MechE PhD Prospectus Defense: Guoyao Shen

TITLE: WHEN DEEP LEARNING MEETS MEDICAL IMAGING: IMAGE RECONSTRUCTION AND BEYOND

ABSTRACT: Deep neural networks have achieved superior performance in multiple fields in natural language processing, computer vision, robotics control, etc. Medical imaging tools such as magnetic resonance imaging (MRI), computed tomography (CT) are essential diagnostic tool that provides invaluable diagnostic information in medicine that profoundly influences medical decision making and helps to improve patient outcomes. While the abundance of works for vision-related tasks, medical imaging tasks typically include various image corruptions and down-samplings, limited dataset for a universal end-to-end training process. This prospectus focuses on developing deep learning-based models that suits the needs of medical imaging. We first propose a cross-domain (k-space domain and image domain) network for accelerated MRI reconstruction, which provides superior performance evaluated by metrics and blind tests by a radiologist. We develop a regularization by neural style transfer (RNST) framework for magnetic strength transformation with limited data, which requires no pre-training and can be deployed out-of-box. We validate our model with multiple pulse sequences (T1-, T2- and proton density-weighted) on 3T and 1.5T scanners. To further explore the potential of generative models, we propose a k-space cold diffusion model that performs image degradation in k-space instead of image space without the need for Gaussian noise. K-space diffusion achieves superior performance in multiple under-sampling setups compared to conventional CNN-based networks. Moreover, a magnetic resonance image processing transformer (MR-IPT) is developed for general accelerated MRI reconstruction. MR-IPT outperforms conventional CNN-based networks with zero- and few-shot learning and can reach steady downstream task performance with rather limited dataset scale. In the future, we intend to perform more studies in model framework design for more general medical imaging tasks with modular design. For example, this would require co-training process for the main skeleton and encoders for multiple prompt encoders. Universal data loaders and down-stream tuning techniques are essential for further deployment.

COMMITTEE: ADVISOR/CHAIR Professor Xin Zhang, ME/ECE/BME/MSE; Professor Lei Tian, ECE/BME; Professor Chuanhua Duan, ME/MSE; Professor Yannis Paschaladis, ECE/SE/BME; Professor Chad W. Farris, BU Chobanian & Avedisian School of Medicine