A Unified Vision-Language Foundation Model for Surgical Intelligence

---

Overview

Surgical intelligence requires more than general visual understanding. It involves perception of surgical scenes, temporal understanding of procedural progress, and reasoning over instruments, anatomy, actions, and safety.

SurgVLM is a unified vision-language model designed for surgical intelligence. It supports diverse surgical tasks within a single modeling pipeline, spanning visual perception, temporal analysis, and high-level reasoning.

Our project includes:

• SurgVLM-DB: a surgical multimodal corpus for training surgical vision-language models
• SurgVLM Models: two specialized variants for instruction following and reasoning
• SurgVLM-Bench: A Comprehensive Surgical Benchmarks for VLMs

---

Highlights

• Built specifically for surgical intelligence
• Supports visual perception, temporal understanding, and reasoning
• Covers 10 surgical tasks
• Includes two specialized models built on Qwen3.5-9B
• Designed for general surgical understanding and complex reasoning scenarios
• Part of the training data is publicly available on Hugging Face

---

SurgVLM-DB

SurgVLM-DB is a multimodal surgical corpus designed for training domain-specific vision-language models.

Key Characteristics

• Integrates diverse surgical data sources
• Covers multiple surgical procedures, anatomical structures, and task types
• Supports model training from low-level perception to high-level reasoning
• Provides a unified foundation for surgical vision-language learning

Public Release

Part of the training data has been publicly released through SurgSigma-DB on Hugging Face:

• Hugging Face: SurgSigma/SurgSigma-DB

Additional resources and updates will be released through the project page and repository.

---

Models

SurgVLM provides two specialized model variants, both built on Qwen3.5-9B, to support different surgical intelligence scenarios.

Model	Backbone	Description
SurgVLM-9B-Instruct	Qwen3.5-9B	An instruction-tuned model for general surgical vision-language understanding
SurgVLM-9B-Reasoning	Qwen3.5-9B	A reasoning-oriented model for more complex surgical analysis

---

Data

Resource	Download	Status	Description
SurgVLM-DB	Hugging Face	Partially released	Surgical multimodal corpus for model training