Machine Learning Methods for Medical and Biological Image Computing

MACHINE LEARNING METHODS FOR MEDICAL AND BIOLOGICAL IMAGE COMPUTING Rongjian Li Old Dominion University, 2016 Director: Dr. Shuiwang Ji Medical and biological imaging technologies provide valuable visualization information of structure and function for an organ from the level of individual molecules to the whole object. Brain is the most complex organ in body, and it increasingly attracts intense research attentions with the rapid development of medical and biological imaging technologies. A massive amount of high-dimensional brain imaging data being generated makes the design of computational methods for efficient analysis on those images highly demanded. The current study of computational methods using hand-crafted features does not scale with the increasing number of brain images, hindering the pace of scientific discoveries in neuroscience. In this thesis, I propose computational methods using high-level features for automated analysis of brain images at different levels. At the brain function level, I develop a deep learning based framework for completing and integrating multi-modality neuroimaging data, which increases the diagnosis accuracy for Alzheimer’s disease. At the cellular level, I propose to use three dimensional convolutional neural networks (CNNs) for segmenting the volumetric neuronal images, which improves the performance of digital reconstruction of neuron structures. I design a novel CNN architecture such that the model training and testing image prediction can be implemented in an endto-end manner. At the molecular level, I build a voxel CNN classifier to capture discriminative features of the input along three spatial dimensions, which facilitate the identification of secondary structures of proteins from electron microscopy images. In order to classify genes specifically expressed in different brain cell-type, I propose to use invariant image feature descriptors to capture local gene expression information from cellular-resolution in situ hybridization images. I build image-level representations by applying regularized learning and vector quantization on generated image descriptors. The developed computational methods in this dissertation are evaluated using images from medical and biological experiments in comparison with baseline methods. Experimental results demonstrate that the developed representations, formulations, and algorithms are effective and efficient in learning from brain imaging data.