RESEARCH
Statistical inference and uncertainty quantification
My research develops statistical learning and inference methods for biomedical imaging and spatial transcriptomics, with a focus on uncertainty quantification. I study how to quantify uncertainty in spatial effects and predictions, preserve informative biological signals, and guide data collection under limited budgets.
Interpretable spatial effects and uncertainty-aware molecular prediction.
Biomedical imaging studies often require both flexible modeling of complex spatial patterns and interpretable inference on covariate effects.
We develop a Deep Spatially Varying Coefficient Model that combines deep representation learning with interpretable spatial coefficient maps, while providing global hypothesis testing and simultaneous uncertainty quantification for spatially heterogeneous associations.
Histology-based spatial gene-expression models typically provide point predictions without indicating when those predictions can be trusted.
We develop an uncertainty-aware virtual spatial transcriptomics framework that predicts a calibrated expression distribution for each spot–gene pair, enabling reliability-aware molecular mapping and supporting downstream tasks such as region prioritization, tumor characterization, and spatial analysis.
Supporting biomedical learning when molecular assays and diagnostic labels are limited.
Spatial transcriptomics is powerful but costly, raising a practical question: which tissue slides should receive molecular profiling when only a limited ST budget is available?
AsST formulates this as a budget-constrained, cross-modal assay-allocation problem and provides a policy-agnostic framework for using histology, or combined with existing ST data, to prioritize samples for profiling while balancing predictive performance, biological coverage, and assay cost.
Self-supervised contrastive learning can overemphasize dominant visual features while suppressing subtler signals that matter for downstream diagnosis.
We develop Fisher Contrastive Learning (FCL) to preserve complementary, class-relevant information in learned representations and apply it to acute promyelocytic leukemia diagnosis, where clinically important nuclear and granule morphology must be learned from limited labeled microscopy data.
Earlier work on structure-preserving sampling and individualized risk prediction.
In the real world, there are increasingly large networks, including disease networks, contact tracing networks, brain networks, etc.
Most networks have thousands or, millions of nodes or edges. However, there are some challenges in large networks: hard to visualize; hard to analyze; and hard for downstream computation; Thus, we propose a manifold-based network sampling method that does not need access to the whole graph and utilizes local features with global information.
We develop a network-based framework for estimating individualized SARS-CoV-2 infection risk from population contact-tracing data, where each person is represented as a node and contacts define the network structure.
Using demographic, exposure, and contact information, we propose to recover missing infection labels and generate personalized risk scores by graph convolutional networks, which can further support resource-aware PCR testing policies that balance testing capacity with outbreak control.