A new approach for pediatric posterior fossa semantic segmentation in magnetic resonance images

• Ano de defesa: 2022
• Autor(a) principal: Dias, Larissa de Oliveira Penteado
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Fossa posterior; Fully convolutional neural networks; Magnetic resonance imaging; Pediatric brain segmentation; Posterior fossa; Redes neurais totalmente convolucionais; Ressonância magnética; Segmentação de cérebro infantil; Segmentação semântica; Semantic segmentation
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45134/tde-08112022-071554/

The analysis of brain magnetic resonance imaging (MRI) exams is an essential task for the diagnosis and treatment of various diseases. The manual examination of such images is time-consuming and prone to inter observer variability. Moreover, the analysis of neonatal and pediatric exams poses intrinsic challenges due to the smaller size of the brain structures and the greater inter patient variability, because of the childrens neurological development, especially during the first two years of life. Therefore, the development of automatic methods to perform the semantic segmentation of MRI data is important to aid the doctors at examining such images. In order to automatically obtain the segmentation of a MRI volume, there are both 2D and 3D methods. Fully Convolutional Neural Networks (FCN) have been presenting increasingly better results at the segmentation of both natural and medical images. In this project, we developed a new approach to perform the segmentation of the posterior fossa and the fourth ventricle regions on pediatric brain MRI data, using the FCN called LiviaNet, which is a patch 3D approach. These are the regions of occurence of the medulloblastoma, a common cancer that affects childrens brains. The identification of this tumor is of interest for the doctors from the Childrens Institute (HC-FMUSP). They provided 32 MRI volumes for this project, from children with ages ranging from less than a year to 18 years. Our method was able to identify the region of interest with a mean dice score of 0.74, thus showing the potential of the proposed approach.