206\u5272 query\u3002\u8fd9\u4e2a\u95ee\u9898\u6bd4\u666e\u901a\u533b\u5b66\u56fe\u50cf\u5206\u5272\u66f4\u8d34\u8fd1\u4e34\u5e8a\u73b0\u5b9e\uff1a\u4e0d\u540c\u533b\u9662\u3001\u626b\u63cf\u4eea\u3001\u6a21\u6001\u3001\u5e8f\u5217\u548c\u89e3\u5256\u533a\u57df\u5e26\u6765\u5f3a domain shift\uff0c\u800c\u9010\u4efb\u52a1\u91cd\u65b0\u6807\u6ce8\u548c\u8bad\u7ec3\u4ee3\u4ef7\u5f88\u9ad8\u3002\u8bba\u6587\u7684 paper map \u53ef\u6982\u62ec\u4e3a\uff1a\u5b83\u7814\u7a76 CD-FSMIS\uff1b\u4e3b\u8981\u52a8\u4f5c\u662f\u4ece support mask \u6784\u9020 distance-to-boundary ordinal strata\uff0c\u5e76\u901a\u8fc7 Ordinal Shape Branch \u4e0e Geometry-Aware Prototype Enrichment \u7ed9 local prototypes \u52a0\u51e0\u4f55\u504f\u79fb\uff1b\u4f5c\u8005\u58f0\u79f0\u8fd9\u80fd\u5728\u8de8\u6a21\u6001\u3001\u8de8\u5e8f\u5217\u3001\u8de8\u4e0a\u4e0b\u6587 setting \u4e2d\u8d85\u8fc7 RobustEMD\u3001FAMNet\u3001C-Graph \u7b49\u65b9\u6cd5\uff1b\u8bc1\u636e\u4e3b\u8981\u6765\u81ea 7 \u4e2a\u6570\u636e\u96c6\u30013 \u7c7b\u8de8\u57df\u8bc4\u4f30\u548c\u6d88\u878d\uff1b\u5173\u952e\u98ce\u9669\u662f\u201c\u51e0\u4f55\u7ed3\u6784\u8de8\u4e2a\u4f53\/\u8de8\u75c5\u7076\u7a33\u5b9a\u201d\u8fd9\u4e2a\u5047\u8bbe\u5728\u5f62\u6001\u53d8\u5f02\u5f3a\u3001\u75c5\u7076\u8fb9\u754c\u4e0d\u89c4\u5219\u6216\u975e\u5668\u5b98\u76ee\u6807\u4e0a\u53ef\u80fd\u5931\u6548\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u8ba4\u4e3a\u73b0\u6709 FSMIS \/ CD-FSMIS \u65b9\u6cd5\u7684\u95ee\u9898\u4e0d\u662f\u6ca1\u6709 prototype\uff0c\u800c\u662f prototype \u4e3b\u8981\u7531 encoder \u7279\u5f81\u5e73\u5747\u6216\u5c40\u90e8\u805a\u5408\u5f97\u5230\uff0c\u5bb9\u6613\u628a anatomy structure \u548c domain-specific appearance \u6df7\u5728\u4e00\u8d77\u3002RobustEMD\u3001FAMNet\u3001C-Graph \u7b49\u8de8\u57df\u65b9\u6cd5\u901a\u8fc7\u7eb9\u7406\u6291\u5236\u3001\u9891\u5e26\u5339\u914d\u6216 prototype graph \u51cf\u8f7b domain shift\uff0c\u4f46\u8bba\u6587\u6307\u51fa\u8fd9\u7c7b\u201c\u6291\u5236\u57df\u7279\u5f81\u201d\u7684\u601d\u8def\u53ef\u80fd\u540c\u65f6\u635f\u5931\u5173\u952e\u7ed3\u6784\u4fe1\u606f\u3002GeoProto \u7684\u66ff\u4ee3\u89c6\u89d2\u662f\uff1a\u4e0d\u8981\u53ea\u538b\u5236\u57df\u5dee\u5f02\uff0c\u800c\u8981\u5f15\u5165\u4e00\u4e2a\u8de8\u57df\u66f4\u7a33\u5b9a\u7684\u5339\u914d\u5750\u6807\u7cfb\uff0c\u5373 organ boundary-to-centroid \u7684\u51e0\u4f55\u5c42\u7ea7\u3002<\/p>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>\u8bba\u6587\u8def\u7ebf\u4e3a method-algorithm\uff0c\u8bc1\u636e\u91cd\u70b9\u662f experimental-eval\u3001ablation-and-mechanism-isolation\u3001robustness-and-ood\u3002\u6574\u4f53\u6846\u67b6\u5982\u4e0b\uff1a<\/p>\n<ol>\n<li>\u4f7f\u7528\u5171\u4eab ResNet-101 backbone \u7f16\u7801 support image \u548c query image\uff0c\u5f97\u5230\u7279\u5f81 <code>F^s, F^q \u2208 R^{C\u00d7h\u00d7w}<\/code>\u3002<\/li>\n<li>\u5bf9 support mask \u8ba1\u7b97 Euclidean distance-to-boundary transform\uff08EDT\uff09\uff0c\u5e76\u628a\u524d\u666f\u50cf\u7d20\u4ece\u8fb9\u754c\u5230\u4e2d\u5fc3\u5747\u5300\u91cf\u5316\u4e3a <code>K<\/code> \u4e2a ordinal bins\uff1bbin 0 \u8fd1\u8fb9\u754c\uff0cbin <code>K-1<\/code> \u8fd1\u4e2d\u5fc3\u3002<\/li>\n<li>Ordinal Shape Branch\uff08OSB\uff09\u4ece support feature \u9884\u6d4b\u6bcf\u4e2a\u524d\u666f\u50cf\u7d20\u7684 bin distribution\uff0c\u5e76\u7528\u524d\u666f cross-entropy \u52a0 ordinal distance penalty \u8bad\u7ec3\uff0c\u4f7f\u9884\u6d4b\u7684\u51e0\u4f55 embedding \u968f\u8ddd\u79bb\u5c42\u7ea7\u5e73\u6ed1\u3001\u5355\u8c03\u53d8\u5316\u3002<\/li>\n<li>Geometry-Aware Prototype Enrichment\uff08GAPE\uff09\u628a support feature map \u5212\u6210 <code>G\u00d7G<\/code> \u5c40\u90e8\u7f51\u683c\uff0c\u6bcf\u4e2a\u6709\u6548\u524d\u666f cell \u505a masked-average pooling \u5f97\u5230 local appearance prototype\u3002<\/li>\n<li>\u5bf9\u6bcf\u4e2a cell \u8ba1\u7b97\u5176\u671f\u671b ordinal stratum <code>\\bar d_i = mean_x \u03a3_k k p_k(x)<\/code>\uff0c\u518d\u7ecf\u4e24\u5c42 MLP <code>\u03d5(\\bar d_i\/(K-1))<\/code> \u5f97\u5230 <code>C<\/code> \u7ef4\u51e0\u4f55 offset <code>e_i<\/code>\u3002<\/li>\n<li>enriched prototype \u4e3a <code>p'_i = p_i^fg + e_i<\/code>\uff1b\u80cc\u666f prototype \u4e0d\u505a\u51e0\u4f55\u589e\u5f3a\uff0c\u56e0\u4e3a\u80cc\u666f\u6ca1\u6709 organ-interior geometry\u3002<\/li>\n<li>query pixel \u7528 softmax-weighted cosine similarity \u4e0e enriched foreground prototypes\u3001background prototypes \u5339\u914d\uff0c\u8f93\u51fa segmentation mask\u3002<\/li>\n<li>\u603b loss \u5305\u62ec query segmentation CE\u3001bidirectional prototype alignment loss\uff0c\u4ee5\u53ca <code>\u03bb_geo L_OSB<\/code>\u3002<\/li>\n<\/ol>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<ul>\n<li><strong>EDT ordinal bin construction<\/strong>\uff1a\u8f93\u5165\u662f support binary mask\uff0c\u8f93\u51fa\u662f\u6bcf\u4e2a\u524d\u666f\u50cf\u7d20\u7684\u79bb\u6563\u51e0\u4f55\u5c42\u7ea7 <code>Z \u2208 {0,\u2026,K-1}<\/code>\u3002\u5b83\u89e3\u51b3\u7684\u662f prototype \u4e0d\u77e5\u9053\u201c\u6765\u81ea\u5668\u5b98\u8fb9\u7f18\u8fd8\u662f\u4e2d\u5fc3\u201d\u7684\u95ee\u9898\u3002\u521b\u65b0\u4e0d\u5728 EDT \u672c\u8eab\uff0c\u800c\u5728\u628a EDT \u53d8\u6210 few-shot prototype \u7684\u663e\u5f0f\u7ed3\u6784\u6761\u4ef6\u3002<\/li>\n<li><strong>Ordinal Shape Branch\uff08OSB\uff09<\/strong>\uff1a\u8f93\u5165 support feature\uff0c\u8f93\u51fa K-bin logits \/ distribution\u3002<code>L_cls<\/code> \u53ea\u770b bin \u5206\u7c7b\u662f\u5426\u6b63\u786e\uff0c<code>L_dist<\/code> \u6309\u9884\u6d4b bin \u4e0e\u771f\u5b9e bin \u7684 ordinal gap \u52a0\u6743\uff0c\u9f13\u52b1\u76f8\u90bb\u5c42\u7ea7\u7684\u5e73\u6ed1\u8fc7\u6e21\u3002\u8fd9\u4e2a\u6a21\u5757\u7684\u521b\u65b0\u70b9\u8f83\u660e\u786e\uff0c\u5c24\u5176\u662f\u628a ordinal consistency \u7528\u4f5c\u7ed3\u6784\u76d1\u7763\uff1b\u4f46\u5b83\u4f9d\u8d56 support mask \u8d28\u91cf\uff0cmask \u566a\u58f0\u6216\u8fb9\u754c\u7c97\u7cd9\u4f1a\u76f4\u63a5\u5f71\u54cd\u51e0\u4f55\u76d1\u7763\u3002<\/li>\n<li><strong>Geometry-Aware Prototype Enrichment\uff08GAPE\uff09<\/strong>\uff1a\u8f93\u5165 local appearance prototype \u548c\u8be5 cell \u7684 expected stratum\uff0c\u8f93\u51fa enriched prototype\u3002\u5b83\u901a\u8fc7\u52a0\u6cd5 offset \u6539\u53d8 prototype feature space\uff0c\u4f7f\u5916\u89c2\u76f8\u4f3c\u4f46\u51e0\u4f55\u4f4d\u7f6e\u4e0d\u540c\u7684\u5c40\u90e8 prototype \u88ab\u533a\u5206\u5f00\u3002\u5bf9 U-Net \/ nnU-Net \u4e3b\u5e72\u4e0d\u662f\u5373\u63d2\u5373\u7528\uff0c\u56e0\u4e3a\u5b83\u670d\u52a1\u4e8e prototype matching\uff1b\u4f46\u5bf9 few-shot\u3001prompt-based \u6216 memory-prototype segmentation \u6846\u67b6\u5f88\u6709\u8fc1\u79fb\u4ef7\u503c\u3002<\/li>\n<li><strong>Bidirectional alignment regularizer<\/strong>\uff1a\u7528 query \u9884\u6d4b mask \u53cd\u8fc7\u6765 segment support\uff0c\u7a33\u5b9a feature space\u3002\u4f5c\u8005\u523b\u610f\u4e0d\u5728 alignment \u5206\u652f\u7528 geometry enrichment\uff0c\u4ee5\u907f\u514d GAPE \u4e0e alignment \u7684\u5faa\u73af\u4f9d\u8d56\u3002<\/li>\n<li><strong>\u662f\u5426\u9002\u5408 polyp segmentation \/ 3D segmentation<\/strong>\uff1a\u5bf9\u666e\u901a supervised polyp segmentation \u7684\u76f4\u63a5\u4ef7\u503c\u6709\u9650\uff0c\u56e0\u4e3a\u606f\u8089\u5f62\u6001\u3001\u906e\u6321\u3001\u6241\u5e73\u75c5\u7076\u548c\u8fb9\u754c\u4e0d\u89c4\u5219\u4f1a\u524a\u5f31\u201c\u540c\u7c7b\u76ee\u6807\u6709\u7a33\u5b9a interior strata\u201d\u7684\u5047\u8bbe\uff1b\u4f46\u5bf9 few-shot polyp\u3001\u8de8\u4e2d\u5fc3 colonoscopy domain shift \u6216 prototype-based polyp adaptation \u53ef\u501f\u9274\u30023D segmentation \u4e2d EDT \u53ef\u4ee5\u81ea\u7136\u63a8\u5e7f\u5230 3D distance transform\uff0c\u4f46\u8ba1\u7b97\u548c memory-prototype \u8bbe\u8ba1\u9700\u91cd\u505a\u3002<\/li>\n<\/ul>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p>\u5b9e\u9a8c\u8986\u76d6\u4e09\u7c7b\u8bbe\u7f6e\uff1a<\/p>\n<ul>\n<li><strong>Cross-modality<\/strong>\uff1aAbdominal CT \u2194 MRI\uff0c\u5668\u5b98\u5305\u62ec liver\u3001left kidney\u3001right kidney\u3001spleen\u3002GeoProto \u5728 CT\u2192MRI mean DSC 69.45%\uff0cMRI\u2192CT mean DSC 69.65%\uff1b\u76f8\u8f83 C-Graph \u5728 CT\u2192MRI \u7684 67.38% \u6709\u5c0f\u5e45\u63d0\u5347\uff0c\u5728 MRI\u2192CT \u76f8\u8f83 C-Graph 67.20% \u4e5f\u6709\u63d0\u5347\u3002<\/li>\n<li><strong>Cross-sequence<\/strong>\uff1aCardiac LGE \u2194 b-SSFP\u3002GeoProto \u5728 LGE\u2192b-SSFP mean DSC 74.23%\uff0c\u4e0e C-Graph 74.20% \u57fa\u672c\u6301\u5e73\uff1b\u5728 b-SSFP\u2192LGE mean DSC 67.10%\uff0c\u9ad8\u4e8e C-Graph 63.33% \u548c FAMNet 61.39%\u3002<\/li>\n<li><strong>Cross-context<\/strong>\uff1a\u8bad\u7ec3\u4e8e Abdominal CT\uff0c\u6d4b\u8bd5\u5230 CXR\u3001ISIC\u3001Cardiac b-SSFP MRI\u3001MI-PMR\u3002GeoProto \u5728 CXR\u3001Cardiac b-SSFP\u3001MI-PMR \u4e0a\u8868\u73b0\u8f83\u5f3a\uff0c\u4f8b\u5982 MI-PMR mean DSC 61.24%\uff0c\u9ad8\u4e8e C-Graph 50.10%\uff1b\u4f46\u5728 ISIC \u4e0a 47.45% \u4f4e\u4e8e C-Graph 50.42%\u3002<\/li>\n<\/ul>\n<p>\u6d88\u878d\u65b9\u9762\uff0c\u4f5c\u8005\u6bd4\u8f83 baseline\u3001\u4f4d\u7f6e\u7f16\u7801 PE\u3001Geo-E without OSB-L\u3001OSB-L without Geo-E \u548c\u5b8c\u6574\u6a21\u578b\u3002\u5b8c\u6574\u6a21\u578b\u5728 Abdominal MRI \u4e0a 69.45%\uff0c\u6bd4 baseline 63.01% \u63d0\u5347 6.44\uff1b\u5728 Cardiac b-SSFP \u4e0a 61.03%\uff0c\u6bd4 baseline 45.59% \u63d0\u5347 15.44\u3002K-bin \u6d88\u878d\u663e\u793a K=10 \u7efc\u5408\u6700\u597d\uff1bfusion mode \u4e2d additive fusion \u6574\u4f53\u4f18\u4e8e concat projection \u548c scale gate\u3002\u6548\u7387\u8868\u663e\u793a GeoProto 43.4M \u53c2\u6570\uff0ctest latency 12.87ms\uff0c\u4f4e\u4e8e FAMNet 22.72ms \u548c C-Graph 147.39ms\u3002<\/p>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u53ef\u4fe1\u4e4b\u5904\uff1a\u6570\u636e\u96c6\u548c transfer settings \u8986\u76d6\u9762\u8f83\u5e7f\uff1bbaseline \u5305\u62ec RobustEMD\u3001FAMNet\u3001C-Graph \u7b49 CD-FSMIS \u76f8\u5173\u65b9\u6cd5\uff1b\u6d88\u878d\u4e0d\u662f\u53ea\u5220\u4e00\u4e2a\u6a21\u5757\uff0c\u800c\u662f\u9a8c\u8bc1\u4e86 geometry prior\u3001OSB-L\u3001bin \u6570\u91cf\u3001fusion mode \u548c\u590d\u6742\u5ea6\u3002\u5c24\u5176 Fig. 3 \u5bf9 EDT bin \u5206\u5e03\u548c support-query Bhattacharyya coefficient \u7684\u5206\u6790\uff0c\u76f4\u63a5\u56de\u5e94\u4e86\u201c\u51e0\u4f55\u662f\u5426\u7a33\u5b9a\u201d\u7684\u6838\u5fc3\u5047\u8bbe\u3002<\/p>\n<p>\u9700\u8981\u8c28\u614e\u4e4b\u5904\uff1a\u7b2c\u4e00\uff0c\u6240\u6709\u7ed3\u679c\u90fd\u662f 1-way 1-shot\uff0c\u5c1a\u4e0d\u6e05\u695a\u591a shot\u3001\u4e0d\u540c support \u91c7\u6837\u65b9\u5dee\u548c\u66f4\u771f\u5b9e\u4e34\u5e8a episode \u4e0b\u662f\u5426\u7a33\u5b9a\uff1b\u7b2c\u4e8c\uff0c\u4e3b\u8981\u6307\u6807\u662f DSC\uff0c\u7f3a\u5c11\u7edf\u8ba1\u663e\u8457\u6027\u68c0\u9a8c\u548c\u7f6e\u4fe1\u533a\u95f4\uff1b\u7b2c\u4e09\uff0cgeometry prior \u5bf9\u9ad8\u5ea6\u53d8\u5f62\u75c5\u7076\u3001\u672f\u540e\u7ed3\u6784\u3001\u80bf\u7624\u6d78\u6da6\u8fb9\u754c\u3001\u6241\u5e73\u606f\u8089\u7b49\u76ee\u6807\u672a\u88ab\u5145\u5206\u9a8c\u8bc1\uff1b\u7b2c\u56db\uff0cbackbone \u4f7f\u7528 MS-COCO \u521d\u59cb\u5316 ResNet-101\uff0c\u548c\u73b0\u4ee3 MedSAM \/ SAM2 \/ DINO \/ Mamba \u7c7b backbone \u7684\u5173\u7cfb\u8fd8\u6ca1\u5c55\u5f00\u3002<\/p>\n<h4>8. \u4e0e\u4e3b\u6d41\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u7684\u5173\u7cfb<\/h4>\n<ul>\n<li><strong>U-Net \/ nnU-Net<\/strong>\uff1aGeoProto \u4e0d\u662f\u7ecf\u5178 encoder-decoder \u5168\u76d1\u7763\u6846\u67b6\uff0c\u4e5f\u4e0d\u662f nnU-Net recipe \u6539\u8fdb\uff1b\u5b83\u5173\u6ce8 few-shot prototype matching\u3002\u53ef\u501f\u9274\u7684\u662f geometry-aware local prototype\uff0c\u800c\u4e0d\u662f\u6574\u4f53\u66ff\u4ee3 nnU-Net\u3002<\/li>\n<li><strong>MedNeXt \/ CNN segmentation<\/strong>\uff1a\u53ef\u628a\u5176\u601d\u60f3\u7528\u4e8e CNN feature prototype \u6216 memory bank\uff0c\u4f7f CNN \u7279\u5f81\u5e26\u4e0a\u8fb9\u754c-\u4e2d\u5fc3\u7ed3\u6784\u6761\u4ef6\u3002<\/li>\n<li><strong>UNETR \/ Swin-UNet \/ TransUNet \/ TransFuse<\/strong>\uff1a\u8bba\u6587\u5f53\u524d backbone \u662f ResNet-101\uff0c\u4e0d\u662f Transformer segmentation \u67b6\u6784\uff1b\u4f46 geometry offset \u53ef\u4ee5\u7406\u8bba\u4e0a\u52a0\u5230 Transformer token prototype \u6216 prompt token \u4e0a\u3002<\/li>\n<li><strong>Mamba \/ VMamba \/ SegMamba \/ DAMamba<\/strong>\uff1aGeoProto \u4e0d\u6d89\u53ca state space module\uff1b\u4f46\u5b83\u63d0\u4f9b\u4e86\u4e00\u4e2a orthogonal idea\uff1a\u5728 Mamba\/CNN encoder \u4e4b\u5916\uff0c\u7528 distance-transform geometry \u5bf9 prototype \u6216 decoder query \u505a\u6761\u4ef6\u5316\u3002\u5bf9 DAMamba \u6539\u9020\u7684\u542f\u53d1\u662f\uff1a\u53ef\u4ee5\u628a boundary-to-centroid ordinal map \u4f5c\u4e3a\u8f85\u52a9\u76d1\u7763\u6216 gating signal\uff0c\u800c\u4e0d\u662f\u53ea\u4f9d\u8d56 sequence scanning\u3002<\/li>\n<li><strong>Foundation model segmentation<\/strong>\uff1a\u548c SAM\/MedSAM \u7684 promptable segmentation \u6709\u6f5c\u5728\u8fde\u63a5\uff1asupport mask \u7684 EDT geometry \u53ef\u4ee5\u4f5c\u4e3a prompt prototype \u7684\u7ed3\u6784\u5148\u9a8c\uff0c\u4f46\u8bba\u6587\u6ca1\u6709\u76f4\u63a5\u9a8c\u8bc1 foundation model \u573a\u666f\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u5982\u679c\u7528\u6237\u5173\u6ce8 polyp segmentation \u548c DAMamba\uff0cGeoProto \u7684\u76f4\u63a5 baseline \u4ef7\u503c\u4e0d\u5982 U-Net\/nnU-Net \u7c7b\u8bba\u6587\uff0c\u4f46\u6982\u5ff5\u4ef7\u503c\u8f83\u9ad8\uff1a\u5b83\u63d0\u9192\u6211\u4eec\u5728\u8de8\u57df\u5206\u5272\u4e2d\uff0cboundary-aware \u6216 center-aware geometry \u53ef\u4ee5\u6bd4\u5355\u7eaf\u52a0\u6ce8\u610f\u529b\u3001Mamba block \u66f4\u6709\u89e3\u91ca\u529b\u3002\u5bf9\u606f\u8089\u4efb\u52a1\uff0c\u53ef\u8003\u8651\u628a EDT\/\u8fb9\u754c\u8ddd\u79bb\u56fe\u4f5c\u4e3a\u8f85\u52a9\u5206\u652f\u3001boundary-to-interior ordinal supervision\u3001\u6216\u5728 prototype\/memory-based polyp adaptation \u4e2d\u52a0\u5165\u51e0\u4f55\u504f\u79fb\u3002\u5bf9 DAMamba\uff0c\u53ef\u5c1d\u8bd5\u628a distance strata \u4f5c\u4e3a decoder-side gating \u6216 scan-order bias\uff0c\u5c24\u5176\u7528\u4e8e\u63d0\u5347\u8fb9\u754c\u4e0e\u5185\u90e8\u4e00\u81f4\u6027\u3002<\/p>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5efa\u8bae\u7cbe\u8bfb<\/strong>\u3002\u7406\u7531\u662f\u65b9\u6cd5\u673a\u5236\u6e05\u695a\u3001\u5b9e\u9a8c\u8986\u76d6\u591a\u4e2a\u8de8\u57df setting\u3001\u6d88\u878d\u80fd\u652f\u6491\u6838\u5fc3\u5047\u8bbe\uff1b\u4f46\u82e5\u5f53\u524d\u76ee\u6807\u662f\u6784\u5efa\u5e38\u89c4\u5168\u76d1\u7763 polyp segmentation \u4e3b\u5e72\uff0c\u5b83\u4e0d\u662f\u6700\u4f18\u5148 baseline\uff0c\u5e94\u4f5c\u4e3a\u201c\u8de8\u57df few-shot \/ geometry prior \/ prototype matching\u201d\u65b9\u5411\u7684\u542f\u53d1\u6587\u732e\u9605\u8bfb\u3002<\/p>\n<hr \/>\n<h2>\u8bba\u6587 2\uff1aXTinyU-Net: Training-Free U-Net Scaling via Initialization-Time Sensitivity<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aXTinyU-Net: Training-Free U-Net Scaling via Initialization-Time Sensitivity<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aAlvin Kimbowa, Moein Heidari, David Liu, Ilker Hacihaliloglu \/ \u7b2c\u4e00\u4f5c\u8005 Alvin Kimbowa<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-10<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint, arXiv:2605.09639v1<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.09639<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.09639v1 \uff08\u5df2\u4e0b\u8f7d\uff1aMEDIA:\/tmp\/medseg_daily_20260513\/xtiny_unet.pdf\uff09<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1a\u8bba\u6587\u548c arXiv \u9875\u9762\u7ed9\u51fa https:\/\/github.com\/alvinkimbowa\/nntinyunet.git \uff0c\u4f46\u4eca\u65e5\u8bbf\u95ee\u5bf9\u5e94 GitHub \u9875\u9762\u8fd4\u56de 404\uff1b\u56e0\u6b64\u4ee3\u7801\u201c\u58f0\u79f0\u516c\u5f00\u201d\uff0c\u4f46\u5f53\u524d\u672a\u786e\u8ba4\u53ef\u7528<\/li>\n<li>\u4efb\u52a1\uff1a\u8f7b\u91cf\u5316\u533b\u5b66\u56fe\u50cf\u5206\u5272\uff1b\u8bad\u7ec3\u524d\u9009\u62e9 dataset-specific U-Net width scaling configuration<\/li>\n<li>\u6570\u636e\u96c6\uff1aBUS-BRA\u3001EchoNet Dynamic\u3001ISIC 2018\u3001FiVES\u3001BraTS2020\u3001ACDC<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1aU-Net \/ nnU-Net width scaling\uff1bzero-cost \/ training-free architecture selection\uff1blightweight 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>XTinyU-Net \u7684\u91cd\u8981\u4ef7\u503c\u4e0d\u662f\u53d1\u660e\u65b0 block\uff0c\u800c\u662f\u63d0\u51fa\u4e00\u4e2a\u5f88\u5b9e\u7528\u7684\u5224\u65ad\uff1a\u5728 nnU-Net \u6846\u67b6\u5185\u4ec5\u505a width scaling\uff0c\u5e76\u7528\u521d\u59cb\u5316\u65f6\u7684 Jacobian sensitivity \u627e\u5230\u201c\u6027\u80fd\u5d29\u584c\u524d\u7684\u6700\u5c0f U-Net\u201d\uff0c\u53ef\u80fd\u6bd4\u8bb8\u591a\u590d\u6742\u8f7b\u91cf\u5316\u67b6\u6784\u66f4\u7701\u53c2\u6570\u4e14\u4e0d\u727a\u7272 Dice\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u8bba\u6587\u7814\u7a76\u8d44\u6e90\u53d7\u9650\u533b\u5b66\u56fe\u50cf\u5206\u5272\uff1aU-Net\/nnU-Net \u5f88\u5f3a\uff0c\u4f46\u5b8c\u6574\u6a21\u578b\u5728\u79fb\u52a8\u8bbe\u5907\u3001\u5e8a\u65c1\u7cfb\u7edf\u3001\u57fa\u5c42\u533b\u9662\u6216\u5b9e\u65f6\u573a\u666f\u4e2d\u53c2\u6570\u548c FLOPs \u8fc7\u9ad8\uff1b\u624b\u5de5\u8bbe\u8ba1 lightweight architecture \u53c8\u5f80\u5f80\u9700\u8981\u5927\u91cf train-and-evaluate sweep\u3002paper map \u53ef\u6982\u62ec\u4e3a\uff1a\u5b83\u7814\u7a76\u5982\u4f55\u5728\u4e0d\u8bad\u7ec3\u5019\u9009\u6a21\u578b\u7684\u60c5\u51b5\u4e0b\uff0c\u4e3a\u7279\u5b9a\u533b\u5b66\u5206\u5272\u6570\u636e\u96c6\u9009\u62e9\u8db3\u591f\u5c0f\u4f46\u5c1a\u672a\u5bb9\u91cf\u5d29\u584c\u7684 U-Net\uff1b\u4e3b\u52a8\u4f5c\u662f\u6784\u9020 width-capped U-Net family\uff0c\u5e76\u7528\u521d\u59cb\u5316\u65f6\u8f93\u5165-\u8f93\u51fa Jacobian sensitivity \u66f2\u7ebf\u7684 total variation \u68c0\u6d4b collapse boundary\uff1b\u4f5c\u8005\u58f0\u79f0 XTinyU-Net \u5728 6 \u4e2a\u6570\u636e\u96c6\u4e0a\u63a5\u8fd1\u6216\u8d85\u8fc7 nnU-Net Dice\uff0c\u540c\u65f6\u53c2\u6570\u5c11 400\u00d7\u20131600\u00d7\uff0c\u5e76\u4f18\u4e8e UNeXt\u3001CMUNeXt-S\u3001TinyU-Net\uff1b\u8bc1\u636e\u6765\u81ea nnU-Net \u534f\u8bae\u4e0b 6 \u4e2a\u516c\u5f00\u6570\u636e\u96c6\u3001\u4e09\u6b21\u8fd0\u884c\u5747\u503c\u65b9\u5dee\u3001\u8f7b\u91cf baseline \u5bf9\u6bd4\u548c NAS metric \u6d88\u878d\uff1b\u4e3b\u8981\u98ce\u9669\u662f Jacobian collapse boundary \u4e0e\u771f\u5b9e\u6027\u80fd\u5d29\u584c\u7684\u5173\u8054\u53ef\u80fd\u4f9d\u8d56\u6a21\u578b\u65cf\u3001\u9884\u5904\u7406\u3001patch size \u548c\u6570\u636e\u96c6\uff0c\u5c1a\u672a\u5728 3D \u5168\u5206\u8fa8\u7387\u6216\u66f4\u590d\u6742\u4efb\u52a1\u4e0a\u5145\u5206\u9a8c\u8bc1\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u6279\u8bc4\u4e24\u7c7b\u73b0\u72b6\uff1a\u7b2c\u4e00\uff0c\u8f7b\u91cf\u5316\u6a21\u578b\u5982 UNeXt\u3001CMUNeXt\u3001TinyU-Net\u3001LeanUNet\u3001MedNCA \u7b49\u901a\u8fc7\u65b0 block\u3001\u7b80\u5316 decoder\u3001token mixing \u7b49\u964d\u4f4e\u53c2\u6570\uff0c\u4f46\u5f80\u5f80\u6ca1\u6709\u89e3\u51b3\u201c\u4e0d\u540c\u6570\u636e\u96c6\u9700\u8981\u4e0d\u540c\u5bb9\u91cf\u201d\u7684\u9009\u62e9\u95ee\u9898\uff1b\u7b2c\u4e8c\uff0czero-cost NAS \u6307\u6807\u591a\u9762\u5411\u5206\u7c7b\u6216\u5b8f\u89c2\u67b6\u6784\u6392\u5e8f\uff0c\u5728\u533b\u5b66 dense prediction \u4e2d\u5bf9\u540c\u4e00 U-Net \u5bb6\u65cf\u7684\u7ec6\u7c92\u5ea6\u5bb9\u91cf\u53d8\u5316\u4e0d\u591f\u654f\u611f\u3002\u8bba\u6587\u8ba4\u4e3a\u771f\u6b63\u5b9e\u7528\u7684\u95ee\u9898\u662f\uff1a\u5bf9\u7ed9\u5b9a\u6570\u636e\u96c6\uff0cU-Net \u53ef\u4ee5\u7f29\u5230\u591a\u5c0f\u800c\u4e0d\u5d29\uff1f<\/p>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>\u8bba\u6587\u8def\u7ebf\u4e3a method-algorithm\uff0c\u8bc1\u636e\u91cd\u70b9\u662f experimental-eval\u3001reproducibility-and-compute\u3001ablation-and-mechanism-isolation\u3002\u65b9\u6cd5\u6b65\u9aa4\u5982\u4e0b\uff1a<\/p>\n<ol>\n<li>\u56fa\u5b9a U-Net \/ nnU-Net \u7684\u7a7a\u95f4\u5c42\u7ea7\u3001stage \u6570\u3001\u4e0b\u91c7\u6837\u8def\u5f84\u548c\u8bad\u7ec3 recipe\uff0c\u53ea\u6539\u53d8 channel width\u3002<\/li>\n<li>\u8bbe baseline \u7b2c <code>l<\/code> \u5c42\u901a\u9053\u6570\u4e3a <code>C_l^base = 2^l C_0<\/code>\uff0cbaseline \u6700\u5927\u901a\u9053 <code>C_max^base = C_{L-1}^base<\/code>\u3002<\/li>\n<li>\u6784\u9020\u79bb\u6563\u6a21\u578b\u65cf\uff1a<code>C_max(i)=C_max^base \/ 2^i<\/code>\uff0c\u6bcf\u5c42\u901a\u9053 <code>C_l(i)=min(2^l C_0, C_max(i))<\/code>\u3002\u8bba\u6587\u91c7\u7528 <code>C_max^base=512<\/code>\uff0c\u5f62\u6210 10 \u4e2a\u5019\u9009\u6a21\u578b\uff0cindex \u8d8a\u5927\u6a21\u578b\u8d8a\u5c0f\u3002<\/li>\n<li>\u5bf9\u6bcf\u4e2a\u672a\u8bad\u7ec3\u6a21\u578b <code>M_i<\/code>\uff0c\u53d6\u5c11\u91cf\u672a\u6807\u6ce8\u8f93\u5165\u56fe\u50cf\uff0c\u8ba1\u7b97\u8f93\u51fa\u5bf9\u8f93\u5165\u7684\u68af\u5ea6\uff1a<code>g_i(x)=\u2207_x \u03a3_u f_i(x;\u03b8_i)_u<\/code>\uff0c\u5e76\u7528 <code>||g_i(x)||_2<\/code> \u7684 RMS \u805a\u5408\u4e3a sensitivity score <code>S_i<\/code>\u3002<\/li>\n<li>\u5bf9\u6240\u6709\u5019\u9009\u7684 <code>S_i<\/code> \u505a min-max normalize\uff0c\u8ba1\u7b97\u76f8\u90bb\u6a21\u578b sensitivity \u5dee\u5206 <code>d_i=|S_{i+1}-S_i|<\/code>\u3002<\/li>\n<li>\u5bf9\u5019\u9009 split <code>k<\/code>\uff0c\u6bd4\u8f83\u5927\u6a21\u578b\u533a\u548c\u5c0f\u6a21\u578b\u533a\u7684 total variation\uff0c\u9009\u62e9\u4f7f pre-collapse \u533a\u7a33\u5b9a\u3001post-collapse \u533a\u6ce2\u52a8\u5927\u7684\u8fb9\u754c <code>k*<\/code>\uff0c\u6700\u7ec8\u9009 <code>M_{k*}<\/code> \u4f5c\u4e3a XTinyU-Net\u3002<\/li>\n<\/ol>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<ul>\n<li><strong>Width-capped U-Net family<\/strong>\uff1a\u8f93\u5165\u662f nnU-Net baseline \u914d\u7f6e\uff0c\u8f93\u51fa\u662f\u4e00\u7ec4\u53ea\u6539\u53d8\u6700\u5927 channel cap \u7684 U-Net\u3002\u5b83\u4fdd\u7559\u7a7a\u95f4\u62d3\u6251\u548c receptive field\uff0c\u56e0\u6b64\u628a\u53d8\u91cf\u5c3d\u91cf\u9650\u5236\u4e3a capacity\uff0c\u800c\u4e0d\u662f\u6df7\u5165 depth\u3001kernel\u3001decoder \u8bbe\u8ba1\u5dee\u5f02\u3002\u8fd9\u4e00\u70b9\u5bf9\u516c\u5e73\u7814\u7a76\u5f88\u5173\u952e\u3002<\/li>\n<li><strong>Jacobian sensitivity score<\/strong>\uff1a\u8f93\u5165\u5c11\u91cf\u672a\u6807\u6ce8\u56fe\u50cf\u548c\u968f\u673a\u521d\u59cb\u5316\u6a21\u578b\uff0c\u8f93\u51fa\u6bcf\u4e2a\u6a21\u578b\u7684 input-output sensitivity\u3002\u76f4\u89c9\u662f\u8fc7\u7a84\u6a21\u578b\u5728\u5bb9\u91cf\u8fb9\u754c\u9644\u8fd1\u4f1a\u5bf9\u8f93\u5165\u6270\u52a8\u66f4\u654f\u611f\uff0c\u66f2\u7ebf\u53d8\u5316\u66f4\u5267\u70c8\u3002\u5b83\u4e0d\u9700\u8981 label\uff0c\u4e5f\u4e0d\u9700\u8981\u8bad\u7ec3\uff0c\u56e0\u6b64\u9002\u5408\u5feb\u901f\u6a21\u578b\u9009\u62e9\u3002<\/li>\n<li><strong>Total-variation collapse detector<\/strong>\uff1a\u4e0d\u662f\u628a sensitivity \u5f53\u201c\u8d8a\u5927\u8d8a\u597d\u201d\u7684 NAS \u5206\u6570\uff0c\u800c\u662f\u627e sensitivity \u66f2\u7ebf\u4ece\u5e73\u7a33\u5230\u5267\u70c8\u53d8\u5316\u7684\u8fb9\u754c\u3002\u8fd9\u4e2a\u8bbe\u8ba1\u662f\u8bba\u6587\u76f8\u5bf9\u666e\u901a zero-cost NAS \u7684\u4e3b\u8981\u5dee\u5f02\u3002<\/li>\n<li><strong>XTinyU-Net selected configuration<\/strong>\uff1a\u4e0d\u662f\u56fa\u5b9a\u67b6\u6784\uff0c\u800c\u662f\u6bcf\u4e2a\u6570\u636e\u96c6\u4e00\u4e2a\u9009\u51fa\u6765\u7684\u6700\u5c0f pre-collapse U-Net\u3002\u56e0\u6b64\u5b83\u66f4\u50cf\u4e00\u4e2a\u9009\u62e9\u7b97\u6cd5\uff0c\u800c\u4e0d\u662f\u5355\u4e00\u6a21\u578b\u3002<\/li>\n<li><strong>\u662f\u5426\u9002\u5408\u8fc1\u79fb\u5230 polyp \/ 3D<\/strong>\uff1a\u5bf9 polyp segmentation \u5f88\u9002\u5408\u5c1d\u8bd5\uff0c\u56e0\u4e3a\u606f\u8089\u5206\u5272\u5e38\u7528 2D U-Net\/UNet++\/PraNet\/\u8f7b\u91cf\u6a21\u578b\uff0c\u4e14\u90e8\u7f72\u5b9e\u65f6\u6027\u91cd\u8981\uff1b\u53ea\u8981\u80fd\u5728 colonoscopy \u6570\u636e\u4e0a\u6784\u9020 width family\uff0c\u5c31\u80fd\u7528\u672a\u6807\u6ce8\u56fe\u50cf\u5148\u7b5b\u4e00\u4e2a\u5bb9\u91cf\u8fb9\u754c\u3002\u5bf9 3D segmentation \u4e5f\u6709\u6f5c\u529b\uff0c\u4f46 3D Jacobian \u8ba1\u7b97\u5185\u5b58\u66f4\u9ad8\uff0c\u4e14\u8bba\u6587\u5f53\u524d\u5b9e\u9a8c\u672a\u8bc1\u660e 3D nnU-Net \u5168\u6d41\u7a0b\u6709\u6548\u3002<\/li>\n<\/ul>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p>\u8bba\u6587\u5728 nnU-Net training\/evaluation protocol \u5185\u6bd4\u8f83\uff0c\u62a5\u544a Dice \u548c HD95\uff0c\u5e76\u7ed9\u51fa\u4e09\u6b21\u8fd0\u884c mean\u00b1std\u3002\u6570\u636e\u96c6\u5305\u62ec\u4e73\u817a\u8d85\u58f0 BUS-BRA\u3001EchoNet Dynamic \u5fc3\u8d85\u3001ISIC2018 \u76ae\u80a4\u75c5\u7076\u3001FiVES \u89c6\u7f51\u819c\u8840\u7ba1\u3001BraTS2020 \u8111\u80bf\u7624\u3001ACDC \u5fc3\u810f MRI\u3002<\/p>\n<p>\u5173\u952e\u7ed3\u679c\uff1a<\/p>\n<ul>\n<li>\u5728 BUS-BRA \u4e0a\uff0cXTinyU-Net \u53ea\u6709 20.39k \u53c2\u6570\u30010.92G FLOPs\uff0c\u4f46 Dice 90.64\u00b17.57\uff0c\u9ad8\u4e8e nnU-Net \u7684 89.90\u00b110.25\uff1b\u53c2\u6570\u4ece 33,472.40k \u964d\u5230 20.39k\uff0c\u8d85\u8fc7 1600\u00d7 \u51cf\u5c11\u3002<\/li>\n<li>EchoNet Dynamic \u4e0a XTinyU-Net Dice 92.31\u00b13.94\uff0c\u63a5\u8fd1 nnU-Net 92.45\u00b14.16\uff0cHD95 3.78 \u4e0e nnU-Net 3.75 \u63a5\u8fd1\u3002<\/li>\n<li>ISIC2018 \u4e0a XTinyU-Net Dice 89.39\u00b112.38\uff0c\u9ad8\u4e8e nnU-Net 87.84\u00b112.19\u3002<\/li>\n<li>FiVES \u4e0a XTinyU-Net Dice 88.97\u00b110.42\uff0c\u7565\u9ad8\u4e8e nnU-Net 88.57\u00b110.03\uff0c\u53c2\u6570 79.55k\u3001FLOPs 1.96G\u3002<\/li>\n<li>BraTS2020 \u4e0a XTinyU-Net Dice 89.47\u00b111.30\uff0c\u4f4e\u4e8e nnU-Net 91.19\u00b111.17\uff0c\u4f46\u8fdc\u9ad8\u4e8e UNeXt\u3001CMUNeXt-S\u3001TinyU-Net\u3002<\/li>\n<li>ACDC \u4e0a XTinyU-Net Dice 91.86\u00b16.06\uff0c\u63a5\u8fd1\/\u7565\u9ad8\u4e8e nnU-Net 91.37\u00b16.01\uff0cHD95 2.53 \u4e0e nnU-Net 2.27 \u63a5\u8fd1\u3002<\/li>\n<\/ul>\n<p>\u6d88\u878d\u663e\u793a\uff0cJacobian-based collapse detection \u5bf9 batch size \u548c random initialization \u6709\u4e00\u5b9a\u9c81\u68d2\u6027\uff1bNASWOT\u3001SWAP \u7b49\u5df2\u6709 zero-cost metrics \u5f80\u5f80\u968f\u6a21\u578b\u5bb9\u91cf\u5355\u8c03\u53d8\u5316\uff0c\u56e0\u6b64\u4e0d\u80fd\u5b9a\u4f4d sharp pre-\/post-collapse transition\u3002<\/p>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u53ef\u4fe1\u4e4b\u5904\uff1a\u8bba\u6587\u907f\u514d\u4e86\u201c\u6362\u65b0\u6a21\u5757 + \u6362\u8bad\u7ec3 recipe\u201d\u7684\u6df7\u6742\uff0c\u628a\u6240\u6709\u6a21\u578b\u653e\u5728 nnU-Net \u534f\u8bae\u4e0b\u6bd4\u8f83\uff1b\u62a5\u544a\u4e86 Dice\/HD95 \u548c\u4e09\u6b21\u8fd0\u884c\u5747\u503c\u65b9\u5dee\uff1bbaseline \u5305\u62ec UNeXt\u3001CMUNeXt-S\u3001TinyU-Net\uff0c\u4e5f\u5305\u62ec\u5b8c\u6574 nnU-Net \u4f5c\u4e3a upper\/reference bound\u3002\u7ed3\u8bba\u201c\u5f88\u591a\u6570\u636e\u96c6\u4e0a\u6807\u51c6 U-Net \u7f29\u5bbd\u540e\u5df2\u8db3\u591f\u5f3a\u201d\u6709\u8f83\u5f3a\u5de5\u7a0b\u610f\u4e49\u3002<\/p>\n<p>\u4e0d\u8db3\u4e4b\u5904\uff1a\u7b2c\u4e00\uff0c\u8bba\u6587\u76ee\u524d\u662f arXiv preprint\uff0c\u4ee3\u7801\u94fe\u63a5\u4eca\u65e5\u672a\u80fd\u8bbf\u95ee\uff0c\u590d\u73b0\u6027\u6682\u65f6\u6253\u6298\uff1b\u7b2c\u4e8c\uff0c\u867d\u7136\u8bf4 within nnU-Net framework\uff0c\u4f46\u5b9e\u9a8c\u770b\u8d77\u6765\u4e3b\u8981\u662f 2D \u6216\u5207\u7247\u7ea7\u4efb\u52a1\uff0c\u5e76\u6ca1\u6709\u5145\u5206\u5c55\u793a 3D full-resolution nnU-Net \u7684\u663e\u5b58\u3001patch\u3001spacing\u3001cascade \u7ec6\u8282\uff1b\u7b2c\u4e09\uff0ccollapse detector \u7684\u6570\u5b66\u5f62\u5f0f\u5728 PDF \u4e2d <code>TVL\/TVR<\/code> \u533a\u57df\u547d\u540d\u7565\u5bb9\u6613\u6df7\u6dc6\uff0c\u4e14\u9009\u62e9 <code>argmax |TVR-TVL|<\/code> \u7684\u7406\u8bba\u89e3\u91ca\u4ecd\u504f\u7ecf\u9a8c\uff1b\u7b2c\u56db\uff0c\u53c2\u6570\u6781\u4f4e\u65f6 Dice \u63a5\u8fd1 nnU-Net \u8fd9\u4e00\u73b0\u8c61\u5f88\u5438\u5f15\u4eba\uff0c\u4f46\u9700\u8981\u5728\u66f4\u591a\u591a\u7c7b\u522b\u3001\u5c0f\u76ee\u6807\u3001\u4f4e\u5bf9\u6bd4\u75c5\u7076\u4efb\u52a1\u4e0a\u9a8c\u8bc1\u662f\u5426\u4ecd\u6210\u7acb\u3002<\/p>\n<h4>8. \u4e0e\u4e3b\u6d41\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u7684\u5173\u7cfb<\/h4>\n<ul>\n<li><strong>U-Net \/ nnU-Net<\/strong>\uff1a\u8fd9\u662f\u8bba\u6587\u7684\u4e2d\u5fc3\u3002\u5b83\u4e0d\u662f\u66ff\u4ee3 U-Net\uff0c\u800c\u662f\u628a U-Net \u7684 width \u5f53\u4f5c dataset-specific capacity knob\uff0c\u5728 nnU-Net recipe \u4e0b\u5bfb\u627e\u6700\u5c0f\u53ef\u7528\u914d\u7f6e\u3002<\/li>\n<li><strong>MedNeXt<\/strong>\uff1aMedNeXt \u4ee3\u8868\u5927 kernel ConvNeXt-style \u73b0\u4ee3 CNN \u5206\u5272\u6846\u67b6\uff1bXTinyU-Net \u63d0\u9192\u6211\u4eec\u5728\u5f15\u5165\u590d\u6742 CNN block \u524d\uff0c\u5e94\u5148\u6bd4\u8f83 properly scaled nnU-Net\uff0c\u5426\u5219\u53ef\u80fd\u9ad8\u4f30\u65b0\u7ed3\u6784\u4ef7\u503c\u3002<\/li>\n<li><strong>UNETR \/ Swin-UNet \/ TransUNet \/ TransFuse<\/strong>\uff1a\u8bba\u6587\u4e0d\u76f4\u63a5\u6d89\u53ca Transformer\uff1b\u4f46\u5b83\u7684\u89c2\u70b9\u5bf9 Transformer-based segmentation \u4e5f\u662f\u8b66\u793a\uff1a\u8f7b\u91cf\u5316\u4e0d\u4e00\u5b9a\u8981\u5148\u6362 token mixer\uff0c\u5148\u505a\u5bb9\u91cf\u8fb9\u754c\u641c\u7d22\u53ef\u80fd\u66f4\u6709\u6027\u4ef7\u6bd4\u3002<\/li>\n<li><strong>Mamba \/ VMamba \/ SegMamba \/ DAMamba<\/strong>\uff1a\u8bba\u6587\u4e0e Mamba \u65e0\u76f4\u63a5\u5173\u7cfb\uff0c\u4f46\u5bf9 DAMamba \u7814\u7a76\u975e\u5e38\u6709\u53c2\u8003\u4ef7\u503c\uff1a\u5982\u679c\u8981\u8bc1\u660e DAMamba block \u6709\u4ef7\u503c\uff0c\u5fc5\u987b\u548c\u5f3a\u4e14\u7ecf\u8fc7\u5bb9\u91cf\u8c03\u4f18\u7684 tiny U-Net \/ nnU-Net \u6bd4\uff0c\u800c\u4e0d\u662f\u53ea\u548c\u672a\u8c03\u4f18 heavy U-Net \u6216\u968f\u610f\u8f7b\u91cf baseline \u6bd4\u3002<\/li>\n<li><strong>Foundation model segmentation<\/strong>\uff1a\u4e0e MedSAM\/SAM \u7c7b\u65b9\u6cd5\u4e0d\u540c\uff0cXTinyU-Net \u8d70\u7684\u662f\u5c0f\u6a21\u578b\u90e8\u7f72\u8def\u7ebf\uff1b\u5728\u65e0 GPU \u6216\u5b9e\u65f6\u573a\u666f\u4e2d\uff0c\u53ef\u80fd\u6bd4 foundation model \u66f4\u5b9e\u9645\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u5bf9\u7528\u6237\u7684 polyp segmentation \/ DAMamba \u65b9\u5411\uff0cXTinyU-Net \u7684\u4ef7\u503c\u5f88\u9ad8\u3002\u7b2c\u4e00\uff0c\u5b83\u53ef\u4ee5\u4f5c\u4e3a\u8f7b\u91cf baseline\uff1a\u5728 CVC-ClinicDB\u3001Kvasir-SEG\u3001CVC-ColonDB\u3001ETIS\u3001EndoScene \u7b49\u606f\u8089\u6570\u636e\u4e0a\u5148\u8dd1 width-scaled U-Net family\uff0c\u518d\u770b DAMamba \u662f\u5426\u771f\u7684\u63d0\u5347\u3002\u7b2c\u4e8c\uff0c\u5b83\u63d0\u4f9b\u4e86\u4e00\u4e2a\u65e0\u9700\u8bad\u7ec3\u7684\u6a21\u578b\u9009\u62e9\u5de5\u5177\uff0c\u53ef\u5feb\u901f\u4f30\u8ba1\u67d0\u4e2a\u6570\u636e\u96c6\u9700\u8981\u7684\u6700\u5c0f channel cap\u3002\u7b2c\u4e09\uff0c\u5b83\u80fd\u5e2e\u52a9 introduction\/related work \u4e2d\u8ba8\u8bba\u201c\u8f7b\u91cf\u5316\u533b\u5b66\u5206\u5272\u4e0d\u4e00\u5b9a\u4f9d\u8d56\u590d\u6742\u6a21\u5757\uff0c\u5bb9\u91cf\u9009\u62e9\u672c\u8eab\u662f\u5173\u952e\u201d\u3002<\/p>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5f3a\u70c8\u5efa\u8bae\u7cbe\u8bfb\u5e76\u5c1d\u8bd5\u590d\u73b0\u6838\u5fc3\u9009\u62e9\u5668<\/strong>\u3002\u867d\u7136\u4ee3\u7801\u5f53\u524d\u672a\u786e\u8ba4\u53ef\u8bbf\u95ee\uff0c\u4f46\u7b97\u6cd5\u5b9e\u73b0\u5e76\u4e0d\u590d\u6742\uff1a\u6784\u9020 10 \u4e2a channel cap\u3001\u8ba1\u7b97\u968f\u673a\u521d\u59cb\u5316 Jacobian sensitivity\u3001\u627e total-variation boundary\u3002\u5bf9\u540e\u7eed\u505a polyp segmentation\u3001DAMamba ablation\u3001\u516c\u5e73 baseline \u8bbe\u8ba1\u90fd\u5f88\u6709\u76f4\u63a5\u5e2e\u52a9\u3002<\/p>\n<hr \/>\n<h2>\u4eca\u65e5\u63a8\u8350\u4f18\u5148\u7ea7<\/h2>\n<ol>\n<li><strong>XTinyU-Net: Training-Free U-Net Scaling via Initialization-Time Sensitivity<\/strong>\uff1a\u6700\u9002\u5408\u7528\u6237\u540e\u7eed\u533b\u5b66\u56fe\u50cf\u5206\u5272\u7814\u7a76\uff0c\u5c24\u5176\u662f polyp segmentation\u3001\u8f7b\u91cf\u5316 baseline\u3001\u516c\u5e73 ablation \u548c DAMamba \u5bf9\u6bd4\u5b9e\u9a8c\u3002\u5b83\u80fd\u76f4\u63a5\u8f6c\u5316\u4e3a\u5b9e\u9a8c\u7b56\u7565\uff1a\u5148\u627e\u5230 tiny U-Net \u5bb9\u91cf\u8fb9\u754c\uff0c\u518d\u8bc4\u4f30\u65b0\u6a21\u5757\u662f\u5426\u771f\u7684\u6709\u6548\u3002<\/li>\n<li><strong>Geometry-aware Prototype Learning for Cross-domain Few-shot Medical Image Segmentation<\/strong>\uff1a\u66f4\u9002\u5408\u4f5c\u4e3a\u8de8\u57df few-shot\u3001geometry prior\u3001prototype matching \u7684\u65b9\u6cd5\u542f\u53d1\u3002\u5bf9\u5e38\u89c4\u5168\u76d1\u7763\u606f\u8089\u5206\u5272\u4e0d\u662f\u76f4\u63a5 baseline\uff0c\u4f46\u5bf9\u8de8\u4e2d\u5fc3\/\u5c11\u6807\u6ce8\/\u8fb9\u754c\u7ed3\u6784\u5efa\u6a21\u6709\u8f83\u9ad8\u53c2\u8003\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 \/ \u63d0\u4f9b PDF \u94fe\u63a5\uff1aMEDIA:\/tmp\/medseg_daily_20260513\/geometry_prototype.pdf\uff1bhttps:\/\/arxiv.org\/pdf\/2605.10885v1<\/li>\n<li>\u8bba\u6587 2\uff1a\u5df2\u9644 PDF \/ \u63d0\u4f9b PDF \u94fe\u63a5\uff1aMEDIA:\/tmp\/medseg_daily_20260513\/xtiny_unet.pdf\uff1bhttps:\/\/arxiv.org\/pdf\/2605.09639v1<\/li>\n<\/ul>\n<h2>\u4eca\u65e5\u53ef\u6267\u884c\u5efa\u8bae<\/h2>\n<ol>\n<li>\u4f18\u5148\u590d\u73b0 XTinyU-Net \u7684 training-free width selection\uff1a\u5728\u4e00\u4e2a polyp segmentation \u6570\u636e\u96c6\u4e0a\u5148\u7528\u672a\u6807\u6ce8\u56fe\u50cf\u8ba1\u7b97 Jacobian sensitivity\uff0c\u9009\u51fa\u6700\u5c0f pre-collapse U-Net\uff0c\u518d\u4f5c\u4e3a DAMamba \u6539\u9020\u7684\u5f3a\u8f7b\u91cf baseline\u3002<\/li>\n<li>\u4ece GeoProto \u4e2d\u62bd\u53d6 EDT ordinal supervision \u601d\u8def\uff1a\u53ef\u5728 polyp segmentation decoder \u589e\u52a0 boundary-to-interior distance bin \u8f85\u52a9\u5934\uff0c\u89c2\u5bdf\u662f\u5426\u63d0\u5347\u8fb9\u754c Dice\u3001HD95 \u6216 mIoU\uff0c\u5c24\u5176\u9002\u5408\u505a\u8fb9\u754c\u8d28\u91cf\u6d88\u878d\u3002<\/li>\n<li>\u5199 related work \u65f6\u628a XTinyU-Net \u653e\u5728 lightweight \/ efficient U-Net selection\uff0c\u628a GeoProto \u653e\u5728 cross-domain few-shot \/ geometry-aware prototype\uff1b\u4e0d\u8981\u628a GeoProto \u5f53\u5e38\u89c4 nnU-Net \u6539\u8fdb\u6765\u5f15\u7528\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u4eca\u65e5\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6700\u65b0\u8bba\u6587\u7cbe\u8bfb\u8ffd\u8e2a \u4eca\u65e5\u7ed3\u8bba \u4eca\u5929 arXiv \u5728 2026-05-10 \u5230 2026-05-11 \u65b0\u589e\u4e86\u591a\u7bc7\u533b\u5b66\u56fe &#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-1047","post","type-post","status-publish","format-standard","hentry","category-85"],"_links":{"self":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1047","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=1047"}],"version-history":[{"count":0,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1047\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/media?parent=1047"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/categories?post=1047"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/tags?post=1047"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}