European Conference on Computer Vision | ECCV 2026 | 2nd Edition Vol. 02 / No. 01 / 2026

Data Curation
& Augmentation
in Medical Imaging.

An ECCV workshop on the data bottleneck for robust medical imaging AI - from curated image and video datasets to multimodal clinical benchmarks spanning diagnosis, intervention, and surgical care.

September 8-9, 2026 | ECCV venue TBA Workshop dates Successor to CVPR 2024

iMED figure showing RGB views, depth maps, aligned point clouds, geometry, and tool labels. — FIG. 01iMED RGB / Depth / Geometry

CLIMB benchmark overview with modality coverage, dataset focus areas, code, and sample data. — FIG. 02CLIMB benchmark overview

SurgVU sample frames across multiple surgical tasks. — FIG. 03SurgVU surgical tasks

IIEditionAfter CVPR 2024

3DatasetsiMED | CLIMB | SurgVU

4Invited speakersConfirmed program

8-9SeptAt ECCV 2026

§ 01About the workshop

PP. 01 - 04

From scarce data
to safer medicine.

A focused program of invited talks, peer-reviewed papers, and dataset spotlights.

Reliable medical systems depend on data quality as much as model novelty. Clinical image, video, and multimodal patient data are scarce, expensive to annotate, heterogeneous, and ethically restricted. DCA-MI advances data-centric methods and benchmarks as first-class research contributions across diagnostic imaging, interventional video, multimodal clinical data, surgical scene understanding, and clinical translation.

"The bottleneck is rarely the model. It is almost always the data."
Organizing Committee

§ 02Featured speakers

Invited speakers

Invited speakers
for DCA-MI 2026.

Sophia Bano

UCL, United Kingdom

Robot vision and scene understanding for minimally invasive surgery.

Lena Maier-Hein

DKFZ / Heidelberg University, Germany

Surgical data science, benchmarking, and reproducible evaluation.

José M. M. Montiel

Universidad de Zaragoza, Spain

Visual SLAM, deformable SLAM for endoscopy, EndoMapper.

Mengya Xu

CUHK, Hong Kong

Medical AI across MICCAI, IPCAI, and ICRA.

§ 03Organizers

Organizing committee

Organized by
the DCA-MI team.

Fengyi Jiang

Primary contact | Intuitive Surgical

Sierra Bonilla

iMED dataset lead | UCL Hawkes Institute

Javier Morlana

CLiMB benchmark lead | Universidad de Zaragoza

Ray Zhang

Organizer | Intuitive Surgical

Mary Jin

Organizer | Intuitive Surgical

Shuoqi Chen

Computer Vision & Medical Imaging engineer | Intuitive Surgical

Jingpei Lu

Research Scientist | Intuitive Surgical

Rogerio Nespolo

SurgVU lead | Intuitive Surgical

§ 04Areas of focus

Aligned with invited talks

What we want
to talk about.

01

Dataset Acquisition, Annotation, and Curation in Medical Imaging

Expand

Acquisition pipelines and protocols under real clinical constraints
Annotation strategy, labelling instructions, annotator consensus and disagreement
Pseudo-ground-truth generation, calibration, and quality control
Multi-site, multi-modal, and longitudinal dataset construction
Curation transparency: documenting acquisition, exclusions, and demographic coverage
Annotation tooling, including interactive and promptable segmentation

02

Benchmarking, Evaluation, and Reproducibility

Expand

Clinically meaningful metric design for classification, detection, segmentation, and tracking
Splits, leakage analysis, and dataset contamination
Dataset cleaning, label-error detection, and de-duplication protocols
Calibration of metric choice to clinical decision-making
Reporting and reproducibility standards for medical CV
Human-in-the-loop and clinician-grounded evaluation methodologies

03

Privacy, Fairness, and Federated Learning in Medical CV

Expand

Anonymization and privacy-preserving dataset release
Federated learning under partial or heterogeneous labels; federated benchmarks
Bias auditing of training data and trained models
Fairness across demographic, anatomical, and rare-subgroup strata
Protected-attribute leakage and demographic-shortcut detection
Privacy-utility trade-offs in synthetic-data release

04

Generative Modeling for Dataset Augmentation

Expand

Diffusion models, world models, and controllable generative pipelines for images and video
Foundation-model-driven and text-conditioned synthesis
Synthetic data for under-represented populations, rare diseases, and minority subgroups
Label and structure augmentation: lesion, polyp, vessel, and surgical scene synthesis
Modality-specific augmentation, including fan-shape preservation for ultrasound
Quantifying generative utility: downstream task validity, controllability, clinical plausibility

05

Simulation, Digital Twins, and Sim-to-Real

Expand

Physics-based and anatomy-informed simulation
Digital twins of patients, organs, operating rooms, and procedures
Closed-loop simulation and surgical simulators for training-data generation
Sim-to-real adaptation methodologies
Patient-specific synthetic priors for downstream tasks

06

SLAM, 3D Reconstruction, and Neural Rendering as a Data-Centric Problem

Expand

SLAM and SfM for endoscopy, laparoscopy, bronchoscopy, endovascular interventions, IVUS, ICE, and microscopy
Neural rendering, Gaussian Splatting, and implicit representations for anatomy
Differentiable rendering and patient-specific synthetic priors
Calibration, depth, and pose estimation under deformation, smoke, and reflection
Tissue tracking, registration, and intra-operative 2D/3D alignment
Dataset and pseudo-GT design choices and their effect on reconstruction benchmarks
Generalization across patients, procedures, and imaging modalities

07

Closing the Domain Gap: Distribution Shift and Clinical Translation

Expand

Sim-to-real gaps and dataset-to-clinic transfer
Distribution shift, calibration, and out-of-distribution detection in deployment
Robust evaluation: when does benchmark performance actually transfer?
Safety, uncertainty quantification, and failure-mode analysis on clinical data
Deployment case studies and bench-to-bedside lessons
Continual learning and model maintenance after deployment
Regulatory and clinically-grounded evaluation methodologies

08

Learning with Limited, Imperfect, or Heterogeneous Data

Expand

Self-, semi-, weakly-, and un-supervised learning for medical imaging
Few-shot, zero-shot, and in-context learning
Foundation-model adaptation: vision and vision-language models for clinical tasks
Domain adaptation, generalization, and continual learning across sites and scanners
Noisy and weak label handling; learning from EMR-derived labels
Active learning and interactive segmentation
Uncertainty quantification under noisy or incomplete supervision

09

Multi-Modal Datasets and Models for Autonomous Surgery

Expand

Beyond imaging: control inputs, robot kinematics, force/torque, eye-gaze, IMU
Synchronized multi-modal acquisition pipelines for surgical procedures
Multi-modal fusion architectures and benchmarks
Datasets and tooling for autonomous surgery research
Workflow recognition, action segmentation, and surgical phase understanding from multi-modal signals
Tool-tissue interaction modelling and contact estimation
Closed-loop control and decision-making informed by multi-modal data
Vision-language models and multimodal reasoning for the operating room

§ 05Advisory board

Confirmed advisors and invited speakers

Featured advisor