Biostatistics Events

Biostatistics Dissertation Defense

Tianyu Ding-Advances in Statistical and Machine Learning Methods for Image Data, with...-ONLINE

Wednesday 5/20 10:00AM - 12:00PM
** Online/Virtual Event **

Tianyu Ding of the Department of Biostatistics defends his dissertation on "Advances in Statistical and Machine Learning Methods for Image Data, with Application to Alzheimer's Disease".

Committee Chairpersons: Robert Krafty, PhD, Department of Biostatistics & Dana Tudorasco, Department of Psychiatry

Committee Members: Stewart Anderson, PhD, Department of Biostatistics & Anne Cohen, Department of Psychiatry


Graduate faculty of the University and all other interested parties are invited to attend via Zoom


The revolutionary development of neuroimage technology allows for the generation of large-scale neuroimage data in modern medical studies. For example, structural magnetic resonance imaging (sMRI) is widely used in segmenting neurodegenerative regions in the brain and positron-emission tomography (PET) is commonly used by clinicians and researchers to quantify the severity of Alzheimer's disease. Motivated by data from studies of aging, this dissertation proposes new statistical methods for two aging-related problems: (1) to automatically segment white matter hyperintensity (WMH), a biomarker of neurodegenerative pathology, (2) to estimate the association between neurodegeneration pathology and vascular measures.

In the first part of this dissertation, we build an improved algorithm of WMH detection in older adults. The method, which we refer to as “OASIS-AD”, is a supervised learning model based on a well-validated automated segmentation tool “OASIS” in multiple sclerosis (MS). OASIS-AD is a major refinement of OASIS that considers the specific challenges raised by WMH in Alzheimer's Disease (AD) to reduce false discoveries. We show that OASIS-AD performs better than several existing automated white matter hyperintensity segmentation approaches. In the second part of this dissertation, we develop an interpretable penalized multivariate high-dimensional method for image-on-scalar regression that can be used for both association studies and predictive modeling between high-dimensional PET images and patients' scalar measures. This method overcomes the lack of interpretability in regularized regression after reduced-rank decomposition through a novel encoder-decoder based penalty to regularize interpretable image characteristics. Empirical properties of the proposed approach are examined and compared to existing methods in simulation studies and in the analysis of PET images from subjects in a study of Alzheimer's Disease. In the third part of this dissertation, we developed ACU-Net, an efficient convolutional networks for medical image segmentation based on U-Net. The proposed deep learning networks relieve the small sample size problem of training a deep neural network when used for medical image segmentation and decrease computation cost by increasing the effective degrees of freedom through data augmentation and several novel ideas on building convolutional layers blocks to compress the model. ACU-Net has 1/20 number of parameters and 1/40 model complexity while without losing performance compared with U-Net when applied to a normal aging cohort WMH segmentation problem.

Last Updated On Thursday, May 14, 2020 by Valenti, Renee Nerozzi
Created On Thursday, May 14, 2020