Study of brain magnetic resonance images reconstruction through convolutional autoencoders

We studied how a variety of 2D convolutional autoencoders of different sizes performed the task of reconstructing healthy brain magnetic resonance images. The images used were the subset of T1-weighted MRI available in the IXI dataset. Two smaller ResNet-like models were created and tested to see ho...

Descripción completa

Detalles Bibliográficos
Autor: Pardo Ginés, Andrés
Tipo de recurso: tesis de maestría
Fecha de publicación:2020
País:España
Institución:Universitat Oberta de Catalunya (UOC)
Repositorio:O2, repositorio institucional de la UOC
OAI Identifier:oai:openaccess.uoc.edu:10609/121266
Acceso en línea:http://hdl.handle.net/10609/121266
Access Level:acceso abierto
Palabra clave:deep learning
MRI
autoencoder
aprenentatge profund
IRM
aprendizaje profundo
Data mining -- TFM
Mineria de dades -- TFM
Minería de datos -- TFM
Descripción
Sumario:We studied how a variety of 2D convolutional autoencoders of different sizes performed the task of reconstructing healthy brain magnetic resonance images. The images used were the subset of T1-weighted MRI available in the IXI dataset. Two smaller ResNet-like models were created and tested to see how the number of parameters in the model affected the reconstruction prowess of the autoencoders. The popular ResNet50 model was also tested under two con gurations: no pre-trained weights and pre-trained weights trained on the ImageNet dataset, to see if the transfer learning process from an unrelated computer vision task helps improve the reconstruction results. All models were trained using mean square error and difference of structural similarity as loss function to explore if the later would assist the models in improving their performances as some publications have pointed out for a range of different computer vision tasks. Both the peak signal-to-noise ratio and structural similarity tests showed that the use of difference of structural similarity as loss function did provide the best results in the test set for models using no pre-trained weights, globally the results of the ResNet-like models were effectively indistinguishable from the original images.