10.07.2025

 

Capturing Complexity in Surgical Environments

MCML Research Insight - With Ege Özsoy, Chantal Pellegrini, Felix Tristram, Kun Yuan, David Bani-Harouni, Matthias Keicher, Benjamin Busam and Nassir Navab

©Özsoy et al.

Figure 1: Visual Summary of a single timepoint in MM-OR, illustrating the multimodal data provided for each sample: RGB-D video from multiple angles, detailed RGB views, low-exposure video, point cloud data, robot screen and tracker logs, audio and speech transcripts, panoptic segmentations, semantic scene graphs, and downstream task annotations such as robot phase, next action, and sterility breach status.

Introduced in a CVPR 2025 paper by MCML Junior Members Ege Özsoy, Chantal Pellegrini, Felix Tristram, Kun Yuan, David Bani-Harouni, Benjamin Busam and Matthias Keicher, MCML PI Nassir Navab and collaborators Tobias Czempiel and Ulrich Eck, MM-OR is a comprehensive multimodal dataset created to significantly enhance our understanding of operating room (OR) dynamics. It captures robotic knee replacement surgeries (see Figure 1) using various sophisticated sensors, including multiple RGB-D cameras, audio recorders, infrared trackers, and real-time robotic system logs.

Semantic Scene Graph Generation with MM2SG

This detailed data collection allows for generating semantic scene graphs - structured representations of interactions between people, tools, and equipment in the OR. Alongside this, the researchers present MM2SG, an innovative multimodal model capable of interpreting and integrating diverse data types to generate detailed semantic scene graphs that accurately represent OR activities (see Figure 2).

©Özsoy et al.

Figure 2: Overview of the proposed MM2SG architecture for multimodal scene graph generation. MM2SG processes a variety of data sources through specialized encoders, projecting them into a shared space. The language model generates scene graph triplets describing SGs with entities E i and predicates p i . Downstream tasks leverage entire sequences of scene graphs rather than individual ones.

A New Benchmark

What sets MM-OR apart from existing datasets is its unprecedented scale, realism, and multimodality, significantly surpassing previous efforts that often suffered from limited size, narrow scope, or lack of diverse data types. By providing panoptic segmentation annotations and supporting complex downstream tasks like sterility breach detection and action anticipation, MM-OR establishes a robust new benchmark for evaluating and developing advanced OR modeling techniques.

Open Challenges

Despite these advancements, open challenges remain, including accurately modeling rare surgical actions and effectively generalizing models to diverse surgical scenarios beyond knee replacement procedures.

Why It Matters

Because better understanding of surgical environments means enhanced situational awareness, improved safety, and more effective surgical assistance.

Curious to Explore More?

The full paper published at the CVPR 2025, one of the highest-ranked AI/ML conferences, provides a more in-depth exploration of the potential of this new dataset and technology, laying the foundations for future advancements in OR systems.

E. Özsoy, C. Pellegrini, T. Czempiel, F. Tristram, K. Yuan, D. Bani-Harouni, U. Eck, B. Busam, M. Keicher and N. Navab.
MM-OR: A Large Multimodal Operating Room Dataset for Semantic Understanding of High-Intensity Surgical Environments.
CVPR 2025 - IEEE/CVF Conference on Computer Vision and Pattern Recognition. Nashville, TN, USA, Jun 11-15, 2025. DOI GitHub

Abstract

Operating rooms (ORs) are complex, high-stakes environments requiring precise understanding of interactions among medical staff, tools, and equipment for enhancing surgical assistance, situational awareness, and patient safety. Current datasets fall short in scale, realism and do not capture the multimodal nature of OR scenes, limiting progress in OR modeling. To this end, we introduce MM-OR, a realistic and large-scale multimodal spatiotemporal OR dataset, and the first dataset to enable multimodal scene graph generation. MM-OR captures comprehensive OR scenes containing RGB-D data, detail views, audio, speech transcripts, robotic logs, and tracking data and is annotated with panoptic segmentations, semantic scene graphs, and downstream task labels. Further, we propose MM2SG, the first multimodal large vision-language model for scene graph generation, and through extensive experiments, demonstrate its ability to effectively leverage multimodal inputs. Together, MM-OR and MM2SG establish a new benchmark for holistic OR understanding, and open the path towards multimodal scene analysis in complex, high-stakes environments.

MCML Authors

Ege Özsoy

→ Group Nassir Navab
 Computer Aided Medical Procedures & Augmented Reality

Chantal Pellegrini

→ Group Nassir Navab
 Computer Aided Medical Procedures & Augmented Reality

Felix Tristram

→ Group Nassir Navab
 Computer Aided Medical Procedures & Augmented Reality

Kun Yuan

→ Group Nassir Navab
 Computer Aided Medical Procedures & Augmented Reality

David Bani-Harouni

→ Group Nassir Navab
 Computer Aided Medical Procedures & Augmented Reality

Benjamin Busam

Dr.

→ Group Nassir Navab
 Computer Aided Medical Procedures & Augmented Reality

Matthias Keicher

Dr.

→ Group Nassir Navab
 Computer Aided Medical Procedures & Augmented Reality

Nassir Navab

Prof. Dr.

Principal Investigator

Computer Aided Medical Procedures & Augmented Reality

Also check out the YouTube Video showing the dataset.

Additional Material: Practical Examples and Recording Setup

©Özsoy et al.

Figure 3: Recording setup and sensors overview. A grey circle by each sensor shows quantity; if absent, the sensor count is one.

To provide deeper insights into how MM-OR and MM2SG work in practice, the following visuals illustrate the technical setup used to collect such comprehensive data (see Figure 3) and showcase a real example of MM2SG generated outputs (see Figure 4).

©Özsoy et al.

Figure 4: Qualitative examples from a test take in MM-OR, illustrating scene graph generation performance of MM2SG. Unlabeled edges indicate the ”close to” predicate.