Degree Name

Master of Engineering by Research


School of Electrical, Computer and Telecommunications Engineering


Accurate segmentation of different brain tissue types is the first step of understanding the neuronal activity in functional magnetic resonance imaging (fMRI). Due to the low spatial resolution of fMRI data and the absence of an automated segmentation approach, human experts require high resolution structural MRI images, which the fMRI data are superimposed on for analysis. The recent advent of high-resolution fMRI, along with temporal characteristic of fMRI data, suggests the possibility of segmenting fMRI image without relying on the high resolution structural MRI image.

This thesis proposes a patch-wise deep learning segmentation method using long-term recurrent convolutional network architecture. The proposed method comprises of three stages: spatial feature extraction with convolutional neural network, temporal feature extraction with long short-term memory, and brain tissue class prediction with softmax classifier.

The proposed method aims to segment five classes in fMRI images, which are gray matter, white matter, blood vessel, non-brain and cerebrospinal fluid. It achieves an average Dice similarity coefficient of 76.99%, which demonstrates that the proposed deep network could be used by specialists for segmenting fMRI data.