Authors:
José A. Viteri
1
;
Francis R. Loayza
2
;
Enrique Pelaez
1
and
Fabricio Layedra
1
Affiliations:
1
Facultad de Ingeniería en Electricidad y Computación, Escuela Superior Politécnica del Litoral, Guayaquil, Ecuador
;
2
Facultad de Ingeniería en Mecánica y Ciencias de la Producción, Escuela Superior Politécnica del Litoral, Guayaquil, Ecuador
Keyword(s):
Convolutional Neural Network, U-Net, WMH Segmentation.
Abstract:
White Matter Hyperintensities (WMH) are lesions observed in the brain as bright regions in Fluid Attenuated Inversion Recovery (FLAIR) images from Magnetic Resonance Imaging (MRI). Its presence is related to conditions such as aging, small vessel diseases, stroke, depression, and neurodegenerative diseases. Currently, WMH detection is done by specialized radiologists. However, deep learning techniques can learn the patterns from images and later recognize this kind of lesions automatically. This team participated in the MICCAI WMH segmentation challenge, which was released in 2017. A dataset of 60 pairs of human MRI images was provided by the contest, which consisted of T1, FLAIR and ground-truth images per subject. For segmenting the images a 21 layer Convolutional Neural Network-CNN with U-Net architecture was implemented. For validating the model, the contest reserved 110 additional images, which were used to test this method’s accuracy. Results showed an average of 78% accuracy a
nd lesion recall, 74% of lesion f1, 6.24mm of Hausdorff distance, and 28% of absolute percentage difference. In general, the algorithm performance showed promising results, with the validation images not used for training. This work could lead other research teams to push the state of the art in WMH images segmentation.
(More)