Precise segmentation of teeth in Cone-Beam Computed Tomography (CBCT) scans is a longstanding challenge in computational dentistry, particularly when it comes to fine anatomical structures such as root apices \u2014 whose accurate delineation is essential for evaluating root resorption in orthodontic practice.<\/p>\n\n\n\n
This study presents GEPAR3D, a new method that combines instance detection and multi-class segmentation into a single unified pipeline, specifically designed to improve root-level segmentation accuracy. The approach incorporates a Statistical Shape Model of dentition as a geometric prior, enabling the model to capture anatomical context and morphological consistency without imposing rigid spatial constraints. To handle the complexity of narrow root structures, GEPAR3D employs a deep watershed technique that represents each tooth as a continuous 3D energy basin encoding voxel distances to boundaries \u2014 an instance-aware formulation that proves particularly effective for challenging apex regions.<\/p>\n\n\n\n
Trained on publicly available CBCT data from a single center and evaluated across four external test sets from both in-house and public medical institutions, GEPAR3D demonstrates strong generalization. The method achieves a mean Dice Similarity Coefficient of 95.0% \u2014 2.8 percentage points above the next best approach \u2014 and a recall of 95.2%, representing a 9.5-point improvement. Qualitative results further confirm meaningful gains in root segmentation quality, pointing to real clinical potential for more reliable root resorption assessment and improved orthodontic decision support. Implementation and dataset are publicly available.<\/p>\n\n\n\n
<\/div>\n\n\n\n
Autors<\/strong>: Tomasz Szczepa\u0144ski, SzymonP\u0142otka,\u00a0Michal K.\u00a0Grzeszczyk, Arleta\u00a0Adamowicz,\u00a0Piotr\u00a0Fudalej, Przemys\u0142aw\u00a0Korzeniowski,\u00a0Tomasz\u00a0Trzci\u0144ski,\u00a0 Arkadiusz\u00a0Sitek<\/p>\n\n\n\n<\/div>\n\n\n","protected":false},"excerpt":{"rendered":"
MICCAI, 2025<\/p>\n","protected":false},"featured_media":0,"template":"","research_type":[8],"research_team":[17],"class_list":["post-31931","research","type-research","status-publish","hentry","research_type-publications","research_team-health-informatics-group-higs"],"yoast_head":"\n
GEPAR3D: Geometry Prior-Assisted Learning for 3D Tooth Segmentation - Centre for Computational Personalized Medicine<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/sano.science\/research\/gepar3d-geometry-prior-assisted-learning-for-3d-tooth-segmentation\/\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"GEPAR3D: Geometry Prior-Assisted Learning for 3D Tooth Segmentation\" \/>\n<meta property=\"og:description\" content=\"MICCAI, 2025\" \/>\n<meta property=\"og:url\" content=\"https:\/\/sano.science\/research\/gepar3d-geometry-prior-assisted-learning-for-3d-tooth-segmentation\/\" \/>\n<meta property=\"og:site_name\" content=\"Centre for Computational Personalized Medicine\" \/>\n<meta property=\"article:publisher\" content=\"https:\/\/www.facebook.com\/sano.science\/\" \/>\n<meta property=\"article:modified_time\" content=\"2026-07-15T09:44:54+00:00\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:site\" content=\"@sanoscience\" \/>\n<meta name=\"twitter:label1\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data1\" content=\"2 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/sano.science\\\/research\\\/gepar3d-geometry-prior-assisted-learning-for-3d-tooth-segmentation\\\/\",\"url\":\"https:\\\/\\\/sano.science\\\/research\\\/gepar3d-geometry-prior-assisted-learning-for-3d-tooth-segmentation\\\/\",\"name\":\"GEPAR3D: Geometry Prior-Assisted Learning for 3D Tooth Segmentation - Centre for Computational Personalized Medicine\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/sano.science\\\/#website\"},\"datePublished\":\"2026-07-15T09:44:06+00:00\",\"dateModified\":\"2026-07-15T09:44:54+00:00\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/sano.science\\\/research\\\/gepar3d-geometry-prior-assisted-learning-for-3d-tooth-segmentation\\\/#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/sano.science\\\/research\\\/gepar3d-geometry-prior-assisted-learning-for-3d-tooth-segmentation\\\/\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/sano.science\\\/research\\\/gepar3d-geometry-prior-assisted-learning-for-3d-tooth-segmentation\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/sano.science\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Research\",\"item\":\"https:\\\/\\\/sano.science\\\/research\\\/\"},{\"@type\":\"ListItem\",\"position\":3,\"name\":\"Publications\",\"item\":\"https:\\\/\\\/sano.science\\\/research-type\\\/publications\\\/\"},{\"@type\":\"ListItem\",\"position\":4,\"name\":\"GEPAR3D: Geometry Prior-Assisted Learning for 3D Tooth Segmentation\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/sano.science\\\/#website\",\"url\":\"https:\\\/\\\/sano.science\\\/\",\"name\":\"Centre for Computational Personalized Medicine\",\"description\":\"Sano \u2013 Centre for Computational Medicine\",\"publisher\":{\"@id\":\"https:\\\/\\\/sano.science\\\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/sano.science\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Organization\",\"@id\":\"https:\\\/\\\/sano.science\\\/#organization\",\"name\":\"Sano \u2013 Centre for Computational Medicine\",\"alternateName\":\"Sano\",\"url\":\"https:\\\/\\\/sano.science\\\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/sano.science\\\/#\\\/schema\\\/logo\\\/image\\\/\",\"url\":\"https:\\\/\\\/sano.science\\\/wp-content\\\/uploads\\\/2024\\\/05\\\/logo_sano_podstawowe.png\",\"contentUrl\":\"https:\\\/\\\/sano.science\\\/wp-content\\\/uploads\\\/2024\\\/05\\\/logo_sano_podstawowe.png\",\"width\":700,\"height\":265,\"caption\":\"Sano \u2013 Centre for Computational Medicine\"},\"image\":{\"@id\":\"https:\\\/\\\/sano.science\\\/#\\\/schema\\\/logo\\\/image\\\/\"},\"sameAs\":[\"https:\\\/\\\/www.facebook.com\\\/sano.science\\\/\",\"https:\\\/\\\/x.com\\\/sanoscience\",\"https:\\\/\\\/www.linkedin.com\\\/company\\\/sanoscience\\\/\",\"https:\\\/\\\/www.youtube.com\\\/channel\\\/UCDZ_8TcjMWUG2ZcgKKgfpwQ\",\"https:\\\/\\\/bsky.app\\\/profile\\\/sanoscience.bsky.social\"]}]}<\/script>\n<!-- \/ Yoast SEO Premium plugin. -->","yoast_head_json":{"title":"GEPAR3D: Geometry Prior-Assisted Learning for 3D Tooth Segmentation - Centre for Computational Personalized Medicine","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/sano.science\/research\/gepar3d-geometry-prior-assisted-learning-for-3d-tooth-segmentation\/","og_locale":"en_US","og_type":"article","og_title":"GEPAR3D: Geometry Prior-Assisted Learning for 3D Tooth Segmentation","og_description":"MICCAI, 2025","og_url":"https:\/\/sano.science\/research\/gepar3d-geometry-prior-assisted-learning-for-3d-tooth-segmentation\/","og_site_name":"Centre for Computational Personalized Medicine","article_publisher":"https:\/\/www.facebook.com\/sano.science\/","article_modified_time":"2026-07-15T09:44:54+00:00","twitter_card":"summary_large_image","twitter_site":"@sanoscience","twitter_misc":{"Est. reading time":"2 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/sano.science\/research\/gepar3d-geometry-prior-assisted-learning-for-3d-tooth-segmentation\/","url":"https:\/\/sano.science\/research\/gepar3d-geometry-prior-assisted-learning-for-3d-tooth-segmentation\/","name":"GEPAR3D: Geometry Prior-Assisted Learning for 3D Tooth Segmentation - Centre for Computational Personalized Medicine","isPartOf":{"@id":"https:\/\/sano.science\/#website"},"datePublished":"2026-07-15T09:44:06+00:00","dateModified":"2026-07-15T09:44:54+00:00","breadcrumb":{"@id":"https:\/\/sano.science\/research\/gepar3d-geometry-prior-assisted-learning-for-3d-tooth-segmentation\/#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/sano.science\/research\/gepar3d-geometry-prior-assisted-learning-for-3d-tooth-segmentation\/"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/sano.science\/research\/gepar3d-geometry-prior-assisted-learning-for-3d-tooth-segmentation\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/sano.science\/"},{"@type":"ListItem","position":2,"name":"Research","item":"https:\/\/sano.science\/research\/"},{"@type":"ListItem","position":3,"name":"Publications","item":"https:\/\/sano.science\/research-type\/publications\/"},{"@type":"ListItem","position":4,"name":"GEPAR3D: Geometry Prior-Assisted Learning for 3D Tooth Segmentation"}]},{"@type":"WebSite","@id":"https:\/\/sano.science\/#website","url":"https:\/\/sano.science\/","name":"Centre for Computational Personalized Medicine","description":"Sano \u2013 Centre for Computational Medicine","publisher":{"@id":"https:\/\/sano.science\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/sano.science\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Organization","@id":"https:\/\/sano.science\/#organization","name":"Sano \u2013 Centre for Computational Medicine","alternateName":"Sano","url":"https:\/\/sano.science\/","logo":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/sano.science\/#\/schema\/logo\/image\/","url":"https:\/\/sano.science\/wp-content\/uploads\/2024\/05\/logo_sano_podstawowe.png","contentUrl":"https:\/\/sano.science\/wp-content\/uploads\/2024\/05\/logo_sano_podstawowe.png","width":700,"height":265,"caption":"Sano \u2013 Centre for Computational Medicine"},"image":{"@id":"https:\/\/sano.science\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/www.facebook.com\/sano.science\/","https:\/\/x.com\/sanoscience","https:\/\/www.linkedin.com\/company\/sanoscience\/","https:\/\/www.youtube.com\/channel\/UCDZ_8TcjMWUG2ZcgKKgfpwQ","https:\/\/bsky.app\/profile\/sanoscience.bsky.social"]}]}},"acf":[],"gutenberg_blocks":[{"blockName":"custom-styles","attrs":{"styles":""}},{"blockName":"core\/paragraph","attrs":{"epAnimationGeneratedClass":"edplus_anim-eTNyM8","epGeneratedClass":"eplus-wrapper"},"innerBlocks":[],"innerHTML":"\n<p class=\" eplus-wrapper\">Precise segmentation of teeth in Cone-Beam Computed Tomography (CBCT) scans is a longstanding challenge in computational dentistry, particularly when it comes to fine anatomical structures such as root apices \u2014 whose accurate delineation is essential for evaluating root resorption in orthodontic practice.<\/p>\n","innerContent":["\n<p class=\" eplus-wrapper\">Precise segmentation of teeth in Cone-Beam Computed Tomography (CBCT) scans is a longstanding challenge in computational dentistry, particularly when it comes to fine anatomical structures such as root apices \u2014 whose accurate delineation is essential for evaluating root resorption in orthodontic practice.<\/p>\n"]},{"blockName":"core\/paragraph","attrs":{"epAnimationGeneratedClass":"edplus_anim-gDMLq9","epGeneratedClass":"eplus-wrapper"},"innerBlocks":[],"innerHTML":"\n<p class=\" eplus-wrapper\">This study presents GEPAR3D, a new method that combines instance detection and multi-class segmentation into a single unified pipeline, specifically designed to improve root-level segmentation accuracy. The approach incorporates a Statistical Shape Model of dentition as a geometric prior, enabling the model to capture anatomical context and morphological consistency without imposing rigid spatial constraints. To handle the complexity of narrow root structures, GEPAR3D employs a deep watershed technique that represents each tooth as a continuous 3D energy basin encoding voxel distances to boundaries \u2014 an instance-aware formulation that proves particularly effective for challenging apex regions.<\/p>\n","innerContent":["\n<p class=\" eplus-wrapper\">This study presents GEPAR3D, a new method that combines instance detection and multi-class segmentation into a single unified pipeline, specifically designed to improve root-level segmentation accuracy. The approach incorporates a Statistical Shape Model of dentition as a geometric prior, enabling the model to capture anatomical context and morphological consistency without imposing rigid spatial constraints. To handle the complexity of narrow root structures, GEPAR3D employs a deep watershed technique that represents each tooth as a continuous 3D energy basin encoding voxel distances to boundaries \u2014 an instance-aware formulation that proves particularly effective for challenging apex regions.<\/p>\n"]},{"blockName":"core\/paragraph","attrs":{"epAnimationGeneratedClass":"edplus_anim-cbPeZ2","epGeneratedClass":"eplus-wrapper"},"innerBlocks":[],"innerHTML":"\n<p class=\" eplus-wrapper\">Trained on publicly available CBCT data from a single center and evaluated across four external test sets from both in-house and public medical institutions, GEPAR3D demonstrates strong generalization. The method achieves a mean Dice Similarity Coefficient of 95.0% \u2014 2.8 percentage points above the next best approach \u2014 and a recall of 95.2%, representing a 9.5-point improvement. Qualitative results further confirm meaningful gains in root segmentation quality, pointing to real clinical potential for more reliable root resorption assessment and improved orthodontic decision support. Implementation and dataset are publicly available.<\/p>\n","innerContent":["\n<p class=\" eplus-wrapper\">Trained on publicly available CBCT data from a single center and evaluated across four external test sets from both in-house and public medical institutions, GEPAR3D demonstrates strong generalization. The method achieves a mean Dice Similarity Coefficient of 95.0% \u2014 2.8 percentage points above the next best approach \u2014 and a recall of 95.2%, representing a 9.5-point improvement. Qualitative results further confirm meaningful gains in root segmentation quality, pointing to real clinical potential for more reliable root resorption assessment and improved orthodontic decision support. Implementation and dataset are publicly available.<\/p>\n"]},{"blockName":"core\/spacer","attrs":{"height":"10px","epAnimationGeneratedClass":"edplus_anim-jSj8gT","epGeneratedClass":"eplus-wrapper"},"innerBlocks":[],"innerHTML":"\n<div style=\"height:10px\" aria-hidden=\"true\" class=\"wp-block-spacer eplus-wrapper\"><\/div>\n","innerContent":["\n<div style=\"height:10px\" aria-hidden=\"true\" class=\"wp-block-spacer eplus-wrapper\"><\/div>\n"]},{"blockName":"core\/paragraph","attrs":{"epAnimationGeneratedClass":"edplus_anim-6Kk4KB","epGeneratedClass":"eplus-wrapper"},"innerBlocks":[],"innerHTML":"\n<p class=\" eplus-wrapper\"><strong>Autors<\/strong>: Tomasz Szczepa\u0144ski, SzymonP\u0142otka,\u00a0Michal K.\u00a0Grzeszczyk, Arleta\u00a0Adamowicz,\u00a0Piotr\u00a0Fudalej, Przemys\u0142aw\u00a0Korzeniowski,\u00a0Tomasz\u00a0Trzci\u0144ski,\u00a0 Arkadiusz\u00a0Sitek<\/p>\n","innerContent":["\n<p class=\" eplus-wrapper\"><strong>Autors<\/strong>: Tomasz Szczepa\u0144ski, SzymonP\u0142otka,\u00a0Michal K.\u00a0Grzeszczyk, Arleta\u00a0Adamowicz,\u00a0Piotr\u00a0Fudalej, Przemys\u0142aw\u00a0Korzeniowski,\u00a0Tomasz\u00a0Trzci\u0144ski,\u00a0 Arkadiusz\u00a0Sitek<\/p>\n"]},{"blockName":"core\/spacer","attrs":{"height":"10px","epAnimationGeneratedClass":"edplus_anim-ObCQst","epGeneratedClass":"eplus-wrapper"},"innerBlocks":[],"innerHTML":"\n<div style=\"height:10px\" aria-hidden=\"true\" class=\"wp-block-spacer eplus-wrapper\"><\/div>\n","innerContent":["\n<div style=\"height:10px\" aria-hidden=\"true\" class=\"wp-block-spacer eplus-wrapper\"><\/div>\n"]},{"blockName":"acf\/button","attrs":{"title":"","button_type":"link","url":"","button_style":"primary","target":"_self","button_extra_classes":""},"innerBlocks":[],"innerHTML":"","innerContent":[]}],"meta_data":{"is_automatically_other_posts":true,"number_of_posts":"3","is_automatically_check_also_posts":true},"_links":{"self":[{"href":"https:\/\/sano.science\/index.php\/wp-json\/wp\/v2\/research\/31931","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/sano.science\/index.php\/wp-json\/wp\/v2\/research"}],"about":[{"href":"https:\/\/sano.science\/index.php\/wp-json\/wp\/v2\/types\/research"}],"version-history":[{"count":6,"href":"https:\/\/sano.science\/index.php\/wp-json\/wp\/v2\/research\/31931\/revisions"}],"predecessor-version":[{"id":31939,"href":"https:\/\/sano.science\/index.php\/wp-json\/wp\/v2\/research\/31931\/revisions\/31939"}],"wp:attachment":[{"href":"https:\/\/sano.science\/index.php\/wp-json\/wp\/v2\/media?parent=31931"}],"wp:term":[{"taxonomy":"research_type","embeddable":true,"href":"https:\/\/sano.science\/index.php\/wp-json\/wp\/v2\/research_type?post=31931"},{"taxonomy":"research_team","embeddable":true,"href":"https:\/\/sano.science\/index.php\/wp-json\/wp\/v2\/research_team?post=31931"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}