April 18, 2025
Bryan Plummer, PhD
Rafik B. Hariri Institute for Computing and Computational Science & Engineering, Boston University
Title: Towards robust and efficient channel-adaptive models for multi-channel imaging.
Time: 10:00-11:00 am
Location: Orchard View Room, Discovery Building
Abstract: Applications using multi-channel imaging like single cell analysis or satellite imaging using sophisticated and heterogeneous sensor and imaging techniques create noisy and complex data that is challenging to train machine learning models on. Specifically, most machine learning techniques assume they are provided with a standardized data input types with relatively clean data from a single data source. Thus, multi-channel imaging represent a significant challenge due to having diverse imaging types (e.g. different sensors or stains) and what labels that are available for training often represent significant noise. What’s more, even when an imaging channel is purported to contain similar information differences in sensors or collection procedures can result in a shift in the distribution of the data between training and inference times. In this talk, Dr. Plummer will showcase some of the work performed in his lab that aims to train adaptive and robust models for multi-channel imaging. In particular, he will begin with a discussion on training channel-adaptive models, and how we can take advantage of self-supervised techniques. He will then discuss how we can mitigate the effect of label noise, especially in the high levels of noise often seen in scientific datasets and when domain shifts that already exist in the training dataset make it more challenging to identify noisy samples. Finally, he will close by discussing the significant energy and storage requirements of training machine learning models, and the increasing need for resource-efficient methods of learning and performing inference.
Past Special Seminars
Faculty Candidate Seminars – Clinical Biostatistics in Cancer Position
Hongyuan Cao, PhD, Florida State University
Title: Kernel meets sieve: transformed hazards models with sparse longitudinal covariates.
Date: Friday, February 28, 2025
Time: 1:15-2:15 pm
Location: 1309 Health Sciences Learning Center
Teams Link:
Meeting ID: 255 440 769 255
Passcode: u7TP2C73
Abstract: We study the transformed hazards model with time-dependent covariates observed intermittently for the censored outcome. Existing work assumes the availability of the whole trajectory of the time-dependent covariates, which is not realistic. We propose combining kernel-weighted log-likelihood and sieve maximum log-likelihood estimation to conduct statistical inference. The method is robust and easy to implement. We establish the asymptotic properties of the proposed estimator and contribute to a rigorous theoretical framework for general kernel-weighted sieve M-estimators. Numerical studies corroborate our theoretical results and show that the proposed method performs favorably over competing methods. The analysis of a data set from a COVID-19 study in Wuhan identifies clinical predictors that otherwise cannot be obtained using existing methods.
——————————————————————————
WILL BE RESCHEDULED: Jie Yang, PhD, Stony Brook University
Title: Phase I/II Adaptive Design for Drug Combination Oncology Trials
Date: Monday, February 17, 2025
Time: 11:45-12:45
Location: Clinical Sciences Center G5/113
Teams Link: TBD
Abstract: Adaptive designs for integrated Phase I/II trials of drug combinations are increasingly utilized to accelerate drug development and optimize therapeutic efficacy, particularly in oncolgy. The model-based Continual Reassessment Method (CRM) is widely used for dose-finding, leveraging accumulated data to guide patient allocation within a Bayesian framework. In this presentation, I will introduce an adaptive Phase I/II trial design for drug combinations using the CRM approach which incorporates both binary toxicity and binary efficacy outcomes, utilizing partial orderings to simultaneously model dose-toxicity and dose-efficacy relationships. To facilitate the implementation of this design, we have developed an R package that supports real-time patient assignment and enables extensive simulation studies to evaluate operating characteristics and generate visual representations. The package offers extensive user-defined customization ensuring flexibility across diverse clinical scenarios. Additionally, I will present a backfilling-integrated design, rigorously assessed through extensive simulations, to further enhance trial efficiency and data utilization.
——————————————————————————
Xiaoqing (Jade) Wang, PhD, University of Michigan
Title: Statistical methods for spatial transcriptomics and brain imaging data with spatial heterogeneity
Date: Friday, February 21
Time: 1 – 2:00 pm
Location: 1309 Health Sciences Learning Center
Teams Link:
Meeting ID: 278 306 938 117
Passcode: 7hV6td6C
Abstract: Revolutionary technologies and treatment protocols have generated a surge in spatially resolved and imaging datasets, such as spatial transcriptomics and brain imaging, which present unique challenges in modeling spatial heterogeneity. This presentation will introduce two novel statistical approaches tailored to these complexities. First, I will present ELLA, a spatial non-homogeneous Poisson process-based method for identifying genes with distinct subcellular expression patterns in spatial transcriptomics data. ELLA detects genes with unique subcellular localizations and associates these patterns with specific mRNA characteristics, supporting the advancement of personalized medicine. Second, I will briefly introduce BI-GMRF, a Gaussian Markov random field-based mediation analysis method for 3D brain imaging. Using BI-GMRF, we identified brain regions vulnerable to treatments that impact late neurocognitive outcomes in a medulloblastoma clinical trial.
—————————————————————————–
Past Special Seminars
Tianyuan Lu, PhD, Schmidt AI in Science Postdoctoral Fellow, Department of Statistical Sciences, University of Toronto
Date: June 30, 2023
Time: Noon to 1 pm
Zoom link: https://uwmadison.zoom.us/j/91593359262?pwd=SXA5YmtPZWdDYmJwZXdHaVJFbm1WQT09
Poster: 230630 Lu, TianyuanPoster
Title: Modeling hidden genetic risk from family history for improving polygenic risk prediction and increasing yield of diagnostic sequencing.
Abstract:
Polygenic risk scores based on common genetic variants have demonstrated significant potential in both research and clinical settings. However, it is important to consider whether family history, a traditional genetic predictor, still provides valuable information. Family history of complex traits and diseases can be influenced by various factors, including the transmission of rare pathogenic variants, shared environmental exposures within families, and a common genetic predisposition.
In this presentation, I will introduce a latent factor model that aims to quantify disease risk beyond what is captured by a common genetic variant-based polygenic risk score but inferable from family history. I will discuss how this model can enhance population-level risk stratification for complex diseases such as cardiovascular diseases, Alzheimer’s disease, and idiopathic short stature. Additionally, I will discuss its potential in prioritizing individuals who are more likely to carry clinically actionable rare pathogenic variants for diagnostic sequencing.
At the end of the presentation, I will provide an overview of other ongoing and future research directions, focusing on the development and implementation of statistical genetics methods for improving the prevention, diagnosis, and treatment of complex diseases.
Relevant publications:
Lu et al. Capturing additional genetic risk from family history for improved polygenic risk prediction. Commun Biol 2022. PMID: 35710731
Lu et al. Polygenic risk score as a possible tool for identifying familial monogenic causes of complex diseases. Genet Med 2022. PMID: 35460399
Lu et al. Individuals with common diseases but with a low polygenic risk score could be prioritized for rare variant screening. Genet Med 2021. PMID: 33110269