{"id":{"repo_id":"oxford-brookes","oai_identifier":"tle:22dc6de1-b80b-4563-91a7-5f52079f3137:d6bd9758-527a-46cd-bfe2-c433766e8fca:1"},"canonical_url":"https://search.dev.ndltd.org/etd/oxford-brookes/tle:22dc6de1-b80b-4563-91a7-5f52079f3137:d6bd9758-527a-46cd-bfe2-c433766e8fca:1","repository":{"repo_id":"oxford-brookes","name":"Oxford Brookes University","base_url":"https://radar.brookes.ac.uk/radar/oai"},"display":{"title":"Developing a User-Friendly and Modular Framework for Deep Learning Methods in 3D Bioimage Segmentation","abstract":"The emergence of deep learning has breathed new life into image analysis, especially for the segmentation, a challenging step required to quantify bidimensional (2D) and tridimensional (3D) objects. Despite deep learning promises, these methods are only slowly spreading in the biological field. In this PhD project, the 3D nucleus of the cell is used as the object of interest to understand how its shape variations contribute to the organisation of the genetic material. First a literature survey showed that very few publicly available methods for 3D nucleus segmentation provide the minimum requirements for their reproducibility. These methods were subsequently benchmarked and only one of them called nnU-Net surpassed the best specialized computer vision tool. Based on these observations, a new development philosophy was designed and, from it, Biom3d, a novel deep learning framework emerged. Biom3d is a user-friendly tool successfully used by biologists involved in 3D nucleus segmentation and provides a new alternative for automatically and accurately computing nuclear shape parameters. Being well optimized, Biom3d also surpasses the performance of cutting-edge methods on a wide variety of biological and medical segmentation problems. Being modular, Biom3d is a sustainable framework compatible with the latest deep learning innovations, such as self-supervised methods. Self-supervision aims at tackling the important need for deep learning methods in manual annotations by pretraining models on large unannotated datasets to extract information first before retraining them on annotated datasets. In this work, a self-supervised approach based on pretraining an entire U-Net model with the Triplet and Arcface losses was developed and demonstrates significant improvements over supervised methods for 3D segmentation. The performance, modularity and interdisciplinary nature of the tools developed during this project will serve as an innovation platform for a wide panel of users ranging from biologist users to future deep learning developers.","abstract_html":"The emergence of deep learning has breathed new life into image analysis, especially for the segmentation, a challenging step required to quantify bidimensional (2D) and tridimensional (3D) objects. Despite deep learning promises, these methods are only slowly spreading in the biological field. In this PhD project, the 3D nucleus of the cell is used as the object of interest to understand how its shape variations contribute to the organisation of the genetic material. First a literature survey showed that very few publicly available methods for 3D nucleus segmentation provide the minimum requirements for their reproducibility. These methods were subsequently benchmarked and only one of them called nnU-Net surpassed the best specialized computer vision tool. Based on these observations, a new development philosophy was designed and, from it, Biom3d, a novel deep learning framework emerged. Biom3d is a user-friendly tool successfully used by biologists involved in 3D nucleus segmentation and provides a new alternative for automatically and accurately computing nuclear shape parameters. Being well optimized, Biom3d also surpasses the performance of cutting-edge methods on a wide variety of biological and medical segmentation problems. Being modular, Biom3d is a sustainable framework compatible with the latest deep learning innovations, such as self-supervised methods. Self-supervision aims at tackling the important need for deep learning methods in manual annotations by pretraining models on large unannotated datasets to extract information first before retraining them on annotated datasets. In this work, a self-supervised approach based on pretraining an entire U-Net model with the Triplet and Arcface losses was developed and demonstrates significant improvements over supervised methods for 3D segmentation. The performance, modularity and interdisciplinary nature of the tools developed during this project will serve as an innovation platform for a wide panel of users ranging from biologist users to future deep learning developers.","abstract_has_math":false,"creators":["Mougeot, Guillaume"],"institution":"Oxford Brookes University","degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":["Graumann, Katja","Desset, Sophie","Chausse, Frédéric"],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2024,"date_issued":"2024","date_published":"2024","updated_at":"2026-07-24T03:42:47Z","subjects":[],"languages":["en"],"rights":["All rights reserved"],"rights_urls":[],"identifier_entries":[]},"links":{"outbound_url":"https://doi.org/10.24384/f92r-9717","outbound_label":"DOI","outbound_source":"dc:identifier"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor","label":"Contributor","values":["Mougeot, Guillaume","Graumann, Katja","Desset, Sophie","Chausse, Frédéric"]},{"key":"dc:creator","label":"Author","values":["Mougeot, Guillaume"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2024"]},{"key":"dc:publisher","label":"Institution","values":["Oxford Brookes University"]},{"key":"dc:type","label":"Dc Type","values":["thesis"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["en"]},{"key":"dc:rights","label":"Dc Rights","values":["All rights reserved"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["https://doi.org/10.24384/f92r-9717","https://radar.brookes.ac.uk/radar/file/22dc6de1-b80b-4563-91a7-5f52079f3137/1/Mougeot2023BioimageSegmentation.pdf"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["The emergence of deep learning has breathed new life into image analysis, especially for the segmentation, a challenging step required to quantify bidimensional (2D) and tridimensional (3D) objects. Despite deep learning promises, these methods are only slowly spreading in the biological field. In this PhD project, the 3D nucleus of the cell is used as the object of interest to understand how its shape variations contribute to the organisation of the genetic material. First a literature survey showed that very few publicly available methods for 3D nucleus segmentation provide the minimum requirements for their reproducibility. These methods were subsequently benchmarked and only one of them called nnU-Net surpassed the best specialized computer vision tool. Based on these observations, a new development philosophy was designed and, from it, Biom3d, a novel deep learning framework emerged. Biom3d is a user-friendly tool successfully used by biologists involved in 3D nucleus segmentation and provides a new alternative for automatically and accurately computing nuclear shape parameters. Being well optimized, Biom3d also surpasses the performance of cutting-edge methods on a wide variety of biological and medical segmentation problems. Being modular, Biom3d is a sustainable framework compatible with the latest deep learning innovations, such as self-supervised methods. Self-supervision aims at tackling the important need for deep learning methods in manual annotations by pretraining models on large unannotated datasets to extract information first before retraining them on annotated datasets. In this work, a self-supervised approach based on pretraining an entire U-Net model with the Triplet and Arcface losses was developed and demonstrates significant improvements over supervised methods for 3D segmentation. The performance, modularity and interdisciplinary nature of the tools developed during this project will serve as an innovation platform for a wide panel of users ranging from biologist users to future deep learning developers."]},{"key":"dc:format","label":"Dc Format","values":["application/pdf"]},{"key":"dc:title","label":"Title","values":["Developing a User-Friendly and Modular Framework for Deep Learning Methods in 3D Bioimage Segmentation"]}]}],"canonical_facts":{"dc:contributor":["Mougeot, Guillaume","Graumann, Katja","Desset, Sophie","Chausse, Frédéric"],"dc:creator":["Mougeot, Guillaume"],"dc:date":["2024"],"dc:description":["The emergence of deep learning has breathed new life into image analysis, especially for the segmentation, a challenging step required to quantify bidimensional (2D) and tridimensional (3D) objects. Despite deep learning promises, these methods are only slowly spreading in the biological field. In this PhD project, the 3D nucleus of the cell is used as the object of interest to understand how its shape variations contribute to the organisation of the genetic material. First a literature survey showed that very few publicly available methods for 3D nucleus segmentation provide the minimum requirements for their reproducibility. These methods were subsequently benchmarked and only one of them called nnU-Net surpassed the best specialized computer vision tool. Based on these observations, a new development philosophy was designed and, from it, Biom3d, a novel deep learning framework emerged. Biom3d is a user-friendly tool successfully used by biologists involved in 3D nucleus segmentation and provides a new alternative for automatically and accurately computing nuclear shape parameters. Being well optimized, Biom3d also surpasses the performance of cutting-edge methods on a wide variety of biological and medical segmentation problems. Being modular, Biom3d is a sustainable framework compatible with the latest deep learning innovations, such as self-supervised methods. Self-supervision aims at tackling the important need for deep learning methods in manual annotations by pretraining models on large unannotated datasets to extract information first before retraining them on annotated datasets. In this work, a self-supervised approach based on pretraining an entire U-Net model with the Triplet and Arcface losses was developed and demonstrates significant improvements over supervised methods for 3D segmentation. The performance, modularity and interdisciplinary nature of the tools developed during this project will serve as an innovation platform for a wide panel of users ranging from biologist users to future deep learning developers."],"dc:format":["application/pdf"],"dc:identifier":["https://doi.org/10.24384/f92r-9717","https://radar.brookes.ac.uk/radar/file/22dc6de1-b80b-4563-91a7-5f52079f3137/1/Mougeot2023BioimageSegmentation.pdf"],"dc:language":["en"],"dc:publisher":["Oxford Brookes University"],"dc:rights":["All rights reserved"],"dc:title":["Developing a User-Friendly and Modular Framework for Deep Learning Methods in 3D Bioimage Segmentation"],"dc:type":["thesis"]},"updated_at":"2026-07-24T03:42:47Z"}