Seminars

Abstract: Recent progress in end-to-end learning has led to highly accurate models for automated radiological image interpretation. Yet, most models lack transparency and fail to ground predictions in a clinically meaningful foundation—contrasting sharply with the way radiologists diagnose by referencing shared medical knowledge. In this talk, I will present CheXomni, a foundation model designed to bridge this gap by aligning model predictions with established clinical observations. Trained on over 0.87 million image-report pairs from 227,835 chest X-ray studies, CheXomni uses large language model (LLM) embeddings to project radiological findings into a shared semantic space. This enables interpretable and auditable predictions while maintaining strong diagnostic accuracy. We demonstrate CheXomni’s effectiveness across four large, physician-annotated chest X-ray datasets spanning the U.S., Europe, and Asia. Beyond superior classification performance, CheXomni introduces novel capabilities: zero-shot pathology detection, radiological confounder auditing, and concept bottleneck modeling—all grounded in real clinical data. These advances mark a significant step toward transparent, generalizable, and trustworthy medical AI.

Bio: Dr. Tianyu Han is an Assistant Professor in the Department of Radiology at the University of Pennsylvania and a core faculty member of the AI2D Center. His research focuses on developing interpretable and generalizable AI systems for radiological data understanding. He has led work on multimodal models, medical image synthesis, and adversarial robustness, with applications spanning chest X-rays, breast MRI, and CT imaging. His work is published in Radiology, Nature Communications, JAMA, Science Advances, Cell Reports Medicine, and Nature Machine Intelligence. Dr. Han is also the lead author of MedAlpaca, an open-source framework for medical conversational AI, and is committed to building transparent, clinically useful AI for medical diagnosis.

Abstract: Individuals vary widely in how their brains age and these differences have important clinical consequences. For example, advanced age is the strongest known risk factor for Alzheimer’s disease (AD). This talk focuses on recent findings from our research suggesting that brain aging, as quantified using structural MRI, can modify a person’s risk for AD. Individuals with signs of advanced brain aging are more likely to progress from normal cognition to cognitive impairment (CI), whereas those with slower brain aging appear more resilient. New results highlight that regional brain aging patterns are indicative of CI conversion. We also find that chronic disease risk factors — such as cardiovascular and metabolic conditions, as well as traumatic brain injury — may lead to earlier or more pronounced brain aging. Importantly, we explore how women's health factors, including menopause and reproductive history, influence brain aging trajectories and cognition. These findings offer promising avenues for improving risk stratification and patient-specific disease prevention strategies based on quantification of aging in the brain.

Bio: Andrei Irimia, PhD is an associate professor in the Leonard Davis School of Gerontology at the University of Southern California, with courtesy appointments in biomedical engineering and quantitative biology. His research focuses on brain aging, traumatic brain injury, and Alzheimer’s disease, using advanced neuroimaging and quantitative methods to understand individual variability in aging trajectories and dementia risk. Dr. Irimia leads several NIH-funded studies examining how chronic disease variables and women's health factors influence brain aging and neurodegeneration. His work bridges population neuroscience and clinical neurology, with the goal of improving early detection and stratification of patients at risk for cognitive decline.

Abstract: TBA

Bio: TBA

Abstract: TBA

Bio: TBA

Abstract: TBA

Bio: Yonghyun Nam is a Research Assistant Professor of Informatics at the University of Pennsylvania’s Perelman School of Medicine. His research focuses on integrating multi-omics data—including genomics, proteomics, and electronic health records—to advance risk prediction for complex diseases. He develops machine learning frameworks that combine polygenic risk scores, proteomic biomarkers, and network-based models to identify individuals at elevated risk for a range of complex disorders.