Learn about NVIDIA's latest research directions in medical AI, spanning models that derive insights directly from raw sensor data, dynamic representations of physiology beyond static anatomy, and multimodal, longitudinal patient world models. This opening session provides a high-level overview of the program and frames the themes and questions explored throughout the day.

NV-Reason-CT is a new open 3D vision language model (VLM) that brings chain-of-thought reasoning to 3D CT analysis. It processes 3D images natively and can generate structured reports, explainable diagnostic traces, and general multi-step QA about chest and abdomen CT abnormalities.

NV-Reason-Multimodal is a unified multimodal medical vision-language model that brings X-ray, 3D CT, surgical video, pathology images, and clinical text into a single framework. With dedicated 2D and video processing and MONAI-based volumetric processing, NV-Reason-Multimodal produces responses and structured outputs supported by task-relevant reasoning while delivering excellent benchmark performance across radiology, surgery, and pathology. Designed as a shared research foundation, NV-Reason-Multimodal unites clinicians, researchers, and developers to explore and advance the next generation of AI-enabled medical imaging and computer-assisted interventions.

NV-Reason-PathWSI is a reasoning vision-language model for whole-slide pathology analysis that brings clinical chain-of-thought to gigapixel WSI. It processes slide overview natively and can propose relevant regions of interest, generate structured findings, and deliver explainable traces grounded in expert-annotated regions while supporting multi-step visual reasoning across 11 organ systems. Designed as a shared research foundation for computational pathology, NV-Reason-PathWSI involves clinicians, researchers, and developers to explore and advance the next generation of AI-enabled slide-level diagnosis, region-guided analysis, and human-aligned clinical AI.

Break

Medical AI on Databricks presentation.

Getting MONAI models from research into a managed cloud workflow can take several paths. This session covers training, fine-tuning, and scaling MONAI architectures such as SegResNet, SwinUNETR, and UNet on Amazon SageMaker, including where JumpStart and bring-your-own-container approaches fit. It also introduces MedAnvil, a newly released open-source dashboard covering the path from S3 ingestion through SageMaker training, deployment, evaluation, drift monitoring, and fairness auditing.

ImPartial and MONAI session.

Break

Learn how coding agents and medical AI skills can support adaptable medical imaging workflows. See how agents can inspect data, run imaging inference, generate reports, and prototype training experiments in a containerized hospital-style environment.

Explore using MONAI to create patient-specific models of cardiac and respiratory motion from 3D CT images. Dive into workflows, do's, and don'ts for creating AI surrogates of physiological processes that generate simulations in minutes rather than hours.

Clinical Application of the MONAI Active Learning Framework session.

Healthcare robotics faces three fundamental challenges: limited real-world data, insufficient experience for robust generalization, and stringent requirements for real-time deployment. Discover recent advancements in healthcare robotics, powered by NVIDIA's three-computer solution, to address them.

Collaborative, community-driven data initiatives are unlocking research and commercial opportunities for healthcare robotics. Explore the latest in open datasets and subsequent model development.

NV-Generate: Foundation Models for 3D Medical Image Synthesis presentation.

NV-Reason-Surg presentation.

Break

Digital twins and shadows are computational representations of the real world and complex processes therein. Examine recent work on creating high-utility digital twins for peri-operative applications and leveraging them to solve some of the most pressing challenges in surgical AI.

Generative AI and physics-based simulation are converging to accelerate the development of intelligent surgery and advance CMR Surgical's mission to expand access to minimal access surgery. The central barrier is data: Surgical robotics has only a fraction of the training data available to language models and autonomous vehicles. This talk explores how Versius data and open data collaborations such as Open-H can begin to close that gap, why real-world data must be complemented by simulation, and how world foundation models such as NVIDIA Cosmos could unlock richer synthetic training environments—accelerating the path toward intraoperative autonomy.

NVIDIA Holoscan 5 provides unprecedented performance for both perception and control pipelines. See how Holoscan combines with Holoscan Sensor Bridge to turn traditional FPGA-centric products into software-defined devices supporting agentic development, validation, regulatory compliance, and deployment.

See how we enable agents to co-develop simulation applications in healthcare, from converting medical data into USD files and building simulation scenes and data collection pipelines to training and evaluating autonomous policies.