Keynote

Abstract

Every diagnostic decision begins with where a clinician's eyes land — yet for decades, this rich cognitive signal has been discarded. In this talk, I trace a ten-year arc of research transforming radiologist gaze from passive behavioral data into an active computational signal that fundamentally reshapes how AI systems learn, segment, and diagnose. Beginning with Gaze2Segment (2016) and C-CAD (2019), which first demonstrated that fixation patterns encode expert knowledge transferable to deep networks, I show how this idea matured through GazeSAM and GazeGNN (2023–2024) into real-time, registration-free integration with foundation models — eliminating the preprocessing bottleneck that long prevented clinical deployment. I then present our latest systems — EyeSee, GazeMind, EyeTune, and GazeAssist (2025–2026) — which close the loop entirely: AI no longer just consumes gaze but predicts, validates, and augments it, creating a bidirectional human-AI cognitive partnership. Across several systems we developed, one principle endures: the most powerful signal in medical AI was never in the pixels — it was in the eyes reading them.

Bio

Prof. Ulas Bagci is the director of the Machine and Hybrid Intelligence Lab and a tenured faculty member at Northwestern University's Radiology, Biomedical Engineering (BME), and Electrical and Computer Engineering (ECE) departments. His research interests are artificial intelligence, machine learning, and their applications in biomedical and clinical imaging. Dr. Bagci has authored more than 520 peer-reviewed articles. Dr. Bagci holds several NIH grants as Principal Investigator and serves as a steering committee member of AIR (Artificial Intelligence Resource) at the NIH. He has served as an area chair for MICCAI for several years and is an associate editor of top medical AI journals including IEEE Transactions on Medical Imaging and Medical Image Analysis. He teaches medical image computing and advanced machine learning courses and has received several international and national recognitions including best paper awards, best reviewer awards, editorial recognitions, and outstanding researcher and mentor awards.

Abstract

Open medical data can provide a foundation for AI systems that perceive, reason, generate, and act. This talk presents our open data initiatives across 3D imaging, longitudinal cancer studies, surgical video, and healthcare robotics. I will show how these resources support foundation models for segmentation, controllable CT and MRI generation, accelerated MRI reconstruction, and radiology reasoning. I will also discuss surgical world models and methods that turn video into training data for robotic policies.

The second half focuses on agents for medical AI research. AutoMedBench evaluates both research outcomes and the stages required to produce them. Building on those stage-level signals, BaT trains the model policy, VERA improves the policy and agent harness through verified environments, and MIAgent makes research decisions and experimental evidence visible to human investigators. Together, these efforts show how open resources, verifiable evaluation, and human oversight can support progress from medical imaging models toward autonomous research and physical AI.

Bio

Dr. Daguang Xu is a Senior Research Manager in Healthcare at NVIDIA, where he leads AI research in healthcare. His work spans medical imaging analysis, multimodal foundation models, and, more recently, physical AI and world models. His team is the primary contributor to the open-source platforms MONAI, and is increasingly focused on releasing open foundation models, and open datasets for generative AI in healthcare.