MedFM-Bench Workshop

About the Workshop

Why Medical Foundation Models & Benchmarks?

Medical foundation models are rapidly reshaping medical image understanding, enabling stronger generalization and more unified learning across modalities (X-ray, CT, MRI, ultrasound) and tasks such as classification, detection, segmentation, report generation, and multimodal reasoning. A key differentiator of this workshop is the explicit focus on the co-evolution of foundation models and benchmarking — not only building stronger models, but also establishing evaluation protocols that diagnose robustness, cross-site generalization, and evidence consistency.

Pillar I

Foundation Model Design & Learning

Model architectures, large-scale pretraining and self-supervised learning, multimodal learning, alignment, cross-modality reasoning, and efficient fine-tuning and adaptation strategies.

ArchitecturesSelf-supervised learningMultimodal alignmentEfficient fine-tuning

Pillar II

Data, Scaling & Generalization

Global-scale multi-institution datasets, collaboration frameworks and governance, privacy-preserving learning (federated learning), data curation, domain generalization under distribution shift, robustness, safety, and trustworthiness.

Data scalingFederated learningDomain shiftSafety & Trust

Pillar III

Benchmarking & Deployment

Evaluation standards, benchmark design, reproducible protocols, and clinical deployment — developing emerging benchmarks that diagnose robustness, generalization, and evidence consistency for high-stakes clinical settings.

BenchmarksReproducibilityClinical deploymentEvaluation protocols

Keynotes & Invited Talks

Invited Speakers

Leading researchers sharing complementary perspectives on medical foundation models, multimodal reasoning, and clinical AI.

Michael Moor

ETH Zürich

Lalithkumar Seenivasan

Johns Hopkins University

Weidi Xie

Shanghai Jiao Tong University

Xiaoxiao Li

UBC

Program

Workshop Schedule

September 8, 2026 · Room: Quality View Oresundssalen - 2.
All times are in CEST (Europe/Stockholm). An afternoon program with four invited talks, a poster session with coffee break, and five oral presentations.

Time	Event	Type
2:00 PM10 min	Opening Remarks Welcome and workshop overview from the organizers
2:10 PM30 min	Invited Talk 1 Michael Moor	Keynote
2:40 PM30 min	Invited Talk 2 Lalithkumar Seenivasan	Keynote
3:10 PM60 min	☕ Poster Session & Coffee Break Interactive poster session — up to 30 posters on display	Poster
4:10 PM10 min	RadMatch: Auditable Radiology Report Evaluation via Finding-Level Matching	Oral
4:20 PM10 min	Consistent View Alignment Improves Vision Foundation Models for 3D Medical Image Analysis	Oral
4:30 PM10 min	MMBU: A Massive Multi-modal Biomedical Understanding Benchmark to Probe the Perception Capabilities of Vision-Language Models	Oral
4:40 PM10 min	Stateful Visual Encoders for Vision-Language Models	Oral
4:50 PM10 min	DistillPath: An Efficient 22M Distilled Pathology Encoder Approaching Large Foundation Model Performance	Oral
5:00 PM30 min	Invited Talk 3 Weidi Xie	Keynote
5:30 PM30 min	Invited Talk 4 Xiaoxiao Li	Keynote

28 Accepted Papers

Accepted Papers

How Medical Multimodal Large Language Models Think in Space
When Do VLM-Generated Labels Help? A Scaling-Law Diagnostic for Medical Imaging Foundation Models
Controlled One-Click Benchmarking and Post-Fusion Recalibration for Support-Conditioned Medical Segmentation
Consistent View Alignment Improves Vision Foundation Models for 3D Medical Image Analysis
How Does SAM-3 Adapt? Capacity, Update Geometry, and Domain Transfer
RadMatch: Auditable Radiology Report Evaluation via Finding-Level Matching
Look, Ask, Diagnose: A Multi-Agent Visual Dialogue Framework for Medical Disease Diagnosis
V-REX: Efficient Specialist VLM Training for Veterinary X-Rays
Anatomy Contextualized Adaption of CT Foundation Models
MMBU: A Massive Multi-modal Biomedical Understanding Benchmark to Probe the Perception Capabilities of Vision-Language Models
PixCell-Text: Adapting a Pathology Image Prior for Text Conditioning
Reliable Normal Projection with Healthy-Reference Calibration for Brain MRI Anomaly Localization
MedDeepSearch: Agentic Search for Medical VQA with Reinforcement-Trained LLMs
What Your CXR Reveals Beyond the Pathology: Causal Evidence for Demographic Leakage in Medical Foundation Model Embeddings and Its Calibration Fairness Consequences
Rethinking Real-World MRI Denoising: Learning from Physical Noise
Stateful Visual Encoders for Vision-Language Models
S²L-Net: Structured Saliency Fusion and Local-Global Recalibration for Medical Image Segmentation
Conserved Immune Topology Improves Pathology Foundation Model Generalization for Cross-Cancer MSI-H Prediction
Discrete Diffusion Language Models for Interactive Radiology Report Drafting
BM1b: Data-Efficient Scaling of Bone Marrow Foundation Models through Domain-Specific Self-Supervised Learning and Dense Morphology-Preserving Representation Learning
A Deep Learning Benchmark for Wound Tissue Segmentation
DistillPath: An Efficient 22M Distilled Pathology Encoder Approaching Large Foundation Model Performance
Beyond Training from Scratch: Foundation Models for Data-Efficient and Generalizable Cardiac MRI Reconstruction
Beyond Natural-Image Foundation Models: Benchmarking Satellite Pretraining for Ophthalmic Image Analysis
Jolia: Concept-Level Vision-Language Alignment for 3D CT Contrastive Learning
Curia-2: Scaling Self-Supervised Learning for Radiology Foundation Models
CheXSpatialVQA: A Multi modal Spatial Understanding Benchmark for Chest X-Ray Interpretation
Do Brain MRI Foundation Models Capture Temporal Change? An Evaluation on Post-Treatment Glioma