DS Dissertation Proposal | Yao Su | Thursday, May 2, 9:30am - 11:00am | Unity Hall 343

Data Science 

Ph.D. Dissertation Proposal 

Yao Su, Ph.D. Candidate 

Thursday, May 2nd

9:30am - 11:00am

Unity Hall 343 Conference Room

Dissertation Committee:

Professor Xiangnan Kong, Worcester Polytechnic Institute, Advisor.

Professor Randy Paffenroth, Worcester Polytechnic Institute.

Professor Yanhua Li, Worcester Polytechnic Institute.

Professor Lifang He, Lehigh University.

Title:  Towards End-to-End Knowledge Discovery in Complex Brain Imaging Data

Abstract:

Brain imaging data mining has emerged as a prevalent paradigm in neuroscience. The goal is to extract and analyze information from neuroimaging data, aiming to reveal the brain’s structural and functional mysteries. Conventional brain imaging analysis tools (e.g., FSL, ANTs, FreeSurfer) have been applied successfully to tasks such as brain extraction, registration, segmentation, parcellation etc. However, these tasks are typically performed separately in a pipeline-based approach, and their performance necessitates visual inspections by experts across tasks for error correction. These processes are particularly burdensome for high-dimensional neuroimages (e.g., 3D sMRI, 4D fMRI), where inspecting detailed voxel-level information and conducting manual quality control are both expensive and time-consuming, presenting substantial obstacles in many medical studies.

The goal of this dissertation is to bridge multiple tasks within brain imaging analysis by studying end-to-end learning solutions that can naturally handle and capture their interrelationships. Specifically, we explore deep learning-based approaches through the following steps.

1. Anti-blur brain image registration. Image registration is an essential task for neuroimaging data analysis, where the goal is to align multiple images into one coordinate system to eliminate the interference during image acquisition (e.g., viewpoints, motion and imaging modalities). Recent research in image registration focuses on improving accuracy using multi-stage methods that often blur the image due to repeated interpolation. In this step, we study the problem of anti-blur image registration and propose a novel solution, called Anti-Blur Network (ABN), for multi-stage image registration. 

2. Unsupervised brain extraction and registration. Following the registration task, we attempt to explore its relationship with the brain extraction task. Brain extraction is the task of removing non-cerebral tissues (e.g., skull, dura and scalp) from the raw MRI scan. Conventional research mainly focuses on developing methods for the extraction and registration tasks separately under supervised settings. The performance of these methods highly depends on the amount of training samples and visual inspections performed by experts for error correction. In this step, we study the problem of unsupervised collective extraction and registration and propose a collective Extraction-Registration Network (ERNet), to jointly optimize the extraction and registration tasks without labeled data. 

3. One-shot brain extraction, registration and segmentation. To further study the task relations in brain imaging analysis, we explore how to integrate the segmentation task with extraction and registration tasks in an end-to-end manner. Brain segmentation is the task of labeling the anatomical regions in the brain MRI image, which often provide guidance for disease diagnosis. In this step, we study the problem of one-shot joint extraction, registration and segmentation in neuroimaging data, which exploits only one labeled template image and unlabeled raw images for training. We propose an end-to-end framework, called JERS, to jointly optimize the extraction, registration and segmentation tasks, allowing feedback among them. 

4. Unified model for comprehensive brain imaging tasks. Besides brain extraction, registration, and segmentation tasks, there are more preliminary yet indispensable tasks (e.g., parcellation, brain network generation, disease prediction, etc. ) that need to be conducted for comprehensive brain imaging analysis. In this step, we study the problem of end-to-end learning for comprehensive brain analysis. We proposed a unified model, UniBrain, to explore how tasks like brain extraction, registration, segmentation, parcellation, network generation, and classification interrelate, aiming to mutually boost their efficacy with minimal labeled data.