White Matter Hyperintensities Segmentation with Prototype Learning

This paper proposes a new method -based on meta-learning- for the WMH Segmentation Challenge, organized by UMC Utrecht, VU Amsterdam, and NUHS Singapore hospitals. The purpose of this challenge is to compare methods for the semantic segmentation of white matter hyperintensities (WMH), which are brai...

Descripción completa

Detalles Bibliográficos
Autor: Alarcón Palomar, Óscar
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:20.500.14468/14590
Acceso en línea:https://hdl.handle.net/20.500.14468/14590
Access Level:acceso abierto
Palabra clave:1203.04 Inteligencia artificial
white matter hyperintensities
white matter lesions
meta learning
few-shot learning
prototype learning
semantic segmentation
convolutional neural network
Descripción
Sumario:This paper proposes a new method -based on meta-learning- for the WMH Segmentation Challenge, organized by UMC Utrecht, VU Amsterdam, and NUHS Singapore hospitals. The purpose of this challenge is to compare methods for the semantic segmentation of white matter hyperintensities (WMH), which are brain white matter lesions, of presumably vascular origin in brain imaging obtained with magnetic resonance. White matter hyperintensities are found in patients with brain diseases like Parkinson, Alzheimer or stroke. Semantic segmentation refers to the process of linking each pixel in an image to a class label. The semantic segmentation of images has had a great advance with convolution neural networks, but they require a large number of images to be able to obtain good results. Convolutional neural networks are a type of neural networks specialized on images which architecture is similar to neurons’ pattern in human brain and they were inspired by the organization of the visual cortex. With the aim to reduce the number of images required in training, in this work, we propose the use of meta-learning algorithms, in particular prototype learning, to do this semantic segmentation. In addition, this approach also allows the network to be used in a different task for which it was not trained, which can improve its potential use. Results obtained suggest that it could be possible to use the network trained to a specific task (i.e., detect WMH in the brain), to another task (i.e. detect any kind of tumors in the brain).