Anatomical Priors for Fully Automated Medical Image Segmentation

In this dissertation, novel ideas regarding the usage of prior anatomical information in fully automated image segmentation pipelines are presented and investigated. In the context of traditional segmentation methods, primitive shape priors are used to construct contour initialization methods, which complement traditional contour based segmentation approaches towards full automation. In the scope of this thesis, these initialization methods, namely Polar Appearance Models (PAMs) and Gradient based Expanding Spherical Appearance Models (GESAMs), are specifically designed for the extraction of the femoral bone in MR volumes. Regarding deep learning, full automation is already implied by the architectural end-to-end design of fully convolutional segmentation networks. Their performance can, however, be increased by sufficient incorporation of prior anatomical knowledge. In regards to shape priors, a cascaded convolutional distance transform is proposed, which directly integrates the distance transform, as a conventional representation for shape, into arbitrary segmentation networks. Moreover, two imitating encoder based architectures are introduced, in which the compressing property of convolutional autoencoders is leveraged to infuse shape information during training. Furthermore, their applicability in cross-modality and one-shot settings is demonstrated. In case of zero-shot domain adaptation, three strategies, i.e. shape priors by Oktay et al.’s ACNN [OF+18], contour infusion by edge enhancement, and feature abstraction by color augmentation, are introduced in this specific setting, all enforcing shape aware feature learning to gap the domain shift to unseen target domains. Additionally, a novel deep learning segmentation approach is presented for small structures with strong shape variations, which considers topographical priors by means of multitask learning. A similar topography aware approach is shown in an excursion to weakly supervised caries detection in smartphone images. The insights from both shape ...

2023-04-28

dissertation ; Text ; Abschlussarbeit ; doc-type:doctoralThesis ; doctoral thesis ; [Doctoral and postdoctoral thesis]

eng

ddc:004 ; Fakultät für Ingenieurwissenschaften » Informatik und Angewandte Kognitionswissenschaft

004 Data processing & computer science

https://creativecommons.org/licenses/by-sa/4.0/ ; info:eu-repo/semantics/openAccess

CC-BY-SA

https://doi.org/10.17185/duepublico/78355
https://nbn-resolving.org/urn:nbn:de:hbz:465-20230428-134430-0
https://duepublico2.uni-due.de/receive/duepublico_mods_00078355
https://duepublico2.uni-due.de/servlets/MCRFileNodeServlet/duepublico_derivate_00077973/Diss_Pham.pdf

Universität Duisburg-Essen: DuEPublico2 (Duisburg Essen Publications online)
Further nameUniversity of Duisburg-Essen: DuEPublico2 (Duisburg Essen Publications online)  

• URL: https://duepublico2.uni-due.de/
• Research Organization Registry (ROR): University of Duisburg-Essen
• Continent: Europe
• Country: de
• Latitude / Longitude: 51.458020 / 7.014760 (Google Maps | OpenStreetMap)
• Number of documents: 14,393
• Open Access: 14,391 (99%)
• Type: Academic publications
• System: MyCoRe
• Content provider indexed in BASE since: 2006-09-15
• BASE URL: https://www.base-search.net/Search/Results?q=coll:ftunivessen