Segmentation of Brain MRI Structures with Deep Machine Learning

Several studies on brain Magnetic Resonance Images (MRI) show relations between neuroanatomical abnormalities of brain structures and neurological disorders, such as Attention Defficit Hyperactivity Disorder (ADHD) and Alzheimer. These abnormalities seem to be correlated with the size and shape of these structures, and there is an active field of research trying to find accurate methods for automatic MRI segmentation. In this project, we study the automatic segmentation of structures from the Basal Ganglia and we propose a new methodology based on Stacked Sparse Autoencoders (SSAE). SSAE is a strategy that belongs to the family of Deep Machine Learning and consists on a supervised learning method based on an unsupervisely pretrained Feed-forward Neural Network. Moreover, we present two approaches based on 2D and 3D features of the images. We compare the results obtained on the different regions of interest with those achieved by other machine learning techniques such as Neural Networks and Support Vector Machines. We observed that in most cases SSAE improves those other methods. We demonstrate that the 3D features do not report better results than the 2D ones as could be thought. Furthermore, we show that SSAE provides state-of-the-art Dice Coefficient results (left, right): Caudate (90.63±1.4, 90.31±1.7), Putamen (91.03±1.4, 90.82±1.4), Pallidus (85.11±1.8, 83.47±2.2), Accumbens (74.26±4.4, 74.46±4.6).