Biomedical signal and image synthesis using deep learning

Abstract

Accurate modeling and characterization of biomedical data is imperative for comprehensive understanding and reliable medical diagnosis. However, developing automated solutions to various biomedical challenges is a complex task as biomedical data is inherently high-dimensional with large variability and non-linear and non-stationary dynamics. Consequently, machine learning algorithms that offer data-driven model-free approach have been used extensively for a multitude of tasks such as synthesis, classification and segmentation. However, conventional methods that use hand-engineered features (based on prior knowledge) do not generalize well due to heterogeneity inherent to biomedical data. In recent years, deep learning models have consistently outperformed previous state-of-the-art machine learning techniques in several fields, with computer vision being one of the most prominent domains. With their ability to exploit compositional structure to find complex patterns in high dimensional and heterogeneous data, deep learning has emerged as a critical tool for analyzing biomedical data. With growing availability of public datasets, several efforts have been directed towards synthesizing biomedical signals and images based on phenomenologic or datadriven models to address diverse range of challenges such as data augmentation, signal encoding and transfer of knowledge across modalities or domains. Inspired by their overwhelming success in the field of computer vision, this thesis investigates the feasibility of using deep learning algorithms for biomedical signal and image synthesis. Specifically, we addressed the following challenges: (a) Synthesizing functional Positron Emission Tomography (PET) scans from structural Magnentic Resonance (MR) images to facilitate accurate diagnosis of Alzheimer’s disease with incomplete multi-modal data (b) Capturing complex temporal patterns underlying Electroencephalography (EEG) microstate sequences for better characterization of neuro-psychiatric disorders and (c) Learning a non-linear transfer function to map EEG signals to corresponding blood oxygenation level dependent (BOLD) signals to investigate the relationship between brain’s electrophysiological and hemodynamic measures. Generic image-to-image translation, signal encoding and sequence-to-sequence models were tailored to address these specific tasks by incorporating domain specific knowledge. Extensive evaluation of our proposed techniques against state-of-the-art algorithms demonstrates their ability to handle complex, multi-modal, multi-variate biomedical data towards providing better understanding and clinical decision support. Finally, we utilized visualization and interpretation strategies to overcome the limited interpretability of the black-box deep learning models. These tools were most effective in identifying key features and meaningful representations that proved valuable in gaining novel insights from the data.