MedCompute 2026 — NeurIPS Workshop

Topic & Motivation

Why This Workshop, Why Now

An estimated 4.2 billion medical imaging studies are performed annually worldwide, yet despite the U.S. Food and Drug Administration clearing over 1,000 AI/ML-enabled medical devices, real-world clinical deployment remains a negligible fraction of global workflows. The compute infrastructure necessary to close that gap now exists: GPU-accelerated reconstruction pipelines, real-time edge inference, and scalable foundation model (FM) pretraining. What the community still lacks is scientific agreement on how to utilize this computational layer to build systems that are robust, verifiable, and safe at global clinical scale. MedCompute 2026 is where that systematic work begins.

Large-scale visual FMs demonstrate remarkable zero-shot transfer across diverse clinical modalities, with specialized architectures matching or exceeding specialist performance on structured diagnostic tasks. Yet generalized robustness remains fragile: a comprehensive 2025 benchmark (LMOD+) evaluating 24 state-of-the-art multimodal large language models found that frontier architectures achieved only ~57.8% accuracy on ophthalmic disease screening under zero-shot conditions, with performance collapsing toward random baselines on complex staging tasks. The vast majority of clinical AI tools remain trapped in the research phase, or fail post-deployment due to uncharacterized distribution shifts and calibration errors. The bottleneck is not data or architecture alone — it is the absence of compute-aware design principles that make FMs robust, efficient, and deployable across heterogeneous real-world clinical environments.

Open Problems This Workshop Addresses

MedCompute 2026 is organized around three open problems the community has not yet systematically solved:

Compute-aware model design

How GPU memory constraints, reconstruction latency, and edge inference budgets should inform FM architecture and training.

Closing the robustness gap

How to build evaluation frameworks that predict real-world clinical performance, not just benchmark accuracy, under distribution shift and scanner heterogeneity.

Verified clinical deployment

What a compute-to-clinic pipeline looks like when it must simultaneously satisfy regulatory, privacy, and equity constraints.

Three Converging Factors in 2026

Edge compute platforms are achieving clinical readiness. 2025 marked a watershed moment for medical AI infrastructure, as high-performance edge platforms achieved regulatory clearance for real-time intraoperative use. Open-source, domain-specific frameworks now support multimodal inference directly at the point of care. The infrastructure is ready; the foundational algorithmic science must now catch up.

The deployment and validation frontier severely lags. While in silico performance metrics advance rapidly, real-world robustness remains brittle. Recent analyses show that fewer than 30% of FDA-authorized radiology AI devices have undergone any clinical testing, with only 5% subject to rigorous prospective evaluation. Closing this trust deficit through verifiable, hardware-aware model design and federated evaluation is the workshop's core scientific mandate.

The Sydney moment. NeurIPS 2026 places a flagship ML conference on Asia-Pacific soil for the first time. MedCompute 2026 is designed to bring Australia's world-class medical imaging institutions and AI-native companies into direct conversation with the global NeurIPS community.

Differentiation from Related Workshops

Compute-first framing

Existing workshops treat compute as background infrastructure. MedCompute 2026 inverts this: compute-aware design — GPU-native reconstruction, hardware-software co-design, and real-time edge inference — is the primary scientific object of study. A paper on MRI acceleration that ignores reconstruction latency on clinical hardware would be out of scope here.

Full pipeline scope

Recent series (e.g., AIM-FM at NeurIPS) focus on FM pretraining; the MICCAI ecosystem targets specific bottlenecks in isolation. MedCompute 2026 covers the full arc from GPU-accelerated reconstruction through FM pretraining to robustness under distribution shift and real-world deployment — treating each stage as an interconnected bottleneck in a single unified pipeline.

3D/volumetric native

CT, MRI, PET/CT, and ultrasound are fundamentally volumetric. The costs of volumetric attention can no longer be bypassed by 2D slicing, which destroys spatial context and introduces avoidable artifacts. Our program centers 3D ViTs, Mamba and State Space Models, and implicit neural representations as first-class topics.

Scope

Topics of Interest

We invite submissions across the following interconnected themes. Full research contributions, position papers, negative results, dataset releases, and infrastructure papers are all explicitly encouraged.

GPU-accelerated image reconstruction

Deep unrolling, learned priors, and diffusion-based methods for CT, MRI, PET/CT, and ultrasound; hardware-software co-design; acceleration benchmarks; handling of high-resolution data including whole-slide microscopy and digital breast tomosynthesis (DBT).

Foundation model pre-training for medical imaging

Self-supervised and contrastive objectives; multimodal (image–report–genomic) pre-training; scaling laws specific to medical domains; open-source FM releases.

Multimodal AI and generative models

Vision–language models integrating imaging with radiology reports, EHRs, and genomics; clinical report generation; diffusion-based synthesis for augmentation, privacy, and rare pathology simulation; usage of synthetic data and digital twins for training, validation, and regulatory evaluation.

3D and volumetric vision

Architectures native to volumetric data (3D ViTs, Mamba and State Space Models, implicit neural representations); efficient volumetric attention; anatomy- and topology-aware losses; cross-modality 3D alignment.

Evaluation, robustness, and interpretability

Distribution shift across scanners, sites, and populations; evaluation beyond aggregate AUC; dataset curation and bias auditing; prospective vs. retrospective validation; calibration under covariate shift; conformal prediction; concept-based explanations for imaging FMs; human–AI collaboration and trust.

Clinical deployment, regulatory pathways, and governance

Federated and privacy-preserving learning; edge inference on surgical and diagnostic devices; workflow integration and clinical validation; Green AI; regulatory pathways for AI/ML-based SaMD (FDA 510(k)/De Novo, CE Mark under the EU AI Act, TGA); post-market surveillance and continual learning.

Program

Workshop Schedule

A full-day in-person workshop structured around alternating focused talks and interactive discussion. The morning establishes the compute foundation; the afternoon translates it to the clinic.

09:00–09:15

Welcome and Framing — Organizers

09:15–10:00

Keynote — Aengus Tran, Harrison.ai, Australia Keynote

10:00–10:30

Coffee Break & Poster Session I

10:30–11:05

Featured Talk: Compute Foundation — Sajid Javed, Khalifa University FeaturedGPU-accelerated reconstruction and FM pre-training

11:05–11:45

Open Discussion — Building on the Compute Foundation

11:45–13:15

Lunch Break

13:15–14:15

Contributed Papers: Spotlight Talks (6 × 10 min) Contributed

14:15–15:00

Coffee Break & Poster Session II

15:00–15:35

Featured Talk: Clinical Translation — Hyungjeong Yang, Chonnam National University FeaturedRobustness, deployment, and regulatory pathways

15:35–16:45

Panel Discussion — From Compute to Clinic: What Will It Actually Take? Panel

16:45–17:00

Best Paper Award and Closing Remarks

Invited Speakers

Speaker Lineup

★

Aengus Tran AU/APAC Keynote · Confirmed

Harrison.ai, Australia

Large-scale clinical AI deployment · AI-assisted radiology · scaling medical AI products

★

Sajid Javed Morning Session · Confirmed

Khalifa University of Science and Technology, UAE

Computational pathology · histopathological image analysis

★

Hyungjeong Yang Afternoon Session · Confirmed

Chonnam National University, South Korea

Medical AI · multimodal healthcare intelligence · clinical AI translation · trustworthy AI

Diversity & Inclusion

Our Commitment

Diversity is a design constraint of this workshop, not an afterthought.

Organizers

The team spans three countries.

Speakers & PC

Selected for gender, geographic, and career-stage diversity: ≥40% female speakers; PC from ≥10 countries with ≥35% female members; panel to include at least one clinician from the Global South.

Submissions

Registration fees waived for accepted authors from low- and middle-income countries, with ≥2 spotlight slots reserved for early-career researchers.

Organizing Team

Organizers

Oge Marques

NVAITC Senior Scientist, NVIDIA AI Technology Centre; Emeritus Professor of Engineering and Computer Science, Florida Atlantic University

Oge Marques works at the interface of academic research and NVIDIA's AI stack, focusing on medical imaging and foundation models. He is Emeritus Professor of Engineering and Computer Science at Florida Atlantic University, where he taught and led research in visual AI and medical image analysis for over two decades, supervising 10 PhD and 17 Master's students to completion. His editorial service includes membership on the Editorial Board of Multimedia Tools and Applications (Springer, 2012–2024), and he serves as Book Series Editor for Applied Machine Learning (Springer, 2021–present). His professional service includes his current role as Chair of the Recognition and Awards Committee of the IEEE Palm Beach Section, Senior Membership in both IEEE and ACM, and active membership in medical imaging societies including SIIM, RSNA, EuSoMII, ESR, and AMIA. He is a Leshner Leadership Institute Public Engagement Fellow of the American Association for the Advancement of Science (AAAS). MedCompute 2026 is his first NeurIPS workshop proposal. His involvement brings direct access to NVIDIA's medical AI compute stack — including MONAI, Holoscan, Clara, and BioNeMo — as well as a long-standing research network spanning four continents.

Tomasz Bednarz

Director of Strategic Research Collaboration, NVIDIA AI Technology Centre; Visiting Scientist at CSIRO; Adjunct Professor at University of Queensland

Tomasz Bednarz is the Director of Strategic Research Collaboration and Special Projects at NVIDIA AI Technology Centre. In this capacity, he leads a team focused on engaging with leading researchers and premier research institutions, driving the adoption of cutting-edge NVIDIA software technologies in innovative, computationally intensive projects designed to tackle some of the world's most complex scientific challenges. His research background spans computational science, physics, computational imaging and visualization, and HPC; he brings to MedCompute 2026 direct access to NVIDIA's medical AI compute stack — including MONAI, Holoscan, Clara, and BioNeMo — as well as deep ties to the Australian research computing ecosystem through CSIRO and national infrastructure partners including NIF and Pawsey. He has co-organized workshops and symposia at ACM and IEEE conferences, eResearch Australasia, Web3D, and has over 20 years of experience bridging academic research and industrial deployment of AI systems.

Judy Wawira Gichoya

Associate Professor, Department of Radiology and Imaging Sciences, Emory University School of Medicine

Judy Wawira Gichoya is an Associate Professor of Radiology and Imaging Sciences at Emory University School of Medicine, where her research focuses on algorithmic fairness, hidden bias in medical imaging AI, and equitable clinical deployment across diverse populations. Trained as both an informatician and an interventional radiologist, she co-leads the Healthcare AI Innovation and Translational Informatics (HITI) Lab at Emory, which builds diverse imaging datasets, validates AI models in real-world settings, and develops multimodal fusion models combining imaging and clinical text. She is internationally recognized for her landmark study demonstrating that AI models silently encode patient race across multiple imaging modalities — published in The Lancet Digital Health — a finding with broad implications for bias auditing and patient safety in deployed clinical AI. She serves on the editorial boards of NEJM AI and Radiology: Artificial Intelligence, and is an NIH DATA Scholar supporting the DSI-Africa program for building data science capacity across the African continent. Her involvement in MedCompute 2026 brings a critical perspective on the equitable deployment of AI across diverse clinical populations, particularly in low- and middle-income settings.

Olivier Salvado

Professor of AI, co-director of the QUT AI in Health Research Network, Faculty of Engineering, Queensland University of Technology

Olivier Salvado is a Professor of AI at Queensland University of Technology, where he leads research to translate advanced machine learning to clinical applications. He is the co-director of the AI in Health research network at QUT to bridge engineering with health. He has been an investigator in one of the largest Australian studies on Alzheimer's biomarkers since 2007 and contributed to several key neuroimaging analysis methods. Professor Salvado was formerly the head of imaging and computer vision at the CSIRO, led large AI initiatives across the organisation, and contributed to numerous large-scale Australian studies and technology development. He was involved in several international conference organisations including as general co-chair of ISBI in 2017, program co-chair of ISBI in 2023, and will contribute to organising MICCAI 2028. His expertise in medical image analysis and computer vision for medical applications is brought to MedCompute 2026, leveraging his network within the Asia-Pacific clinical AI ecosystem that the workshop is designed to engage.

Soon Ki Jung

Professor, School of Computer Science and Engineering, Kyungpook National University

Soon Ki Jung is a Professor in the School of Computer Science and Engineering at Kyungpook National University, South Korea, where he leads the Visual AI Lab research group. His recent research focuses on human behavior analysis, facial privacy and privacy-preserving facial image analysis, medical image segmentation, and multimodal foundation models for medical image analysis, including 3D CT-based body composition analysis. He has published more than 290 peer-reviewed papers and has served on the program committees of ACM RACS, CGI, RSL-CV, FCV, and IVCNZ. He has co-organized major international research events, including serving as the General Chair of IEEE VR 2026, the Conference Chair of SIGGRAPH Asia 2022, and the General Chair of the 27th International Workshop on Frontiers of Computer Vision. His role in MedCompute 2026 strengthens the Asia-Pacific representation of the workshop and its connections to the communities of computer vision, medical image segmentation, and multimodal medical foundation models.

Program Committee

Confirmed Members

We are targeting a final PC of 25–30 members with representation from at least 10 countries and ≥35% female members.

Ge Wang

Rensselaer Polytechnic Institute, USA

CT/MRI reconstruction · imaging AI

Alexandre Falcão

University of Campinas (UNICAMP), Brazil

Image processing · pattern recognition

Felipe Kitamura

Eden, USA / UNIFESP, Brazil

Radiology AI · clinical deployment

Leticia Rittner

University of Campinas (UNICAMP), Brazil

Brain MRI analysis · white matter segmentation · diffusion MRI

Peter van Ooijen

UMCG, Netherlands

AI in radiotherapy · medical imaging informatics

Elena Sizikova

FDA, USA

Regulatory science · robustness

Issam El Naqa

Moffitt Cancer Center, USA

Radiotherapy AI · interpretability

Veronika Cheplygina

IT University of Copenhagen, Denmark

Evaluation methodology · label-efficient learning

Michael Riegler

Simula Research Laboratory, Norway

Medical multimedia AI · AI safety and robustness

Yu Tian

University of Central Florida, USA

Interpretability · uncertainty quantification

Jayashree Kalpathy-Cramer

University of Colorado Anschutz, USA

Medical imaging AI · fairness