03/18/2024
By Danielle Fretwell

The Francis College of Engineering, Department of Electrical and Computer Engineering, invites you to attend a Doctoral Dissertation defense by Yongshun Xu on: "Medical Imaging Data Synthesis and Its Deep Learning based Applications."

Candidate Name: Yongshun Xu
Degree: Doctoral
Defense Date: Wednesday, March 27, 2024
Time: 2-4 p.m.
Location: Ball Hall 302

Committee:

  • Advisor: Hengyong Yu, Professor, Department of Electrical and Computer Engineering, UMass Lowell
  • Yu Cao, Professor, Department of Computer and Information Sciences, UMass Lowell
  • Tricia Chigan, Department of Electrical and Computer Engineering, UMass Lowell
  • Yan Luo, Department of Electrical and Computer Engineering, UMass Lowell

Brief Abstract:

Medical imaging has been a critical topic for many years, with commonly used imaging modalities including Computed Tomography (CT) and Ultrasound image. The imaging techniques significantly aid in clinical physicians' diagnoses. With the rapid evolution of deep learning techniques in recent years, including convolutional neural networks (CNNs), transformers, and diffusion models, the applications in medical imaging and medical image analysis have expanded significantly, encompassing tasks such as classification, image denoising, super-resolution, and image restoration. However, due to the regulations of clinical practices, it is very challenging to collect sufficient medical images/labels to train a neural network for practical applications. To overcome this challenge, this dissertation explores the possibility to synthesize/generate realistic medical images, and apply them to different medical imaging tasks.

In this dissertation, we first introduce the fundamentals of CT imaging and its associated challenges in Chapter one. Then, in chapter two, we propose to use synthesized CT image datasets to augment real datasets, addressing the common challenge of lacking large, well-labeled datasets. Initially, we generate synthetic image datasets through simulation and apply semi-automatic labeling methods to create labels. Subsequently, we compare the performance of the same model trained with and without the inclusion of synthetic image data. This approach leverages the potential of synthetic images to overcome the limitations associated with the need for extensive labeled datasets. In chapter three, the research focus on improving the quality of limited-angle reconstruction images by leveraging the strengths of both CNNs and transformers. The results demonstrated a significant reduction in artifacts and noise, achieving superior image quality compared to the existing state-of-the-art competing reconstruction algorithms. In chapter four, the research focuses on a generative model, diffusion model, to address the challenge of angular range flexibility in limited-angle reconstruction, an area where the supervised learning typically faces limitations. We propose to integrate the diffusion model with an iterative refinement framework. The method incorporates medical image sparsity, achieving notable performance improvements. In chapter five, our research extends to ultrasound imaging, with a focus on image segmentation and classification tasks to identify fatty liver conditions. This demonstrates the wide applicability and potential of the methods developed throughout the study. In the future, we will explore the potential of those data synthesis approached to reduce radiation doses without sacrificing image quality, and we will also investigate the applications of the diffusion model across different imaging modalities.