Deep Learning Techniques for Image Segmentation of Whole Body MRI

Recently, deep learning methods have achieved state of the art results across many fields of research. Deep learning has become prevalent in medical image analysis and plays an integral part of contemporary healthcare. In particular, contemporary image segmentation techniques make extensive use of deep artificial neural networks. This thesis presents two deep learning architectures for segmenting 3D multiparametric MRIs of patients with limb girdle muscular dystrophy. Muscular dystrophy is a genetic disease that causes muscle to be replaced by fatty tissue over time. Segmenting images of patients with muscular dystrophy has not been done prior to this work, and no labeled datasets of muscular dystrophy images are openly available. Image segmentations provide quantitative metrics, such as the fraction of muscle voxels out of all non-background voxels, that can help clinical researchers track the progression of the disease in a patient. The fact that a labeled dataset does not exist poses a challenge. A relatively small, manually labeled dataset of tissue signatures is created to train convolutional neural networks and stacked sparse autoencoders. The success of these deep learning models relies on the high contrast provided by multiparametric MRIs for distinguishing voxels of different classes. The models achieve excellent segmentation accuracy in slices of the thighs. In addition, a deep reinforcement learning framework for locating anatomical landmarks is developed as a step toward improving the segmentation models' performance. A single agent model and a multi-agent model for landmark localization are demonstrated to successfully operate in multiple anatomical regions in a diverse set of imaging modalities and image acquisition parameters. Automatic landmark localization can be used for improving segmentations in low contrast areas as well as for automatic image registration algorithms. This thesis establishes a baseline for a universal deep reinforcement learning framework for such applications