{"id":1045,"date":"2026-05-12T23:16:39","date_gmt":"2026-05-12T15:16:39","guid":{"rendered":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/12\/2026-05-12-%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%9adino-mvr-%e4%b8%8e%e9%87%8f%e5%8c%96-nnunet\/"},"modified":"2026-05-12T23:16:39","modified_gmt":"2026-05-12T15:16:39","slug":"2026-05-12-%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%9adino-mvr-%e4%b8%8e%e9%87%8f%e5%8c%96-nnunet","status":"publish","type":"post","link":"https:\/\/www.eutaboo.com\/index.php\/2026\/05\/12\/2026-05-12-%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%9adino-mvr-%e4%b8%8e%e9%87%8f%e5%8c%96-nnunet\/","title":{"rendered":"2026-05-12 \u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bba\u6587\u7cbe\u8bfb\uff1aDINO-MVR \u4e0e\u91cf\u5316 nnUNet"},"content":{"rendered":"<h1>\u4eca\u65e5\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6700\u65b0\u8bba\u6587\u7cbe\u8bfb\u8ffd\u8e2a<\/h1>\n<blockquote>\n<p>\u672c\u6587\u4ec5\u4fdd\u7559\u8bba\u6587\u9875\u9762\u4e0e PDF URL\uff0c\u4e0d\u4e0a\u4f20 PDF \u9644\u4ef6\u3002<\/p>\n<\/blockquote>\n<h2>\u4eca\u65e5\u7ed3\u8bba<\/h2>\n<p>\u4eca\u5929\u6ca1\u6709\u68c0\u7d22\u5230\u5df2\u786e\u8ba4\u9876\u4f1a \/ \u9876\u520a\u6b63\u5f0f\u63a5\u6536\u4e14\u672a\u91cd\u590d\u7684\u533b\u5b66\u56fe\u50cf\u5206\u5272\u65b0\u8bba\u6587\uff1b\u56e0\u6b64\u4ece 2026 \u5e74 5 \u6708\u4e0a\u65ec arXiv \u6700\u65b0\u9884\u5370\u672c\u4e2d\u7b5b\u9009\u4e86 2 \u7bc7\u66f4\u503c\u5f97\u8ddf\u8e2a\u7684\u5de5\u4f5c\u3002\u6574\u4f53\u8d8b\u52bf\u4ecd\u7136\u5f88\u6e05\u6670\uff1a\u4e00\u7c7b\u5de5\u4f5c\u8bd5\u56fe\u628a DINO \/ SAM \u7b49 foundation model \u7684\u51bb\u7ed3\u8868\u5f81\u8f6c\u5316\u4e3a\u4f4e\u6807\u6ce8\u5206\u5272\u80fd\u529b\uff0c\u53e6\u4e00\u7c7b\u5219\u56f4\u7ed5 nnU-Net \u7684\u90e8\u7f72\u6548\u7387\u3001\u62d3\u6251\u7ea6\u675f\u548c\u4e34\u5e8a\u53ef\u7528\u6027\u505a\u5de5\u7a0b\u5316\u589e\u5f3a\u3002<\/p>\n<p>\u4eca\u5929\u6700\u503c\u5f97\u4f18\u5148\u770b\u7684\u662f <strong>DINO-MVR<\/strong>\uff1a\u5b83\u4e0d\u63d0\u51fa\u65b0 backbone\uff0c\u800c\u662f\u628a\u201c\u51bb\u7ed3 DINOv3 \u7279\u5f81\u5982\u4f55\u8bfb\u51fa\u6210\u533b\u5b66\u5206\u5272 mask\u201d\u4f5c\u4e3a\u6838\u5fc3\u95ee\u9898\uff0c\u5b9e\u9a8c\u8bbe\u8ba1\u76f8\u5bf9\u514b\u5236\uff0c\u4e5f\u4e3b\u52a8\u627f\u8ba4\u6bd4\u8f83\u5e76\u975e\u4e25\u683c head-to-head\u3002\u7b2c\u4e8c\u7bc7 <strong>Topology-Constrained Quantized nnUNet<\/strong> \u4e0e nnU-Net\u30013D \u5206\u5272\u548c\u6a21\u578b\u538b\u7f29\u9ad8\u5ea6\u76f8\u5173\uff0c\u4f46\u8bc1\u636e\u94fe\u660e\u663e\u5f31\u4e00\u4e9b\uff0c\u9002\u5408\u4f5c\u4e3a\u201c\u62d3\u6251\u7ea6\u675f + \u91cf\u5316\u90e8\u7f72\u201d\u601d\u8def\u53c2\u8003\uff0c\u800c\u4e0d\u5b9c\u76f4\u63a5\u5f53\u5f3a baseline\u3002<\/p>\n<h2>\u68c0\u7d22\u8bf4\u660e<\/h2>\n<p>\u4eca\u65e5\u68c0\u7d22\u8303\u56f4\u8986\u76d6 arXiv \u6700\u65b0 2026 \u5e74 5 \u6708\u8bba\u6587\uff0c\u5e76\u91cd\u70b9\u4f7f\u7528\u5173\u952e\u8bcd\u7ec4\u5408\uff1a<code>medical image segmentation<\/code>\u3001<code>3D medical image segmentation<\/code>\u3001<code>nnUNet<\/code>\u3001<code>Mamba<\/code>\u3001<code>Transformer<\/code>\u3001<code>polyp segmentation<\/code>\u3001<code>foundation medical segmentation<\/code>\u3001<code>universal medical image segmentation<\/code>\u3002\u7531\u4e8e\u672a\u53d1\u73b0\u5f53\u5929 UTC 00:30 \u9644\u8fd1\u521a\u53d1\u5e03\u4e14\u6ee1\u8db3\u201c\u9876\u4f1a \/ \u9876\u520a\u6b63\u5f0f\u63a5\u6536 + \u672a\u91cd\u590d + \u5206\u5272\u4e3b\u4efb\u52a1\u201d\u7684\u8bba\u6587\uff0c\u5411\u524d\u56de\u6eaf\u5230 2026 \u5e74 5 \u6708 5\u20138 \u65e5\u7684 arXiv \u9884\u5370\u672c\u3002\u4e24\u7bc7\u5165\u9009\u8bba\u6587\u5747\u4e3a 2025 \u5e74\u4ee5\u540e\u8bba\u6587\uff0c\u4e14\u5747\u4e3a arXiv preprint\uff0c\u5c1a\u672a\u786e\u8ba4\u9876\u4f1a \/ \u9876\u520a\u63a5\u6536\u3002<\/p>\n<p>\u5df2\u68c0\u67e5\u5386\u53f2\u63a8\u8350\u8bb0\u5f55\u5e76\u6392\u9664\u4e86\u91cd\u590d\u8bba\u6587\uff1b\u5386\u53f2\u4e2d\u5df2\u63a8\u8350\u6216\u7cbe\u8bfb\u8fc7\u7684\u91cd\u590d\u5019\u9009\u5305\u62ec\uff1a<strong>TopoMamba: Topology-Aware Scanning and Fusion for Segmenting Heterogeneous Medical Visual Media<\/strong>\u3001<strong>ESICA: A Scalable Framework for Text-Guided 3D Medical Image Segmentation<\/strong>\u3001<strong>Sharpening Lightweight Models for Generalized Polyp Segmentation: A Boundary Guided Distillation from Foundation Models<\/strong>\u3001<strong>SemiSAM-O1<\/strong>\u3001<strong>PanGuide3D<\/strong>\u3001<strong>UHR-Net<\/strong> \u7b49\u3002<\/p>\n<hr \/>\n<h2>\u8bba\u6587 1\uff1aDINO-MVR: Multi-View Readout of Frozen DINOv3 for Annotation-Efficient Medical Segmentation<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aDINO-MVR: Multi-View Readout of Frozen DINOv3 for Annotation-Efficient Medical Segmentation<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aWei Jiang, Feng Liu, Nan Ye, Hongfu Sun \/ \u7b2c\u4e00\u4f5c\u8005 Wei Jiang<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-08<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.07221<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.07221v1<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1a\u672a\u83b7\u53d6\uff1bPDF \u4e0e arXiv \u5143\u6570\u636e\u4e2d\u672a\u786e\u8ba4\u5b98\u65b9 GitHub<\/li>\n<li>\u4efb\u52a1\uff1aannotation-efficient medical image segmentation\uff1b2D polyp \/ skin lesion \u5206\u5272\uff1bslice-wise 3D BraTS FLAIR whole-tumor \u5206\u5272<\/li>\n<li>\u6570\u636e\u96c6\uff1aKvasir-SEG\u3001ISIC 2018\u3001BraTS 2021 FLAIR<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1afrozen foundation model feature readout\uff1bDINOv3 frozen backbone\uff1blightweight MLP probe\uff1bmulti-resolution + TTA + entropy fusion + CRF \/ z-axis smoothing<\/li>\n<\/ul>\n<h3>paper-deep-reader \u7cbe\u8bfb\u7ed3\u679c<\/h3>\n<h4>1. \u4e00\u53e5\u8bdd\u7ed3\u8bba<\/h4>\n<p>\u8fd9\u7bc7\u8bba\u6587\u6700\u6709\u4ef7\u503c\u7684\u5730\u65b9\u4e0d\u662f\u63d0\u51fa\u65b0\u578b U-Net \/ Transformer \/ Mamba backbone\uff0c\u800c\u662f\u628a\u533b\u5b66\u5206\u5272\u91cd\u65b0\u8868\u8ff0\u4e3a\u201c\u5982\u4f55\u4ece\u51bb\u7ed3 DINOv3 dense features \u4e2d\u8bfb\u51fa mask\u201d\u7684\u95ee\u9898\uff0c\u5e76\u7528\u8f7b\u91cf MLP probe\u3001\u591a\u5c3a\u5ea6\u6d4b\u8bd5\u89c6\u89d2\u548c\u71b5\u5f15\u5bfc\u878d\u5408\u8bc1\u660e\u51bb\u7ed3\u81ea\u7136\u56fe\u50cf\u81ea\u76d1\u7763\u7279\u5f81\u5728\u4f4e\u6807\u6ce8\u533b\u5b66\u5206\u5272\u4e2d\u4ecd\u6709\u53ef\u7528\u6f5c\u529b\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u8bba\u6587\u7814\u7a76\u7684\u95ee\u9898\u662f\uff1a\u5728\u533b\u5b66\u56fe\u50cf\u50cf\u7d20\u7ea7 \/ \u4f53\u7d20\u7ea7\u6807\u6ce8\u6602\u8d35\u3001\u533b\u5b66 foundation model \u9002\u914d\u6210\u672c\u9ad8\u7684\u60c5\u51b5\u4e0b\uff0c\u662f\u5426\u53ef\u4ee5\u4e0d\u5fae\u8c03\u5927\u6a21\u578b backbone\uff0c\u53ea\u8bad\u7ec3\u5f88\u5c0f\u7684 readout head\uff0c\u5c31\u5f97\u5230\u53ef\u7528\u7684\u533b\u5b66\u5206\u5272\u6027\u80fd\u3002<\/p>\n<p>\u4f5c\u8005\u628a\u95ee\u9898\u8bbe\u7f6e\u5f97\u6bd4\u8f83\u514b\u5236\uff1a\u7ed9\u5b9a\u4e00\u4e2a\u9884\u8bad\u7ec3 <strong>DINOv3 ViT-B\/16<\/strong>\uff0c\u6574\u4e2a encoder \u53c2\u6570\u51bb\u7ed3\uff0c\u53ea\u8bad\u7ec3\u5206\u8fa8\u7387\u76f8\u5173\u7684\u4e24\u5c42 MLP probe\u3002\u8fd9\u6837\u53ef\u4ee5\u628a\u4e24\u4e2a\u95ee\u9898\u5206\u5f00\uff1a<\/p>\n<ol>\n<li>DINOv3 \u8fd9\u6837\u7684\u81ea\u7136\u56fe\u50cf\u81ea\u76d1\u7763\u89c6\u89c9\u6a21\u578b\u662f\u5426\u5df2\u7ecf\u542b\u6709\u53ef\u8fc1\u79fb\u5230\u533b\u5b66\u56fe\u50cf\u7684\u7ed3\u6784 \/ \u8fb9\u754c\u4fe1\u606f\uff1b<\/li>\n<li>\u5982\u679c\u542b\u6709\u8fd9\u4e9b\u4fe1\u606f\uff0c\u600e\u6837\u8bfb\u51fa\u8fd9\u4e9b\u4fe1\u606f\u624d\u6700\u6709\u6548\u3002<\/li>\n<\/ol>\n<p>\u8fd9\u4e0e\u4f20\u7edf U-Net \/ nnU-Net \u8bad\u7ec3\u8303\u5f0f\u4e0d\u540c\u3002U-Net \/ nnU-Net \u5047\u8bbe\u6709\u8db3\u591f in-domain \u6807\u6ce8\uff0c\u5e76\u901a\u8fc7\u7aef\u5230\u7aef\u8bad\u7ec3\u5f97\u5230\u4efb\u52a1\u4e13\u7528 segmenter\uff1bDINO-MVR \u5219\u5047\u8bbe backbone \u5df2\u7ecf\u56fa\u5b9a\uff0c\u4e3b\u8981\u4f18\u5316\u4ece frozen feature \u5230 mask \u7684\u8bfb\u51fa\u548c\u6d4b\u8bd5\u65f6\u878d\u5408\u7b56\u7565\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u4f5c\u8005\u6307\u51fa\u73b0\u6709\u8def\u7ebf\u4e3b\u8981\u6709\u51e0\u7c7b\u4e0d\u8db3\uff1a<\/p>\n<ul>\n<li><strong>U-Net \/ nnU-Net \u7c7b\u76d1\u7763\u65b9\u6cd5<\/strong>\uff1a\u5728\u8db3\u591f\u6807\u6ce8\u4e0b\u5f88\u5f3a\uff0c\u4f46\u8fc1\u79fb\u5230\u65b0\u5668\u5b98\u3001\u65b0\u6a21\u6001\u3001\u65b0\u533b\u9662\u65f6\u5e38\u9700\u8981\u91cd\u65b0\u5bc6\u96c6\u6807\u6ce8\u548c\u8bad\u7ec3\u3002<\/li>\n<li><strong>SAM \/ MedSAM \u7b49 promptable foundation model<\/strong>\uff1a\u5177\u5907\u5f3a\u6cdb\u5316\u6f5c\u529b\uff0c\u4f46\u533b\u5b66\u573a\u666f\u901a\u5e38\u9700\u8981 prompt\uff1b\u5982\u679c prompt \u6765\u81ea\u4eba\u5de5\u70b9\u3001\u6846\u6216 oracle mask\uff0c\u4f1a\u6539\u53d8\u81ea\u52a8\u5206\u5272\u4efb\u52a1\u5b9a\u4e49\u3002<\/li>\n<li><strong>\u53c2\u6570\u9ad8\u6548\u9002\u914d \/ adapter \/ prompt tuning<\/strong>\uff1a\u51cf\u5c11\u4e86\u5fae\u8c03\u6210\u672c\uff0c\u4f46\u4ecd\u7136\u8981\u4fee\u6539\u6a21\u578b\u7684\u4e00\u90e8\u5206\u53c2\u6570\uff0c\u5e76\u53ef\u80fd\u9700\u8981\u8f83\u591a\u533b\u5b66\u57df\u6570\u636e\u3002<\/li>\n<li><strong>\u5df2\u6709 frozen-DINO segmentation\uff0c\u5982 SegDINO<\/strong>\uff1a\u8bf4\u660e DINO \u7279\u5f81\u53ef\u7528\u4e8e\u5206\u5272\uff0c\u4f46\u4f5c\u8005\u8ba4\u4e3a\u8fd8\u6ca1\u6709\u5145\u5206\u63a2\u7d22\u201c\u591a\u5c42\u7279\u5f81\u3001\u591a\u5206\u8fa8\u7387\u3001\u591a\u6d4b\u8bd5\u53d8\u6362\u201d\u8fd9\u4e9b readout \u7ef4\u5ea6\u3002<\/li>\n<\/ul>\n<p>\u8bba\u6587\u7684\u6838\u5fc3\u7acb\u573a\u662f\uff1a\u533b\u5b66\u5206\u5272\u80fd\u529b\u53ef\u80fd\u5df2\u7ecf\u90e8\u5206\u5b58\u5728\u4e8e\u51bb\u7ed3 DINOv3 dense feature \u4e2d\uff0c\u74f6\u9888\u4e0d\u4e00\u5b9a\u662f backbone \u5fae\u8c03\uff0c\u800c\u662f readout \u8bbe\u8ba1\u4e0d\u8db3\u3002<\/p>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>DINO-MVR \u7684\u6574\u4f53\u6846\u67b6\u53ef\u4ee5\u5206\u4e3a\u8bad\u7ec3\u548c\u63a8\u7406\u4e24\u9636\u6bb5\u3002<\/p>\n<p><strong>\u8bad\u7ec3\u9636\u6bb5\uff1a<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\u56fe\u50cf <code>x<\/code> resize \u5230\u4e0d\u540c\u5206\u8fa8\u7387 <code>r<\/code>\uff1b<\/li>\n<li>\u4f7f\u7528\u51bb\u7ed3 DINOv3 ViT-B\/16 \u63d0\u53d6 patch token\uff1b<\/li>\n<li>\u53d6\u6700\u540e 3 \u4e2a transformer block \u7684\u7279\u5f81\u5e76 concat\uff1a<\/li>\n<li>\u6bcf\u5c42\u7279\u5f81\u7ef4\u5ea6\u4e3a <code>d=768<\/code>\uff1b<\/li>\n<li>concat \u540e\u6bcf\u4e2a patch token \u4e3a <code>3d=2304<\/code> \u7ef4\uff1b<\/li>\n<li>\u5bf9\u6bcf\u4e2a\u5206\u8fa8\u7387\u8bad\u7ec3\u4e00\u4e2a\u4e24\u5c42 MLP\uff1a<\/li>\n<li><code>2304 \u2192 256 \u2192 1<\/code>\uff1b<\/li>\n<li>\u6bcf\u4e2a\u5206\u8fa8\u7387\u7ea6 0.59M \u53ef\u8bad\u7ec3\u53c2\u6570\uff1b<\/li>\n<li>MLP \u8f93\u51fa patch-level foreground logit\uff1b<\/li>\n<li>\u4e0a\u91c7\u6837\u56de mask \u5c3a\u5bf8\uff1b<\/li>\n<li>\u4f7f\u7528 BCE + soft Dice loss \u8bad\u7ec3\uff1b<\/li>\n<li>DINOv3 backbone \u5b8c\u5168\u51bb\u7ed3\uff0c\u4e0d\u53cd\u4f20\u68af\u5ea6\u3002<\/li>\n<\/ul>\n<p><strong>\u63a8\u7406\u9636\u6bb5\uff1a<\/strong><\/p>\n<ul>\n<li>\u5bf9\u540c\u4e00\u56fe\u50cf\u4f7f\u7528\u591a\u4e2a\u89c6\u89d2\uff1a<\/li>\n<li>\u5206\u8fa8\u7387\uff1a<code>512<\/code> \u4e0e <code>1024<\/code>\uff1b<\/li>\n<li>test-time augmentation\uff1aidentity\u3001horizontal flip\u3001vertical flip\uff1b<\/li>\n<li>\u6bcf\u4e2a\u89c6\u89d2\u7ecf\u8fc7 frozen DINOv3 + \u5bf9\u5e94 MLP probe \u5f97\u5230\u6982\u7387\u56fe\uff1b<\/li>\n<li>\u5bf9\u540c\u4e00\u5206\u8fa8\u7387\u4e0b\u4e0d\u540c TTA \u7ed3\u679c\u6c42\u5e73\u5747\uff1b<\/li>\n<li>\u5bf9 512 \u548c 1024 \u4e24\u4e2a\u5206\u8fa8\u7387\u7684\u7ed3\u679c\u505a\u71b5\u5f15\u5bfc\u878d\u5408\uff1b<\/li>\n<li>\u53ef\u9009 DenseCRF \u8fdb\u884c\u8fb9\u754c \/ \u7a7a\u95f4\u4e00\u81f4\u6027 refinement\uff1b<\/li>\n<li>\u5bf9 3D volume\uff0c\u9010 slice \u5904\u7406\u540e\u6cbf z \u8f74\u505a Gaussian smoothing\uff0c\u4ee5\u51cf\u5c11\u8de8\u5c42\u4e0d\u4e00\u81f4\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u662f\u4e00\u79cd\u5178\u578b\u7684\u201c\u51bb\u7ed3\u8868\u5f81 + \u8f7b\u91cf probe + \u6d4b\u8bd5\u65f6 ensemble \/ \u540e\u5904\u7406\u201d\u65b9\u6cd5\uff0c\u800c\u4e0d\u662f\u7aef\u5230\u7aef\u53ef\u5b66\u4e60 3D segmentation architecture\u3002<\/p>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<p><strong>\u6a21\u5757 1\uff1aFrozen DINOv3 feature stack<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\uff1aresize \u540e\u56fe\u50cf <code>x(r)<\/code>\uff1b<\/li>\n<li>\u8f93\u51fa\uff1a\u6700\u540e\u4e09\u4e2a transformer block \u7684 patch feature concat\uff1b<\/li>\n<li>\u4f5c\u7528\uff1a\u5229\u7528 DINOv3 \u540e\u671f\u5c42\u4e2d\u4e0d\u540c\u7a0b\u5ea6\u7684\u8bed\u4e49 \/ \u8fb9\u754c \/ \u5c40\u90e8\u7ed3\u6784\u4fe1\u606f\uff1b<\/li>\n<li>\u521b\u65b0\u6027\u5224\u65ad\uff1a\u4e0d\u662f\u65b0 backbone\uff0c\u800c\u662f\u628a\u6700\u540e 3 \u5c42 concat \u4f5c\u4e3a readout object\uff1b\u521b\u65b0\u4e2d\u7b49\u504f\u5de5\u7a0b\u5316\u3002<\/li>\n<li>\u53ef\u8fc1\u79fb\u6027\uff1a\u8f83\u9ad8\u3002\u82e5\u7528\u6237\u60f3\u6d4b\u8bd5 frozen DINO \/ MAE \/ SAM encoder feature\uff0c\u5bf9 polyp segmentation \u53ef\u76f4\u63a5\u590d\u7528\u8fd9\u4e00 readout \u8bbe\u5b9a\u3002<\/li>\n<li>\u5bf9 3D \u5206\u5272\u9002\u914d\uff1a\u539f\u6587\u4ecd\u662f slice-wise 2D DINO \u7279\u5f81\uff0c\u4e0d\u80fd\u66ff\u4ee3\u771f\u6b63 3D encoder\u3002<\/li>\n<\/ul>\n<p><strong>\u6a21\u5757 2\uff1aScale-specific MLP probes<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\uff1a\u6bcf\u4e2a patch \u7684 2304 \u7ef4 concat feature\uff1b<\/li>\n<li>\u8f93\u51fa\uff1aforeground logit\uff1b<\/li>\n<li>\u4f5c\u7528\uff1a\u628a frozen feature \u8f6c\u4e3a dense binary mask\uff1b<\/li>\n<li>\u5173\u952e\u70b9\uff1a\u6bcf\u4e2a\u5206\u8fa8\u7387\u4e00\u4e2a probe\uff0c\u53c2\u6570\u91cf\u5c0f\uff1b<\/li>\n<li>\u521b\u65b0\u6027\u5224\u65ad\uff1aMLP probe \u672c\u8eab\u7b80\u5355\uff0c\u4f46\u5b83\u4f5c\u4e3a\u8bca\u65ad frozen representation \u7684\u5b9e\u9a8c\u8bbe\u8ba1\u662f\u5408\u7406\u7684\u3002<\/li>\n<li>\u9002\u5408\u8fc1\u79fb\u5230 polyp segmentation\uff1a\u9002\u5408\uff0c\u5c24\u5176\u9002\u5408\u4f4e\u6807\u6ce8 polyp \u6570\u636e\u6216\u5feb\u901f\u9a8c\u8bc1 foundation feature \u662f\u5426\u6709\u7528\u3002<\/li>\n<\/ul>\n<p><strong>\u6a21\u5757 3\uff1aMulti-resolution inference<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\uff1a512 \u4e0e 1024 \u4e24\u4e2a\u5206\u8fa8\u7387\u4e0b\u7684\u9884\u6d4b\uff1b<\/li>\n<li>\u8f93\u51fa\uff1a\u4e24\u7c7b\u6982\u7387\u56fe\uff1b<\/li>\n<li>\u4f5c\u7528\uff1a\u4f4e\u5206\u8fa8\u7387\u66f4\u7a33\u5b9a\uff0c\u9ad8\u5206\u8fa8\u7387\u4fdd\u7559\u8fb9\u754c\u7ec6\u8282\uff1b<\/li>\n<li>\u5b9e\u9a8c\u8bc1\u636e\uff1aKvasir ablation \u4e2d\u5355\u72ec 512 \u6216 1024 \u5206\u652f\u6bd4\u5b8c\u6574 512+1024 \u5dee\u3002<\/li>\n<li>\u5c40\u9650\uff1a\u63a8\u7406\u6210\u672c\u660e\u663e\u589e\u52a0\uff1b\u8fd9\u4e0d\u662f\u8f7b\u91cf\u5b9e\u65f6\u65b9\u6848\u3002<\/li>\n<\/ul>\n<p><strong>\u6a21\u5757 4\uff1aEntropy-guided fusion<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\uff1a512 \u9884\u6d4b <code>p_lo<\/code> \u4e0e 1024 \u9884\u6d4b <code>p_hi<\/code>\uff1b<\/li>\n<li>\u64cd\u4f5c\uff1a\u8ba1\u7b97 binary entropy\uff1b\u5f53\u4f4e\u5206\u8fa8\u7387\u5206\u652f\u71b5\u4f4e\u4e8e\u9608\u503c <code>\u03c4=0.3<\/code> \u65f6\u7528\u4f4e\u5206\u8fa8\u7387\uff0c\u5426\u5219\u7528\u9ad8\u5206\u8fa8\u7387\uff1b<\/li>\n<li>\u4f5c\u7528\uff1a\u8ba9\u7a33\u5b9a\u7684 coarse prediction \u4e3b\u5bfc\u5927\u533a\u57df\uff0c\u8ba9 high-res prediction \u5904\u7406\u4e0d\u786e\u5b9a\u8fb9\u754c\uff1b<\/li>\n<li>\u521b\u65b0\u6027\u5224\u65ad\uff1a\u89c4\u5219\u7b80\u5355\uff0c\u4e0d\u662f\u6df1\u5c42\u7406\u8bba\u8d21\u732e\uff0c\u4f46\u673a\u5236\u6e05\u695a\u3001\u53ef\u590d\u73b0\u3002<\/li>\n<li>\u5bf9 DAMamba \/ hybrid \u6846\u67b6\u542f\u53d1\uff1a\u53ef\u4f5c\u4e3a decoder \u8f93\u51fa\u540e\u5904\u7406\u6216\u591a\u5c3a\u5ea6\u9884\u6d4b\u878d\u5408\u7b56\u7565\uff0c\u800c\u4e0d\u662f\u66ff\u4ee3 Mamba block\u3002<\/li>\n<\/ul>\n<p><strong>\u6a21\u5757 5\uff1aDenseCRF \u4e0e z-axis smoothing<\/strong><\/p>\n<ul>\n<li>DenseCRF\uff1a\u7528\u4e8e 2D mask \u8fb9\u754c\u7a7a\u95f4\u4e00\u81f4\u6027\uff1b<\/li>\n<li>z-axis smoothing\uff1a\u5bf9 BraTS slice-wise \u6982\u7387\u56fe\u6cbf\u6df1\u5ea6\u8f74\u505a Gaussian smoothing\uff1b<\/li>\n<li>\u5b9e\u9a8c\u8bc1\u636e\uff1az-axis smoothing \u5728 BraTS pilot \u4e2d Dice \u4ece 0.8958 \u63d0\u5230\u7ea6 0.9057\uff0cHD95 \u660e\u663e\u964d\u4f4e\uff1b<\/li>\n<li>\u5c40\u9650\uff1a\u8fd9\u662f\u975e\u5b66\u4e60\u5f0f\u5e73\u6ed1\uff0c\u65e0\u6cd5\u5efa\u6a21\u590d\u6742 3D anatomical context\uff1b\u5bf9\u5c0f\u75c5\u7076\u6216\u975e\u8fde\u7eed\u7ed3\u6784\u53ef\u80fd\u8fc7\u5e73\u6ed1\u3002<\/li>\n<\/ul>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p><strong>\u6570\u636e\u96c6\u4e0e\u4efb\u52a1\uff1a<\/strong><\/p>\n<ul>\n<li>Kvasir-SEG\uff1a2D endoscopy polyp segmentation\uff1b\u7528\u4e8e\u591a\u89c6\u89d2 readout \u4e0e ablation\uff1b<\/li>\n<li>ISIC 2018\uff1a2D dermoscopy skin lesion segmentation\uff1b<\/li>\n<li>BraTS 2021 FLAIR\uff1avolumetric MRI whole-tumor segmentation\uff0c\u91c7\u7528 slice-wise readout + z smoothing\u3002<\/li>\n<\/ul>\n<p><strong>\u6307\u6807\uff1a<\/strong><\/p>\n<ul>\n<li>Dice \/ DSC\uff1b<\/li>\n<li>IoU\uff1b<\/li>\n<li>HD95\uff1b<\/li>\n<li>BraTS \u8fd8\u62a5\u544a K-patient annotation efficiency learning curve\u3002<\/li>\n<\/ul>\n<p><strong>\u4e3b\u8981\u7ed3\u679c\uff1a<\/strong><\/p>\n<ul>\n<li>Kvasir-SEG\uff1aDINO-MVR reported DSC 0.8946\uff0cIoU 0.8375\uff0cHD95 15.0\uff1b<\/li>\n<li>ISIC 2018\uff1aDSC 0.8976\uff0cIoU 0.8270\uff0cHD95 12.4\uff1b<\/li>\n<li>BraTS FLAIR\uff1a40-patient reference DSC 0.9082\uff1b5 annotated patients \u8fbe\u5230 0.8937\uff0c\u7ea6\u4e3a reference \u7684 98.4%\u3002<\/li>\n<\/ul>\n<p><strong>\u6d88\u878d\u5b9e\u9a8c\uff1a<\/strong><\/p>\n<p>Kvasir-SEG matched ablation \u6bd4\u8f83\u4e86\uff1a<\/p>\n<ul>\n<li>last-3 blocks vs last-1 \/ last-2\uff1b<\/li>\n<li>MLP probe vs linear probe\uff1b<\/li>\n<li>512+1024 vs 512-only \/ 1024-only\uff1b<\/li>\n<li>\u6709\u65e0 flip-based TTA\uff1b<\/li>\n<li>\u6709\u65e0 DenseCRF\uff1b<\/li>\n<li>raw single-branch readout\u3002<\/li>\n<\/ul>\n<p>\u5173\u952e\u89c2\u5bdf\uff1a<\/p>\n<ul>\n<li>\u7ebf\u6027 probe \u964d\u5e45\u6700\u5927\uff0c\u8bf4\u660e\u975e\u7ebf\u6027\u8bfb\u51fa\u91cd\u8981\uff1b<\/li>\n<li>\u53bb\u6389 TTA \u6216\u591a\u5206\u8fa8\u7387\u4f1a\u4e0b\u964d\uff1b<\/li>\n<li>DenseCRF \u5bf9 Dice \u4e0d\u4e00\u5b9a\u589e\u52a0\uff0c\u4f46\u6539\u5584 HD95\uff1b<\/li>\n<li>last-3 block concat \u6bd4 last-1 \/ last-2 \u7565\u4f18\uff0c\u4f46\u5dee\u8ddd\u4e0d\u5927\uff0c\u8bf4\u660e DINOv3 \u6700\u540e\u4e00\u5c42\u5df2\u7ecf\u5f88\u5f3a\u3002<\/li>\n<\/ul>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u603b\u4f53\u6765\u8bf4\uff0cDINO-MVR \u7684\u8bc1\u636e\u94fe\u6bd4\u5f88\u591a arXiv \u5de5\u7a0b\u62fc\u6a21\u5757\u8bba\u6587\u66f4\u53ef\u4fe1\uff0c\u4f46\u8fd8\u4e0d\u80fd\u89c6\u4e3a\u4e25\u683c SOTA \u8bc1\u660e\u3002<\/p>\n<p><strong>\u53ef\u4fe1\u4e4b\u5904\uff1a<\/strong><\/p>\n<ul>\n<li>\u4f5c\u8005\u660e\u786e\u8bf4\u660e 2D baseline \u8868\u683c\u4e0d\u662f\u4e25\u683c head-to-head\uff0c\u56e0\u4e3a split\u3001metric\u3001backbone size\u3001model selection \u53ef\u80fd\u4e0d\u540c\uff1b<\/li>\n<li>\u6709 Kvasir matched ablation\uff0c\u80fd\u652f\u6301\u4e3b\u8981\u8bbe\u8ba1\u9009\u62e9\uff1b<\/li>\n<li>\u6ca1\u6709\u628a prompt-based SAM \u4e0e\u81ea\u52a8\u5206\u5272\u65b9\u6cd5\u76f4\u63a5\u6df7\u4e3a\u540c\u4e00\u4efb\u52a1\uff1b<\/li>\n<li>\u5bf9\u63a8\u7406\u6210\u672c\u3001\u975e\u4e25\u683c\u6bd4\u8f83\u3001slice-wise 3D \u7684\u9650\u5236\u6709\u6e05\u695a\u8ba8\u8bba\u3002<\/li>\n<\/ul>\n<p><strong>\u4e3b\u8981\u5f31\u70b9\uff1a<\/strong><\/p>\n<ul>\n<li>2D \u8868\u683c\u4e2d\u7684 U-Net\u3001TransUNet\u3001SegDINO \u7b49\u591a\u4e3a\u6587\u732e reported numbers\uff0c\u4e0d\u662f\u7edf\u4e00\u590d\u73b0\u5b9e\u9a8c\uff1b<\/li>\n<li>\u7f3a\u5c11 nnU-Net \/ MedNeXt \/ UNetR \/ Swin-UNETR \u5728\u540c\u4e00 split\u3001\u540c\u4e00\u8bad\u7ec3\u9884\u7b97\u4e0b\u7684\u76f4\u63a5\u6bd4\u8f83\uff1b<\/li>\n<li>3D BraTS \u662f slice-wise \u65b9\u6cd5\u52a0 z smoothing\uff0c\u4e0d\u662f\u771f\u6b63 3D context \u6a21\u578b\uff1b<\/li>\n<li>\u63a8\u7406\u65f6\u4f7f\u7528 2 \u4e2a\u5206\u8fa8\u7387 \u00d7 3 \u4e2a TTA + CRF\uff0c\u8bad\u7ec3\u53c2\u6570\u867d\u5c11\uff0c\u4f46 inference cost \u4e0d\u4f4e\uff1b<\/li>\n<li>\u76ee\u524d\u672a\u786e\u8ba4\u4ee3\u7801\u5f00\u6e90\uff0c\u590d\u73b0\u5b9e\u7528\u6027\u53d7\u9650\u3002<\/li>\n<\/ul>\n<p>\u56e0\u6b64\uff0c\u8be5\u8bba\u6587\u66f4\u9002\u5408\u88ab\u7406\u89e3\u4e3a\u201cfrozen representation probing + readout strategy\u201d\u7684\u6709\u4ef7\u503c\u7814\u7a76\uff0c\u800c\u4e0d\u662f\u53ef\u76f4\u63a5\u66ff\u4ee3 nnU-Net \u7684\u4e34\u5e8a\u7ea7 segmenter\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>\u4e0e U-Net \/ nnU-Net<\/strong>\uff1a\u4e0d\u662f encoder-decoder \u4ece\u5934\u8bad\u7ec3\u8def\u7ebf\uff1b\u53ef\u4f5c\u4e3a\u4f4e\u6807\u6ce8\u573a\u666f\u4e0b\u7684\u66ff\u4ee3 baseline \u6216 feature probing \u65b9\u6cd5\u3002nnU-Net \u4ecd\u662f fully supervised setting \u7684\u5f3a\u57fa\u7ebf\u3002<\/li>\n<li><strong>\u4e0e MedNeXt<\/strong>\uff1aMedNeXt \u5c5e\u4e8e ConvNeXt-style 3D convolutional segmentation backbone\uff1bDINO-MVR \u4e0d\u5b66\u4e60 3D convolutional hierarchy\uff0c\u4e8c\u8005\u5b9a\u4f4d\u4e0d\u540c\u3002<\/li>\n<li><strong>\u4e0e UNetR \/ Swin-UNETR \/ nnFormer<\/strong>\uff1a\u8fd9\u4e9b\u662f transformer-based 3D medical segmentation framework\uff1bDINO-MVR \u4f7f\u7528 2D ViT feature\uff0c\u5e76\u51bb\u7ed3 backbone\uff0c\u7f3a\u5c11 3D transformer \u5efa\u6a21\u3002<\/li>\n<li><strong>\u4e0e TransUNet \/ TransFuse<\/strong>\uff1aTransUNet \u7aef\u5230\u7aef\u7ed3\u5408 CNN \/ Transformer\uff1bDINO-MVR \u66f4\u50cf foundation feature readout + test-time fusion\u3002<\/li>\n<li><strong>\u4e0e Mamba \/ VMamba \/ SegMamba \/ DAMamba<\/strong>\uff1a\u6ca1\u6709\u72b6\u6001\u7a7a\u95f4\u6a21\u5757\uff1b\u5bf9 Mamba \u7814\u7a76\u7684\u542f\u53d1\u5728\u4e8e\u201c\u591a\u5c3a\u5ea6\u9884\u6d4b\u878d\u5408\u201d\u548c\u201c\u51bb\u7ed3 foundation feature + \u5c0f decoder\u201d\u53ef\u4e0e Mamba decoder \u6216 DAMamba block \u7ed3\u5408\u3002<\/li>\n<li><strong>\u4e0e medical segmentation foundation model<\/strong>\uff1a\u76f8\u6bd4 SAM \/ MedSAM\uff0cDINO-MVR \u4e0d\u4f9d\u8d56 prompt\uff0c\u800c\u662f\u8bad\u7ec3\u81ea\u52a8\u5206\u5272 probe\uff1b\u76f8\u6bd4 SuPreM \u7b49 3D supervised pretraining\uff0c\u5b83\u4f9d\u8d56\u81ea\u7136\u56fe\u50cf DINOv3 dense representation\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u5bf9\u7528\u6237\u7684 polyp segmentation\u3001DAMamba \u6539\u9020\u548c\u533b\u5b66\u56fe\u50cf\u5206\u5272\u6846\u67b6\u9009\u62e9\uff0c\u8fd9\u7bc7\u8bba\u6587\u6709\u8f83\u9ad8\u53c2\u8003\u4ef7\u503c\uff1a<\/p>\n<ul>\n<li><strong>polyp segmentation<\/strong>\uff1aKvasir-SEG \u662f\u4e3b\u8981\u5b9e\u9a8c\u4e4b\u4e00\uff0cDINO-MVR \u7684 readout \/ TTA \/ entropy fusion \u53ef\u76f4\u63a5\u4f5c\u4e3a polyp \u4f4e\u6807\u6ce8\u6216 frozen encoder baseline\u3002<\/li>\n<li><strong>DAMamba \u6539\u9020<\/strong>\uff1a\u4e0d\u76f4\u63a5\u63d0\u4f9b Mamba \u6a21\u5757\uff0c\u4f46\u53ef\u501f\u9274\u5176\u591a\u5206\u8fa8\u7387\u9884\u6d4b\u3001entropy-guided fusion\u3001test-time view aggregation\uff0c\u628a\u8fd9\u4e9b\u4f5c\u4e3a DAMamba decoder \u8f93\u51fa\u5c42\u6216 ensemble \u7b56\u7565\u3002<\/li>\n<li><strong>baseline \u4ef7\u503c<\/strong>\uff1a\u9002\u5408\u4f5c\u4e3a foundation-feature-readout baseline\uff0c\u5c24\u5176\u7528\u4e8e\u8bf4\u660e\u201c\u53ea\u8bad\u7ec3\u8f7b\u91cf readout \u7684\u4f4e\u6807\u6ce8\u4e0a\u9650\u201d\u3002<\/li>\n<li><strong>related work \u4ef7\u503c<\/strong>\uff1a\u53ef\u653e\u5728 foundation model \/ frozen self-supervised representation \/ annotation-efficient segmentation \u76f8\u5173\u5de5\u4f5c\u4e2d\u3002<\/li>\n<li><strong>\u590d\u73b0\u4ef7\u503c<\/strong>\uff1a\u5982\u679c\u4ee3\u7801\u672a\u5f00\u6e90\uff0c\u590d\u73b0\u4ecd\u53ef\u884c\uff0c\u56e0\u4e3a\u65b9\u6cd5\u76f8\u5bf9\u7b80\u5355\uff1b\u5173\u952e\u5728\u4e8e DINOv3 feature extraction\u3001\u5206\u8fa8\u7387\u8bbe\u7f6e\u3001TTA\u3001CRF \u4e0e z smoothing \u7ec6\u8282\u3002<\/li>\n<\/ul>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5f3a\u70c8\u5efa\u8bae\u7cbe\u8bfb\u3002<\/strong><\/p>\n<p>\u7406\u7531\uff1a\u8fd9\u7bc7\u8bba\u6587\u4e0e\u201cfoundation model for medical segmentation\u201d\u201cannotation-efficient segmentation\u201d\u201cpolyp segmentation\u201d\u201c\u8f7b\u91cf readout \/ decoder \u8bbe\u8ba1\u201d\u9ad8\u5ea6\u76f8\u5173\uff0c\u800c\u4e14\u65b9\u6cd5\u673a\u5236\u6e05\u695a\u3001\u6d88\u878d\u6bd4\u8f83\u5b8c\u6574\u3001\u9650\u5236\u5199\u5f97\u76f8\u5bf9\u8bda\u5b9e\u3002\u5efa\u8bae\u4f18\u5148\u9605\u8bfb Method \u4e0e Experiments\uff0c\u5c24\u5176\u662f\u6700\u540e 3 \u5c42\u7279\u5f81 concat\u3001MLP probe\u3001\u591a\u5206\u8fa8\u7387\u71b5\u878d\u5408\u548c Kvasir ablation\u3002\u82e5\u7528\u6237\u8ba1\u5212\u505a DAMamba \u6216 polyp segmentation\uff0c\u53ef\u4ee5\u628a\u5b83\u4f5c\u4e3a\u4e00\u4e2a\u975e Mamba \u4f46\u5f88\u6709\u53c2\u8003\u4ef7\u503c\u7684 frozen-foundation baseline\u3002<\/p>\n<hr \/>\n<h2>\u8bba\u6587 2\uff1aTopology-Constrained Quantized nnUNet for Efficient and Anatomically Accurate 3D Tooth Segmentation<\/h2>\n<h3>\u57fa\u672c\u4fe1\u606f<\/h3>\n<ul>\n<li>\u6807\u9898\uff1aTopology-Constrained Quantized nnUNet for Efficient and Anatomically Accurate 3D Tooth Segmentation<\/li>\n<li>\u4f5c\u8005 \/ \u7b2c\u4e00\u4f5c\u8005\uff1aPaarth Prasad, Ruchika Malhotra \/ \u7b2c\u4e00\u4f5c\u8005 Paarth Prasad<\/li>\n<li>\u65f6\u95f4\uff1a2026-05-05<\/li>\n<li>\u6765\u6e90\uff1aarXiv preprint<\/li>\n<li>\u8bba\u6587\u9875\u9762\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.04201<\/li>\n<li>PDF \u6587\u4ef6 \/ PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.04201v1<\/li>\n<li>\u4ee3\u7801\u94fe\u63a5\uff1a\u672a\u83b7\u53d6\uff1bPDF \u4e0e arXiv \u5143\u6570\u636e\u4e2d\u672a\u786e\u8ba4\u5b98\u65b9\u4ee3\u7801<\/li>\n<li>\u4efb\u52a1\uff1a3D tooth segmentation from CBCT\uff1b32 tooth classes + background<\/li>\n<li>\u6570\u636e\u96c6\uff1a\u8bba\u6587\u79f0\u4f7f\u7528 public dental CBCT dataset\uff0c200 scans\uff0c\u5f15\u7528 3DTeethSeg\u201922 \/ 3D teeth scan segmentation and labeling challenge\uff1b\u4f46\u6570\u636e\u63cf\u8ff0\u4e0e\u5f15\u7528\u9700\u8fdb\u4e00\u6b65\u6838\u67e5<\/li>\n<li>\u65b9\u6cd5\u7c7b\u578b\uff1annU-Net \u6539\u8fdb\uff1b8-bit quantization-aware training\uff1btopological loss\uff1b3D medical segmentation\uff1bmodel compression \/ deployment<\/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\u4e3b\u8981\u4ef7\u503c\u5728\u4e8e\u63d0\u51fa\u201c\u91cf\u5316 nnU-Net \u65f6\u4e0d\u4ec5\u4fdd\u6301 Dice\uff0c\u8fd8\u8981\u663e\u5f0f\u4fdd\u6301\u7259\u9f7f\u62d3\u6251\u7ed3\u6784\u201d\u7684\u65b9\u5411\uff0c\u4f46\u5f53\u524d\u8bba\u6587\u7684\u5b9e\u9a8c\u4e0e\u65b9\u6cd5\u7ec6\u8282\u53ef\u4fe1\u5ea6\u4e0d\u8db3\uff0c\u66f4\u9002\u5408\u4f5c\u4e3a\u62d3\u6251\u7ea6\u675f\u548c\u90e8\u7f72\u538b\u7f29\u601d\u8def\u7684\u53c2\u8003\uff0c\u800c\u4e0d\u5efa\u8bae\u76f4\u63a5\u4f5c\u4e3a\u5f3a\u8bc1\u636e baseline\u3002<\/p>\n<h4>2. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u95ee\u9898<\/h4>\n<p>\u8bba\u6587\u7814\u7a76 3D CBCT \u7259\u9f7f\u5206\u5272\u3002\u7259\u9f7f\u5206\u5272\u7528\u4e8e\u6b63\u7578\u89c4\u5212\u3001\u79cd\u690d\u3001\u6cd5\u533b\u5b66\u8bc6\u522b\u7b49\u573a\u666f\uff1b\u8fd9\u7c7b\u4efb\u52a1\u4e0d\u4ec5\u8981\u6c42 voxel-level overlap \u9ad8\uff0c\u8fd8\u8981\u6c42\u7259\u9f7f\u6570\u91cf\u3001\u76f8\u90bb\u7259\u5173\u7cfb\u548c\u5185\u90e8\u7ed3\u6784\u5408\u7406\u3002<\/p>\n<p>\u4f5c\u8005\u5173\u6ce8\u7684\u95ee\u9898\u662f\uff1a<strong>nnU-Net \u5728 3D tooth segmentation \u4e0a\u6548\u679c\u597d\uff0c\u4f46 full precision \u6a21\u578b\u5927\u3001\u63a8\u7406\u6162\uff1b\u666e\u901a 8-bit \u91cf\u5316\u867d\u7136\u51cf\u5c0f\u6a21\u578b\u548c\u52a0\u901f\u63a8\u7406\uff0c\u5374\u53ef\u80fd\u7834\u574f\u7259\u9f7f\u62d3\u6251\u7ed3\u6784<\/strong>\u3002\u4f8b\u5982\uff1a<\/p>\n<ul>\n<li>\u7259\u9f7f\u5b9e\u4f8b\u65ad\u88c2\uff1b<\/li>\n<li>\u76f8\u90bb\u7259\u9519\u8bef\u7c98\u8fde\uff1b<\/li>\n<li>\u7259\u4f53\u5185\u90e8\u51fa\u73b0 spurious cavity\uff1b<\/li>\n<li>tooth count \u4e0d\u6b63\u786e\u3002<\/li>\n<\/ul>\n<p>\u56e0\u6b64\u4f5c\u8005\u8bd5\u56fe\u628a topology-specific constraints \u653e\u5165 quantization-aware training\uff0c\u4f7f\u91cf\u5316\u540e\u7684 nnU-Net \u4ecd\u4fdd\u6301\u89e3\u5256\u5408\u7406\u6027\u3002<\/p>\n<h4>3. \u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3<\/h4>\n<p>\u8bba\u6587\u8ba4\u4e3a\u73b0\u6709\u65b9\u6cd5\u4e3b\u8981\u6709\u4e09\u7c7b\u4e0d\u8db3\uff1a<\/p>\n<ul>\n<li><strong>\u6807\u51c6 nnU-Net \/ dental segmentation \u65b9\u6cd5<\/strong>\uff1a\u7cbe\u5ea6\u8f83\u9ad8\uff0c\u4f46\u6a21\u578b\u5927\uff0c3D volume \u63a8\u7406\u6210\u672c\u9ad8\uff0c\u4e0d\u5229\u4e8e\u8d44\u6e90\u53d7\u9650\u4e34\u5e8a\u90e8\u7f72\u3002<\/li>\n<li><strong>\u666e\u901a post-training quantization \/ QAT<\/strong>\uff1a\u80fd\u628a\u6a21\u578b\u538b\u7f29\u5230 8-bit\uff0c\u63d0\u9ad8 CPU \/ edge device \u63a8\u7406\u6548\u7387\uff0c\u4f46\u4f18\u5316\u76ee\u6807\u4e3b\u8981\u662f\u6570\u503c\u8bef\u5dee\u6216 Dice\uff0c\u4e0d\u4fdd\u8bc1 topology\u3002<\/li>\n<li><strong>\u4e00\u822c topology-aware segmentation loss<\/strong>\uff1a\u591a\u7528\u4e8e full precision \u6a21\u578b\uff0c\u672a\u4e13\u95e8\u5904\u7406\u91cf\u5316\u5f15\u5165\u7684 spatial discontinuity\uff0c\u4e5f\u6ca1\u6709\u7259\u9f7f\u7279\u5f02\u7684 count \/ adjacency \/ cavity \u8bbe\u8ba1\u3002<\/li>\n<\/ul>\n<p>\u8bba\u6587\u7684 main move \u662f\u628a dental topology invariants \u663e\u5f0f\u5199\u6210 loss\uff0c\u5e76\u548c QAT loss \u4e00\u8d77\u4f18\u5316\u3002<\/p>\n<h4>4. \u65b9\u6cd5\u603b\u89c8<\/h4>\n<p>\u6574\u4f53\u6846\u67b6\u662f\u4e00\u4e2a <strong>8-bit quantized nnU-Net + topology loss<\/strong>\u3002<\/p>\n<p><strong>Backbone\uff1a<\/strong><\/p>\n<ul>\n<li>\u4f7f\u7528 nnU-Net v2 \u4f5c\u4e3a\u57fa\u7840\uff1b<\/li>\n<li>\u4ecd\u662f 3D U-Net encoder-decoder\uff1b<\/li>\n<li>\u5305\u542b skip connection\u3001\u5377\u79ef\u5c42\u3001instance normalization\u3001LeakyReLU\u3001deep supervision \u7b49 nnU-Net \u5e38\u89c4\u7ec4\u4ef6\uff1b<\/li>\n<li>\u4e0d\u6539\u53d8\u539f\u59cb nnU-Net \u67b6\u6784\uff0c\u53ea\u66ff\u6362\u4e3a\u91cf\u5316\u611f\u77e5\u8bad\u7ec3\u540e\u7684 8-bit \u6743\u91cd\u548c\u6fc0\u6d3b\u3002<\/li>\n<\/ul>\n<p><strong>\u91cf\u5316\uff1a<\/strong><\/p>\n<ul>\n<li>\u4f7f\u7528 8-bit symmetric uniform quantization\uff1b<\/li>\n<li>weights \u548c activations \u5747\u91cf\u5316\uff1b<\/li>\n<li>\u5377\u79ef\u6743\u91cd per-channel quantization\uff1b<\/li>\n<li>activation per-tensor quantization\uff1b<\/li>\n<li>\u4f7f\u7528 QAT\uff0c\u5728 forward \u4e2d\u6a21\u62df\u91cf\u5316\uff0c\u5728 backward \u4e2d\u7528 straight-through estimator \u8fd1\u4f3c rounding \u68af\u5ea6\uff1b<\/li>\n<li>\u91cf\u5316\u51fd\u6570\u5927\u81f4\u4e3a <code>Q(x)=round(x\/s)*s<\/code>\u3002<\/li>\n<\/ul>\n<p><strong>\u62d3\u6251\u7ea6\u675f\uff1a<\/strong><\/p>\n<p>\u4f5c\u8005\u5b9a\u4e49\u603b\u62d3\u6251\u635f\u5931\uff1a<\/p>\n<p><code>L_topo = \u03bb1 L_count + \u03bb2 L_adj + \u03bb3 L_hole<\/code><\/p>\n<p>\u5176\u4e2d\uff1a<\/p>\n<ul>\n<li><code>L_count<\/code>\uff1a\u60e9\u7f5a\u9884\u6d4b connected component \u6570\u91cf\u4e0e ground truth \u6570\u91cf\u4e0d\u4e00\u81f4\uff1b<\/li>\n<li><code>L_adj<\/code>\uff1a\u60e9\u7f5a\u9884\u6d4b\u7259\u9f7f\u90bb\u63a5\u5173\u7cfb \/ \u8fb9\u754c\u4e0e ground truth \u90bb\u63a5\u4e0d\u4e00\u81f4\uff1b<\/li>\n<li><code>L_hole<\/code>\uff1a\u4f7f\u7528 Betti number \/ persistent homology \u601d\u8def\u60e9\u7f5a\u7259\u9f7f\u5185\u90e8 spurious holes\u3002<\/li>\n<\/ul>\n<p><strong>\u603b\u635f\u5931\uff1a<\/strong><\/p>\n<p><code>L_total = L_CE + \u03b1 L_quant + \u03b2 L_topo<\/code><\/p>\n<p>\u4e5f\u5c31\u662f cross-entropy segmentation loss + quantization regularization + topology loss\u3002<\/p>\n<p>\u63a8\u7406\u9636\u6bb5\u4e0d\u989d\u5916\u505a\u62d3\u6251\u540e\u5904\u7406\uff1b\u4f5c\u8005\u58f0\u79f0\u62d3\u6251\u7ea6\u675f\u5df2\u9690\u5f0f\u7f16\u7801\u5230\u91cf\u5316\u6743\u91cd\u4e2d\uff0c\u56e0\u6b64\u63a8\u7406\u4ecd\u662f\u5355\u6b21 quantized nnU-Net forward pass\u3002<\/p>\n<h4>5. \u6838\u5fc3\u6a21\u5757\u62c6\u89e3<\/h4>\n<p><strong>\u6a21\u5757 1\uff1a8-bit QAT nnU-Net<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\uff1a3D CBCT volume\uff1b<\/li>\n<li>\u8f93\u51fa\uff1a\u591a\u7c7b\u522b tooth segmentation mask\uff1b<\/li>\n<li>\u4f5c\u7528\uff1a\u5728\u4fdd\u7559 nnU-Net \u5206\u5272\u80fd\u529b\u7684\u540c\u65f6\u964d\u4f4e\u6a21\u578b\u5927\u5c0f\u548c\u63a8\u7406\u65f6\u95f4\uff1b<\/li>\n<li>\u521b\u65b0\u6027\u5224\u65ad\uff1aQAT \u672c\u8eab\u4e0d\u662f\u65b0\u65b9\u6cd5\uff0c\u4f46\u548c dental topology loss \u7ed3\u5408\u662f\u8be5\u8bba\u6587\u7684\u4e3b\u8981\u5356\u70b9\u3002<\/li>\n<li>\u5bf9 3D medical segmentation\uff1a\u65b9\u5411\u6709\u7528\uff0c\u5c24\u5176\u9002\u5408 clinical deployment \/ edge inference\u3002<\/li>\n<li>\u5bf9 polyp segmentation\uff1a\u76f4\u63a5\u4ef7\u503c\u6709\u9650\uff0c\u56e0\u4e3a polyp \u662f 2D \/ endoscopy foreground mask\uff0c\u4e0d\u5177\u5907\u7259\u9f7f count \/ adjacency \u8fd9\u79cd\u5f3a\u62d3\u6251\u7ed3\u6784\u3002<\/li>\n<\/ul>\n<p><strong>\u6a21\u5757 2\uff1aTooth count loss<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\uff1a\u9884\u6d4b\u6982\u7387\u56fe\u4e0e ground truth mask\uff1b<\/li>\n<li>\u8f93\u51fa\uff1aconnected component count \u5dee\u5f02\uff1b<\/li>\n<li>\u89e3\u51b3\u95ee\u9898\uff1amissing tooth \/ extra segment \/ fragmented tooth\uff1b<\/li>\n<li>\u521b\u65b0\u6027\u5224\u65ad\uff1a\u6982\u5ff5\u5408\u7406\uff0c\u4f46\u8bba\u6587\u4e2d\u201cdifferentiable thresholding + connected component count\u201d\u7684\u5b9e\u73b0\u7ec6\u8282\u4e0d\u591f\u5145\u5206\uff0c\u771f\u5b9e\u53ef\u5fae\u6027\u548c\u7a33\u5b9a\u6027\u9700\u8981\u4ee3\u7801\u9a8c\u8bc1\u3002<\/li>\n<li>\u53ef\u8fc1\u79fb\u6027\uff1a\u53ef\u8fc1\u79fb\u5230 vertebrae\u3001rib\u3001multi-organ instance count \u7b49\u4efb\u52a1\u3002<\/li>\n<\/ul>\n<p><strong>\u6a21\u5757 3\uff1aAdjacency consistency loss<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\uff1a\u6bcf\u4e2a\u7259\u9f7f\u7c7b\u522b\u7684\u9884\u6d4b\u533a\u57df\u548c ground truth \u90bb\u63a5\u56fe\uff1b<\/li>\n<li>\u8f93\u51fa\uff1a\u90bb\u63a5\u8fb9\u754c\u5dee\u5f02\uff1b<\/li>\n<li>\u89e3\u51b3\u95ee\u9898\uff1a\u76f8\u90bb\u7259\u7c98\u8fde\u6216\u9519\u8bef\u65ad\u5f00\uff1b<\/li>\n<li>\u521b\u65b0\u6027\u5224\u65ad\uff1a\u533b\u5b66\u7ed3\u6784\u7ea6\u675f\u6709\u610f\u4e49\uff0c\u4f46\u5b9e\u73b0\u4f9d\u8d56 ground-truth adjacency \u4e0e\u8fb9\u754c\u63d0\u53d6\uff0c\u8bba\u6587\u6ca1\u6709\u5145\u5206\u8bf4\u660e\u590d\u6742\u75c5\u4f8b\u3001\u7f3a\u7259\u3001implant\u3001\u5f02\u5e38\u6392\u5217\u65f6\u5982\u4f55\u5904\u7406\u3002<\/li>\n<li>\u8fc1\u79fb\u6027\uff1a\u5bf9 spine \/ vessel \/ dental \/ airway \u7b49\u7ed3\u6784\u66f4\u6709\u7528\uff1b\u5bf9 polyp \u5206\u5272\u57fa\u672c\u4e0d\u9002\u7528\u3002<\/li>\n<\/ul>\n<p><strong>\u6a21\u5757 4\uff1aCavity \/ Betti number loss<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\uff1a\u6bcf\u4e2a tooth segment \u7684\u9884\u6d4b mask\uff1b<\/li>\n<li>\u8f93\u51fa\uff1a\u9884\u6d4b\u4e0e\u771f\u503c\u4e00\u9636 Betti number \u5dee\u5f02\uff1b<\/li>\n<li>\u89e3\u51b3\u95ee\u9898\uff1a\u91cf\u5316\u5bfc\u81f4\u7684\u5185\u90e8\u5b54\u6d1e\u6216\u7ed3\u6784\u7834\u574f\uff1b<\/li>\n<li>\u521b\u65b0\u6027\u5224\u65ad\uff1atopological data analysis \u5728 medical segmentation \u4e2d\u5df2\u6709\u57fa\u7840\uff1b\u672c\u6587\u7684\u7259\u9f7f\u7279\u5f02\u7ec4\u5408\u6709\u4e00\u5b9a\u4ef7\u503c\uff0c\u4f46\u201cdifferentiable persistent homology approximation\u201d\u63cf\u8ff0\u504f\u6982\u5ff5\u5316\u3002<\/li>\n<li>\u5bf9 3D segmentation\uff1a\u9002\u5408\u8840\u7ba1\u3001\u6c14\u9053\u3001\u7259\u9f7f\u3001\u690e\u4f53\u7b49\u62d3\u6251\u654f\u611f\u5bf9\u8c61\uff0c\u4f46\u8bad\u7ec3\u6210\u672c\u548c\u68af\u5ea6\u7a33\u5b9a\u6027\u9700\u8981\u4e25\u67e5\u3002<\/li>\n<\/ul>\n<p><strong>\u6a21\u5757 5\uff1aImplicit topology at inference<\/strong><\/p>\n<ul>\n<li>\u8f93\u5165\uff1a\u8bad\u7ec3\u597d\u7684 quantized nnU-Net\uff1b<\/li>\n<li>\u8f93\u51fa\uff1a\u65e0\u9700\u540e\u5904\u7406\u7684 mask\uff1b<\/li>\n<li>\u4f5c\u7528\uff1a\u4fdd\u6301\u63a8\u7406\u901f\u5ea6\uff1b<\/li>\n<li>\u53ef\u4fe1\u5ea6\u5224\u65ad\uff1a\u8fd9\u4e2a\u76ee\u6807\u5408\u7406\uff0c\u4f46\u8bba\u6587\u8bc1\u636e\u4e0d\u8db3\u4ee5\u5b8c\u5168\u8bc1\u660e\u201c\u62d3\u6251\u7ea6\u675f\u88ab\u53ef\u9760\u7f16\u7801\u8fdb\u6743\u91cd\u201d\uff0c\u5c24\u5176\u7f3a\u5c11\u5916\u90e8\u6570\u636e\u3001\u8de8\u4e2d\u5fc3\u6d4b\u8bd5\u548c\u4ee3\u7801\u3002<\/li>\n<\/ul>\n<h4>6. \u5b9e\u9a8c\u8bbe\u8ba1\u4e0e\u7ed3\u679c<\/h4>\n<p><strong>\u6570\u636e\u96c6\uff1a<\/strong><\/p>\n<p>\u8bba\u6587\u79f0\u4f7f\u7528 200 \u4e2a\u9ad8\u5206\u8fa8\u7387 CBCT scans\uff0c32 tooth classes\uff0c\u5e73\u5747\u4f53\u79ef 512\u00d7512\u00d7256\uff0c0.3mm isotropic spacing\uff0c\u5e76\u5212\u5206\u4e3a\uff1a<\/p>\n<ul>\n<li>train\uff1a140 scans\uff1b<\/li>\n<li>validation\uff1a30 scans\uff1b<\/li>\n<li>test\uff1a30 scans\u3002<\/li>\n<\/ul>\n<p>\u6570\u636e\u96c6\u88ab\u79f0\u4e3a public dental CBCT dataset\uff0c\u5e76\u5f15\u7528 3DTeethSeg\u201922 challenge\u3002\u4f46\u8bba\u6587\u6b63\u6587\u4e2d\u7684\u6570\u636e\u96c6\u63cf\u8ff0\u3001\u5f15\u7528\u548c\u5b9e\u9645 3DTeethSeg \u6570\u636e\u89c4\u683c\u9700\u8981\u8fdb\u4e00\u6b65\u6838\u5bf9\uff0c\u56e0\u6b64\u8fd9\u91cc\u4e0d\u628a\u6570\u636e\u8bbe\u5b9a\u89c6\u4e3a\u5b8c\u5168\u5df2\u9a8c\u8bc1\u4e8b\u5b9e\u3002<\/p>\n<p><strong>\u8bad\u7ec3\u8bbe\u7f6e\uff1a<\/strong><\/p>\n<ul>\n<li>PyTorch 1.10\uff1b<\/li>\n<li>nnU-Net v2\uff1b<\/li>\n<li>Adam\uff0c\u521d\u59cb learning rate <code>3e-4<\/code>\uff1b<\/li>\n<li>batch size 2\uff1b<\/li>\n<li>1000 epochs\uff1b<\/li>\n<li>early stopping by validation Dice\uff1b<\/li>\n<li>A100 40GB \u8bad\u7ec3\uff1b<\/li>\n<li>Intel Core i9-10900K CPU \u4e0a\u7528 ONNX Runtime \u6d4b\u91cf integer-only inference\uff1b<\/li>\n<li>random seed 42\u3002<\/li>\n<\/ul>\n<p><strong>\u8bc4\u4ef7\u6307\u6807\uff1a<\/strong><\/p>\n<ul>\n<li>segmentation accuracy\uff1aDSC\u3001IoU\u3001Boundary F1\uff1b<\/li>\n<li>topological fidelity\uff1aTooth Count Accuracy, Adjacency Consistency Score, Cavity Error Rate\uff1b<\/li>\n<li>efficiency\uff1amodel size\u3001inference time\u3001MACs\u3002<\/li>\n<\/ul>\n<p><strong>baseline\uff1a<\/strong><\/p>\n<ul>\n<li>Full-Precision nnUNet\uff1b<\/li>\n<li>Post-Training Quantized nnUNet\uff1b<\/li>\n<li>QAT-nnUNet\uff1b<\/li>\n<li>TopoNet adapted for dental data\u3002<\/li>\n<\/ul>\n<p><strong>\u4e3b\u8981\u7ed3\u679c\uff1a<\/strong><\/p>\n<p>\u8bba\u6587\u62a5\u544a\uff1a<\/p>\n<ul>\n<li>Full-Precision nnUNet\uff1aDSC 92.3%\uff0csize 1024MB\uff0cCPU time 8.2s\uff1b<\/li>\n<li>Post-Training Quant\uff1aDSC 88.7%\uff0csize 256MB\uff0ctime 2.1s\uff1b<\/li>\n<li>QAT-nnUNet\uff1aDSC 90.1%\uff0csize 256MB\uff0ctime 2.3s\uff1b<\/li>\n<li>TopoNet\uff1aDSC 91.8%\uff0csize 896MB\uff0ctime 7.5s\uff1b<\/li>\n<li>Proposed\uff1aDSC 91.5%\uff0csize 256MB\uff0ctime 2.4s\uff0cTCA 93.8%\uff0cACS 91.0%\uff0cCER 3.9%\u3002<\/li>\n<\/ul>\n<p><strong>\u6d88\u878d\uff1a<\/strong><\/p>\n<p>\u4ece QAT-only \u5230\u52a0\u5165 count \/ adjacency \/ cavity \/ full topology loss\uff1a<\/p>\n<ul>\n<li>QAT-only\uff1aDSC 90.1\uff0cTCA 85.4\uff0cACS 83.7\uff0cCER 9.5\uff1b<\/li>\n<li>\n<ul>\n<li>Count Loss\uff1aTCA \u63d0\u5347\u5230 89.2\uff1b<\/li>\n<\/ul>\n<\/li>\n<li>\n<ul>\n<li>Adjacency Loss\uff1aACS \u63d0\u5347\u5230 89.3\uff1b<\/li>\n<\/ul>\n<\/li>\n<li>\n<ul>\n<li>Cavity Loss\uff1aCER \u964d\u5230 5.1\uff1b<\/li>\n<\/ul>\n<\/li>\n<li>Full Topo Loss\uff1aDSC 91.5\uff0cTCA 93.8\uff0cACS 91.0\uff0cCER 3.9\u3002<\/li>\n<\/ul>\n<p>\u4f5c\u8005\u636e\u6b64\u8ba4\u4e3a\u4e09\u7c7b\u62d3\u6251\u7ea6\u675f\u6709\u534f\u540c\u4f5c\u7528\u3002<\/p>\n<h4>7. \u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u5224\u65ad<\/h4>\n<p>\u8fd9\u7bc7\u8bba\u6587\u7684\u60f3\u6cd5\u6709\u53c2\u8003\u4ef7\u503c\uff0c\u4f46\u5b9e\u9a8c\u53ef\u4fe1\u5ea6\u9700\u8981\u8c28\u614e\u5bf9\u5f85\u3002<\/p>\n<p><strong>\u76f8\u5bf9\u53ef\u4fe1\u7684\u90e8\u5206\uff1a<\/strong><\/p>\n<ul>\n<li>\u7814\u7a76\u95ee\u9898\u771f\u5b9e\u5b58\u5728\uff1a\u91cf\u5316\u53ef\u80fd\u7834\u574f\u8fb9\u754c\u548c\u7ec6\u7ed3\u6784\uff1b<\/li>\n<li>\u7259\u9f7f\u5206\u5272\u786e\u5b9e\u9700\u8981 count\u3001adjacency\u3001hole \/ cavity \u7b49\u7ed3\u6784\u5408\u7406\u6027\uff1b<\/li>\n<li>\u4e0e nnU-Net\u3001QAT\u3001topology loss \u7684\u7ed3\u5408\u65b9\u5411\u6e05\u695a\uff1b<\/li>\n<li>\u6307\u6807\u8bbe\u8ba1\u8986\u76d6\u4e86 Dice \u4e4b\u5916\u7684\u62d3\u6251\u8d28\u91cf\uff0c\u6bd4\u53ea\u62a5 Dice \u66f4\u6709\u610f\u4e49\u3002<\/li>\n<\/ul>\n<p><strong>\u4e3b\u8981\u95ee\u9898\uff1a<\/strong><\/p>\n<ul>\n<li>PDF \u4e2d\u6ca1\u6709\u786e\u8ba4\u5b98\u65b9\u4ee3\u7801\uff1b<\/li>\n<li>\u6570\u636e\u96c6\u63cf\u8ff0\u8f83\u7b3c\u7edf\uff0c\u867d\u7136\u5f15\u7528 public CBCT dataset \/ 3DTeethSeg\u201922\uff0c\u4f46\u5177\u4f53\u6570\u636e\u7248\u672c\u3001\u4e0b\u8f7d\u6e90\u3001preprocessing \u548c challenge protocol \u4e0d\u591f\u53ef\u6838\u67e5\uff1b<\/li>\n<li>\u62d3\u6251\u635f\u5931\u7684\u53ef\u5fae\u5b9e\u73b0\u63cf\u8ff0\u504f\u6982\u5ff5\u5316\uff0c\u4f8b\u5982 connected component count\u3001Betti number\u3001boundary adjacency \u7684 differentiable approximation \u9700\u8981\u4ee3\u7801\u6216\u66f4\u4e25\u8c28\u9644\u5f55\u652f\u6491\uff1b<\/li>\n<li>\u8868\u683c\u7ed3\u679c\u5f88\u6574\u9f50\uff0c\u4f46\u7f3a\u5c11\u7f6e\u4fe1\u533a\u95f4\u3001\u75c5\u4f8b\u7ea7\u7edf\u8ba1\u3001\u5931\u8d25\u6848\u4f8b\u7ec6\u8282\uff1b<\/li>\n<li>\u867d\u7136\u58f0\u79f0 paired t-test\uff0c\u4f46\u6b63\u6587\u6ca1\u6709\u5c55\u793a p-value \u6216\u7edf\u8ba1\u68c0\u9a8c\u8868\uff1b<\/li>\n<li>\u6ca1\u6709\u591a\u4e2d\u5fc3 \/ \u5916\u90e8\u6d4b\u8bd5\uff0c\u65e0\u6cd5\u786e\u8ba4 clinical robustness\uff1b<\/li>\n<li>\u672a\u4e0e ToothSeg\u3001\u6700\u65b0 dental segmentation SOTA \u6216\u771f\u5b9e nnU-Net strong training recipe \u505a\u5145\u5206\u7edf\u4e00\u6bd4\u8f83\uff1b<\/li>\n<li>\u91cf\u5316\u63a8\u7406 speedup \u5728 CPU \/ Jetson \u4e0a\u7ed9\u51fa\uff0c\u4f46 ONNX \/ integer-only \u6267\u884c\u7ec6\u8282\u4e0d\u8db3\uff0c\u90e8\u7f72\u590d\u73b0\u96be\u5ea6\u8f83\u9ad8\u3002<\/li>\n<\/ul>\n<p>\u56e0\u6b64\uff0c\u8fd9\u7bc7\u8bba\u6587\u7684\u7ed3\u8bba\u9700\u8981\u5f31\u5316\uff1a\u5b83\u652f\u6301\u201c\u62d3\u6251\u7ea6\u675f\u53ef\u80fd\u7f13\u89e3\u91cf\u5316 nnU-Net \u7684\u7ed3\u6784\u9519\u8bef\u201d\u8fd9\u4e2a\u65b9\u5411\uff0c\u4f46\u8fd8\u4e0d\u8db3\u4ee5\u8bc1\u660e\u5176\u65b9\u6cd5\u5df2\u662f\u53ef\u9760 SOTA\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>\u4e0e U-Net \/ nnU-Net<\/strong>\uff1a\u76f4\u63a5\u57fa\u4e8e nnU-Net\uff0c\u662f nnU-Net \u7684 quantization-aware + topology-aware \u53d8\u4f53\uff1b\u6ca1\u6709\u6539\u53d8 encoder-decoder \u4e3b\u4f53\u3002<\/li>\n<li><strong>\u4e0e MedNeXt<\/strong>\uff1aMedNeXt \u662f\u73b0\u4ee3\u5377\u79ef 3D backbone\uff1b\u672c\u6587\u6ca1\u6709\u5f15\u5165 ConvNeXt block\uff0c\u4f46\u201c\u538b\u7f29 + \u62d3\u6251 loss\u201d\u601d\u60f3\u53ef\u8fc1\u79fb\u5230 MedNeXt\u3002<\/li>\n<li><strong>\u4e0e UNetR \/ Swin-UNETR \/ TransUNet<\/strong>\uff1a\u672c\u6587\u4e0d\u662f Transformer-based segmentation\uff1b\u5982\u679c\u76ee\u6807\u662f volumetric context modeling\uff0c\u5b83\u4e0d\u5982 UNETR \/ Swin-UNETR \u63d0\u4f9b\u65b0\u7684\u5efa\u6a21\u7ed3\u6784\u3002<\/li>\n<li><strong>\u4e0e Mamba \/ VMamba \/ SegMamba \/ DAMamba<\/strong>\uff1a\u65e0 Mamba \/ state space \u6a21\u5757\uff1b\u4f46 topology loss \u53ef\u4f5c\u4e3a DAMamba \u6216 SegMamba \u5728 3D anatomical segmentation \u4e2d\u7684\u8f85\u52a9 loss\u3002<\/li>\n<li><strong>\u4e0e foundation model segmentation<\/strong>\uff1a\u4e0d\u662f SAM \/ MedSAM \/ DINO \u7c7b foundation model\uff1b\u5b83\u66f4\u504f\u90e8\u7f72\u5de5\u7a0b\u548c\u7ed3\u6784\u5148\u9a8c\u3002<\/li>\n<li><strong>\u4e0e lightweight segmentation<\/strong>\uff1a\u76f8\u5173\u6027\u5f3a\u3002\u5b83\u5173\u6ce8 8-bit inference\u3001model size\u3001CPU \/ edge \u63a8\u7406\uff0c\u662f lightweight medical segmentation \u7684\u4e00\u6761\u8def\u7ebf\u3002<\/li>\n<\/ul>\n<h4>9. \u5bf9\u6211\u8bfe\u9898\u7684\u4ef7\u503c<\/h4>\n<p>\u5bf9\u7528\u6237\u8bfe\u9898\u7684\u4ef7\u503c\u8981\u5206\u65b9\u5411\u770b\uff1a<\/p>\n<ul>\n<li><strong>polyp segmentation<\/strong>\uff1a\u76f4\u63a5\u4ef7\u503c\u8f83\u4f4e\u3002Polyp segmentation \u66f4\u5173\u5fc3\u8fb9\u754c\u3001\u4f2a\u88c5\u3001domain generalization\u3001\u5b9e\u65f6\u6027\uff0c\u800c\u4e0d\u662f tooth count \/ adjacency\u3002\u53ef\u4ee5\u501f\u9274 boundary \/ topology-inspired loss\uff0c\u4f46\u4e0d\u5b9c\u7167\u642c\u3002<\/li>\n<li><strong>DAMamba \u6539\u9020<\/strong>\uff1a\u4e2d\u7b49\u4ef7\u503c\u3002\u53ef\u628a topology-aware loss \u4f5c\u4e3a DAMamba \u5904\u7406 3D tooth \/ vessel \/ spine \/ multi-instance anatomy \u65f6\u7684 training regularizer\uff1b\u4f46\u4e0d\u662f Mamba block \u7ed3\u6784\u521b\u65b0\u3002<\/li>\n<li><strong>3D medical image segmentation<\/strong>\uff1a\u6709\u4e00\u5b9a\u53c2\u8003\u4ef7\u503c\u3002\u5c24\u5176\u9002\u5408\u4f5c\u4e3a\u201cnnU-Net \u90e8\u7f72\u538b\u7f29 + anatomical constraint\u201d\u7684 related work\u3002<\/li>\n<li><strong>baseline \u4ef7\u503c<\/strong>\uff1a\u4e0d\u5efa\u8bae\u4f5c\u4e3a\u5f3a baseline\uff1b\u5982\u679c\u5f15\u7528\uff0c\u5e94\u6807\u6ce8\u4e3a arXiv preprint\uff0c\u5e76\u5f3a\u8c03\u4ee3\u7801\u548c\u6570\u636e\u534f\u8bae\u672a\u786e\u8ba4\u3002<\/li>\n<li><strong>introduction \/ related work<\/strong>\uff1a\u53ef\u7528\u4e8e\u8bf4\u660e Dice-only optimization \u4e0e deployment compression \u4e0d\u8db3\u4ee5\u4fdd\u8bc1 anatomical fidelity\u3002<\/li>\n<li><strong>\u590d\u73b0\u5b9e\u9a8c\u53c2\u8003<\/strong>\uff1a\u5982\u679c\u7528\u6237\u5173\u5fc3\u91cf\u5316 \/ edge deployment\uff0c\u53ef\u4ee5\u53c2\u8003\u5176 loss \u8bbe\u8ba1\uff1b\u4f46\u590d\u73b0\u524d\u5fc5\u987b\u5148\u786e\u8ba4\u6570\u636e\u96c6\u3001\u4ee3\u7801\u3001QAT \u5b9e\u73b0\u548c\u62d3\u6251 loss \u53ef\u5fae\u5b9e\u73b0\u3002<\/li>\n<\/ul>\n<h4>10. \u9605\u8bfb\u5efa\u8bae<\/h4>\n<p><strong>\u5efa\u8bae\u7565\u8bfb\u5230\u4e2d\u7b49\u7cbe\u8bfb\u3002<\/strong><\/p>\n<p>\u5efa\u8bae\u9605\u8bfb Introduction\u3001Method \u4e2d loss \u8bbe\u8ba1\u3001Experiment tables\uff0c\u4f46\u4e0d\u5efa\u8bae\u628a\u5b83\u5f53\u4f5c\u53ef\u9760 SOTA \u8bba\u6587\u5168\u9762\u7cbe\u8bfb\u3002\u5b83\u9002\u5408\u4f5c\u4e3a\u4e00\u4e2a\u7814\u7a76\u60f3\u6cd5\u6765\u6e90\uff1a\u5982\u4f55\u5728 3D \u533b\u5b66\u5206\u5272\u4e2d\u628a topology-aware regularization \u4e0e quantization-aware training \u7ed3\u5408\u3002\u82e5\u7528\u6237\u8fd1\u671f\u7814\u7a76\u91cd\u70b9\u662f polyp segmentation \u6216 DAMamba backbone \u6539\u9020\uff0c\u4f18\u5148\u7ea7\u4f4e\u4e8e DINO-MVR\uff1b\u82e5\u8f6c\u5411 3D \u7259\u9f7f\u3001\u810a\u67f1\u3001\u8840\u7ba1\u3001\u6c14\u9053\u6216\u8f7b\u91cf\u90e8\u7f72\uff0c\u5219\u503c\u5f97\u8fdb\u4e00\u6b65\u6838\u67e5\u548c\u590d\u73b0\u3002<\/p>\n<hr \/>\n<h2>\u4eca\u65e5\u63a8\u8350\u4f18\u5148\u7ea7<\/h2>\n<ol>\n<li>\n<p><strong>DINO-MVR: Multi-View Readout of Frozen DINOv3 for Annotation-Efficient Medical Segmentation<\/strong><br \/>\n   \u6700\u503c\u5f97\u4f18\u5148\u8bfb\u3002\u5b83\u4e0e foundation model\u3001\u4f4e\u6807\u6ce8\u533b\u5b66\u5206\u5272\u3001polyp segmentation\u3001\u8f7b\u91cf decoder\/readout \u8bbe\u8ba1\u90fd\u6709\u76f4\u63a5\u5173\u7cfb\uff1b\u65b9\u6cd5\u7b80\u5355\u4f46\u95ee\u9898\u5b9a\u4e49\u6e05\u695a\uff0c\u5b9e\u9a8c\u9650\u5236\u4e5f\u5199\u5f97\u76f8\u5bf9\u8bda\u5b9e\u3002<\/p>\n<\/li>\n<li>\n<p><strong>Topology-Constrained Quantized nnUNet for Efficient and Anatomically Accurate 3D Tooth Segmentation<\/strong><br \/>\n   \u4f5c\u4e3a\u201c\u62d3\u6251\u7ea6\u675f + nnU-Net \u91cf\u5316\u90e8\u7f72\u201d\u7684\u601d\u8def\u503c\u5f97\u4fdd\u7559\uff0c\u4f46\u8bc1\u636e\u94fe\u8f83\u5f31\uff0c\u4e0d\u5efa\u8bae\u4f18\u5148\u6295\u5165\u5927\u91cf\u65f6\u95f4\u590d\u73b0\u3002\u66f4\u9002\u5408\u5728\u505a 3D \u7ed3\u6784\u654f\u611f\u5206\u5272\u6216\u6a21\u578b\u538b\u7f29\u65f6\u4f5c\u4e3a\u542f\u53d1\u6027\u53c2\u8003\u3002<\/p>\n<\/li>\n<\/ol>\n<h2>\u4eca\u65e5\u8bba\u6587\u94fe\u63a5\u83b7\u53d6\u60c5\u51b5<\/h2>\n<ul>\n<li>\u8bba\u6587 1\uff1a\u5df2\u9644 PDF\u3002\u540c\u65f6\u63d0\u4f9b arXiv PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.07221v1<\/li>\n<li>\u8bba\u6587 2\uff1a\u5df2\u9644 PDF\u3002\u540c\u65f6\u63d0\u4f9b arXiv PDF \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2605.04201v1<\/li>\n<\/ul>\n<h2>\u4eca\u65e5\u53ef\u6267\u884c\u5efa\u8bae<\/h2>\n<ol>\n<li>\n<p><strong>\u5148\u7cbe\u8bfb DINO-MVR\uff0c\u5e76\u628a\u5b83\u52a0\u5165 foundation model \/ annotation-efficient segmentation \u76f8\u5173\u5de5\u4f5c\u3002<\/strong><br \/>\n   \u91cd\u70b9\u770b frozen DINOv3 last-3-layer feature concat\u3001MLP probe\u3001\u591a\u5206\u8fa8\u7387 inference\u3001entropy-guided fusion \u4e0e Kvasir ablation\u3002\u5b83\u5bf9 polyp segmentation \u548c\u4f4e\u6807\u6ce8\u5b9e\u9a8c\u8bbe\u8ba1\u5f88\u6709\u53c2\u8003\u4ef7\u503c\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u53ef\u4ee5\u5c1d\u8bd5\u628a DINO-MVR \u7684 entropy-guided multi-resolution fusion \u79fb\u690d\u5230\u81ea\u5df1\u7684 U-Net \/ DAMamba decoder \u8f93\u51fa\u7aef\u3002<\/strong><br \/>\n   \u4f8b\u5982\u8bad\u7ec3 DAMamba \u4e3b\u5206\u652f\u540e\uff0c\u5728 512 \/ 1024 \u6216\u591a\u5c3a\u5ea6\u8f93\u51fa\u4e0a\u505a entropy-based branch selection\uff0c\u89c2\u5bdf polyp boundary\u3001HD95 \u6216 mDice \u662f\u5426\u6539\u5584\u3002<\/p>\n<\/li>\n<li>\n<p><strong>Topology-Constrained Quantized nnUNet \u6682\u4e0d\u5efa\u8bae\u4f5c\u4e3a\u5f3a baseline\uff0c\u4f46\u53ef\u4f5c\u4e3a 3D \u5206\u5272\u90e8\u7f72\u4e0e topology loss \u7684 idea bank\u3002<\/strong><br \/>\n   \u5982\u679c\u540e\u7eed\u505a 3D tooth \/ spine \/ vessel \/ airway segmentation\uff0c\u53ef\u4ee5\u501f\u9274 count \/ adjacency \/ cavity loss\uff1b\u82e5\u4ecd\u805a\u7126 polyp segmentation\uff0c\u5219\u53ea\u9700\u7565\u8bfb\u5176 loss \u601d\u60f3\uff0c\u4e0d\u5fc5\u6295\u5165\u590d\u73b0\u3002<\/p>\n<\/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 \u672c\u6587\u4ec5\u4fdd\u7559\u8bba\u6587\u9875\u9762\u4e0e PDF URL\uff0c\u4e0d\u4e0a\u4f20 PDF \u9644\u4ef6\u3002 \u4eca\u65e5\u7ed3\u8bba \u4eca\u5929\u6ca1\u6709\u68c0\u7d22\u5230\u5df2\u786e\u8ba4\u9876\u4f1a &#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-1045","post","type-post","status-publish","format-standard","hentry","category-85"],"_links":{"self":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1045","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=1045"}],"version-history":[{"count":0,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/posts\/1045\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/media?parent=1045"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/categories?post=1045"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.eutaboo.com\/index.php\/wp-json\/wp\/v2\/tags?post=1045"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}