Enhanced preprocessing and adaptive weighted loss function for improved for white matter hyperintensity segmentation with convolutional neural networks.

There is a great interest in automating White Matter Hyperintensities (WMH) segmentation due to their importance in the medical eld as well as the great amount of inter- and intra-observer variability that appears when it is manually segmented in magnetic resonance imaging. In this work we present a...

Descripción completa

Detalles Bibliográficos
Autor: Duque Asens, Pablo
Tipo de recurso: tesis de maestría
Fecha de publicación:2020
País:España
Institución:Universidad Nacional de Educación a Distancia
Repositorio:e-spacio. Repositorio Institucional de la UNED
Idioma:inglés
OAI Identifier:oai:e-spacio.uned.es:20.500.14468/14523
Acceso en línea:https://hdl.handle.net/20.500.14468/14523
Access Level:acceso abierto
Palabra clave:1203.04 Inteligencia artificial
Descripción
Sumario:There is a great interest in automating White Matter Hyperintensities (WMH) segmentation due to their importance in the medical eld as well as the great amount of inter- and intra-observer variability that appears when it is manually segmented in magnetic resonance imaging. In this work we present a multistep tailored preprocessing consisting mainly of brain extraction, intensity contrast enhancement, subject based slice cropping and intensity standardization. The segmentation task is then performed by a fully convolutional neural network with attention gates which employs a customized loss function based on the dice similarity coecient and the F1 score. Experimental results on the white matter hyperintensities segmentation challenge [Kuijf et al., 2019] show that our proposed preprocessing improves segmentation, that attention gated U-Net further improves segmentation tasks compared to the original U-Net and our proposed loss function has the potential to improve lesion-wise F1 on DSC based segmentations.