Abstract

Brain tissue segmentation is a common medical image processing problem that deals with identifying a region of interest in the human brain from medical scans. It is a fundamental step towards neuroscience research and clinical diagnosis. Magnetic resonance (MR) images are widely used for segmentation in view of their non-invasive acquisition, and high spatial resolution and various contrast information. Accurate segmentation of brain tissues from MR images is very challenging due to the presence of motion artifacts, low signal-to-noise ratio, intensity overlaps, and intra- and inter-subject variability. Convolutional neural networks (CNNs) recently employed for segmentation provide remarkable advantages over the traditional and manual segmentation methods, however, their complex architectures and the large number of parameters make them computationally expensive and difficult to optimize. In this thesis, a novel learning-based algorithm using a three-dimensional deep convolutional neural network is proposed for efficient parameter reduction and compact feature representation to learn end-to-end mapping of T1-weighted (T1w) and/or T2-weighted (T2w) brain MR images to the probability scores of each voxel belonging to the different labels of brain tissues, namely, white matter (WM), gray matter (GM) and cerebrospinal fluid (CSF) for segmentation. The basic idea in the proposed method is to use densely connected convolutional layers and residual skip-connections to increase representation capacity, facilitate better gradient flow, improve learning, and significantly reduce the number of parameters in the network. The network is independently trained on three different loss functions, cross-entropy, dice similarity, and a combination of the two and the results are compared with each other to investigate better loss function for the training. The model has the number of network parameters reduced by a significant amount compared to that of the state-of-the-art methods in brain tissue segmentation. Experiments are performed using the single-modality IBSR18 dataset containing high-resolution T1-weighted MR scans of diverse age groups, and the multi-modality iSeg-2017 dataset containing T1w and T2w MR scans of infants. It is shown that the proposed method provides the best performance on the test sets of both datasets amongst all the existing deep-learning based methods for brain tissue segmentation using the MR images and achieves competitive performance in the iSeg-2017 challenge with the number of parameters that is 47% to 98% lower than that of the other deep-learning based architectures.