STATISTICAL LEARNING METHODS IN IMAGING GENOMICS
Public Deposited- Creator
-
Luo, Tianyou
- Gillings School of Global Public Health, Department of Biostatistics
-
Luo, Tianyou
- Abstract
- The fields of medical imaging and statistical genomics have experienced significant advancements in recent years, benefiting from both the rapid development of deep learning and the invention of novel biotechnologies. Spatial transcriptomics has unlocked the potential to bridge image analysis techniques with genomics, thereby elevating our understandings of biological mechanisms to a new level. Due to the high dimensionality of both imaging and genomics data, numerous challenges persist in efficiently analyzing these novel data structures and transferring the recent success of natural image analysis to the field of imaging genomics. In this dissertation, we develop novel statistical learning methods to tackle these challenges. In the first project, we propose a robust adaptive two-sample (RATS) testing procedure in high dimensional settings to test the equality of underlying distributions. We propose a family of ℓγ−ED tests as natural extensions of the Energy Distance. Our proposed RATS test incorporates a series of ℓγ−ED tests targeting different signal patterns. To extend beyond testing marginal distributions, we also propose a rotational RATS test. Theoretical properties are derived under mild conditions. The proposed RATS test shows superior performances compared with existing methods in simulations and real data analysis spanning a wide range of signal patterns. In the second project, We propose Multi-scale Adaptive ST Deconvolution (MAST-Decon) for cell-type deconvolution in spatial transcriptomics data. MAST-Decon leverages the concept of smoothing in medical image analysis and incorporates spatial neighborhood information to improve deconvolution performance. Unlike existing spatially-unaware deconvolution methods, MAST-Decon calculates a weighted likelihood and adaptive weights without additional parametric assumptions. We carry out simulations and real data analysis to demonstrate that MAST-Decon exhibits better and more robust performance compared to state-of-the-art methods under different sequencing depths.In the third project, we propose a deep generative model for representation learning on human brain structural connectome data. We utilize separate inferential Wasserstein Generative Adversarial Networks (iWGANs) to obtain latent representations for nodes and edges in brain structural networks, and subsequently develop a conditional iWGAN framework to obtain conditional latent representations. We apply the framework to brain structural connectome data derived from UK Biobank.
- Date of publication
- 2023
- Keyword
- DOI
- Resource type
- Rights statement
- In Copyright - Educational Use Permitted
- Advisor
- Zhu, Hongtu
- Li, Yun
- Li, Tengfei
- Li, Didong
- Stein, Jason
- Degree
- Doctor of Philosophy
- Degree granting institution
- University of North Carolina at Chapel Hill Graduate School
- Graduation year
- 2023
- Language
Relations
- Parents:
This work has no parents.
Items
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
Luo_unc_0153D_22117.pdf | 2024-01-11 | Public | Download |