Convolutional Neural Network on DTI data for Sub-cortical Brain Structure Segmentation

Pinheiro G. R., Carmo D. S., Yasuda C., Lotufo R. A., and Rittner L.

Abstract:

Convolutional neural networks have become a powerful tool
for MRI brain analysis and are the state-of-the-art in the matter of brain
structure segmentation. Despite the deep learning power and advantages,
most of the work is still done in classical methods, such as atlas based
segmentation. The majority of those methods also uses only anatom-
ical MRI sequences, e.g. T1- and T2-weighted images, however, other
sequences of MRI could lead to much more interesting results. In this
work, we are proposing the use of Convolutional Neural Networks, in a
multitask approach, which is a tendency to the deep learning commu-
nity, in order to segment a variety of brain structures. We used over 100
subjects with 32 directions diffusion data and manual annotation, drawn
on T1 images, of 8 different brain structures. We have tested variations
in the CNN architecture and input data configurations to ensure the
best performance. Our results show the results of a particular CNN to
segment sub-cortical structures such as Ventricle, Thalamus, Putamen,
and Caudate Nucleus.