Hippocampus Segmentation Methodology From MRI Based on kernels and Local Representation

The effectiveness of brain Magnetic Resonance imaging (MRI) as a useful evaluation tool strongly depends on the performed segmentation of involved tissues or anatomical structures. The traditional approach to segment structure of interest on MRI comprise manual delineation by experts, which is a very time consuming task, becoming not appropriate for employing on large datasets or in applications where time is a critical factor, such as treatment planning. Lately, Multi-Atlas Segmentation (MAS) methods have emerge, offering an efficient alternative to deal with structural anatomical variability by fusing a set of manually labeled atlases, becoming the state of the art on automatic segmentation. However, the accuracy of these approaches is essentially influenced by atlas selection and label fusion stages. Hence, choosing a set of atlases that relate better with the input image, and thus provide more appropriate segmentations, at the same time that exclude irrelevant ones that might misguide the labeling procedure, leads obtain a better segmentation estimate than one that uses the full atlas database, at the same time that improves computational efficiency. Finally, label fusion offers the possibility to combine candidate labels from a set of warped atlases into a unique final label. In this regard, strengths of patch based approaches are proposed to cope with registration errors and high uncertainty regions, improving the segmentation accuracy. This work introduces as first instance an image representation technique based on kernels that is able to code suitably intrinsic relationships within the image domain. Furthermore, two label fusion methodologies are proposed in order to enhance state of the art label fusion methods, the former based on CKA projection of intensity patches onto label space and the latter rely on Bayesian inference. Particularly, the developed methods are evaluated in two publicly available image collections, resulting in an increased performance (assessed through the Dice Index) as compared to other recent Works

• Autores: Tobar Rodríguez, Andrés David
• Fecha de publicación: 2019
• Institución: Universidad Nacional de Colombia
• Repositorio: Universidad Nacional de Colombia
• Idioma: spa
• OAI Identifier: oai:repositorio.unal.edu.co:unal/76703
• Acceso en línea: https://repositorio.unal.edu.co/handle/unal/76703; http://bdigital.unal.edu.co/73401/
• Palabra clave: Brain Tissue Segmentation; Bayesian Segmentation; Inter-Slice Kernel; Inter-Patch Kernel, Inter-Patch Kernel; Segmentación de estructuras cerebrales; Segmentación Bayesiana; Selección de atlas
• Rights: openAccess
• License: Atribución-NoComercial 4.0 Internacional